| 1 | """v3.0.0 orchestration pipeline.""" |
| 2 | |
| 3 | from __future__ import annotations |
| 4 | |
| 5 | import copy |
| 6 | import math |
| 7 | import queue |
| 8 | import re |
| 9 | import sqlite3 |
| 10 | import sys |
| 11 | import threading |
| 12 | import time |
| 13 | from collections import Counter |
| 14 | from concurrent.futures import ThreadPoolExecutor, as_completed |
| 15 | from dataclasses import dataclass, field, replace |
| 16 | from datetime import date, datetime, timedelta, timezone |
| 17 | from pathlib import Path |
| 18 | from shutil import which |
| 19 | from typing import Any |
| 20 | |
| 21 | from . import ( |
| 22 | arxiv, |
| 23 | bird_x, |
| 24 | bluesky, |
| 25 | corpus, |
| 26 | dates, |
| 27 | dedupe, |
| 28 | digg, |
| 29 | dripstack, |
| 30 | entity_extract, |
| 31 | env, |
| 32 | github, |
| 33 | grounding, |
| 34 | hackernews, |
| 35 | health, |
| 36 | hiring_signals, |
| 37 | http, |
| 38 | instagram, |
| 39 | jobs, |
| 40 | linkedin, |
| 41 | library, |
| 42 | library_index, |
| 43 | log, |
| 44 | normalize, |
| 45 | permission_preflight, |
| 46 | perplexity, |
| 47 | pinterest, |
| 48 | planner, |
| 49 | polymarket, |
| 50 | providers, |
| 51 | query, |
| 52 | reddit, |
| 53 | reddit_listing, |
| 54 | reddit_public, |
| 55 | relevance, |
| 56 | rerank, |
| 57 | schema, |
| 58 | signals, |
| 59 | snippet, |
| 60 | stocktwits, |
| 61 | techmeme, |
| 62 | threads, |
| 63 | tiktok, |
| 64 | topic_shape, |
| 65 | truthsocial, |
| 66 | trustpilot, |
| 67 | xai_x, |
| 68 | xiaohongshu_api, |
| 69 | xquik, |
| 70 | xurl_x, |
| 71 | youtube_yt, |
| 72 | ) |
| 73 | from .cluster import cluster_candidates |
| 74 | from .fusion import weighted_rrf |
| 75 | |
| 76 | DISCOVERY_SOURCES = ("reddit", "hackernews", "digg", "x") |
| 77 | _DISCOVERY_GENERIC_DOMAIN_TERMS = { |
| 78 | "ai", "artificial", "intelligence", "tech", "technology", "trending", "trend", |
| 79 | } |
| 80 | |
| 81 | DEPTH_SETTINGS = { |
| 82 | "quick": {"per_stream_limit": 6, "pool_limit": 15, "rerank_limit": 12}, |
| 83 | "default": {"per_stream_limit": 12, "pool_limit": 40, "rerank_limit": 40}, |
| 84 | "deep": {"per_stream_limit": 20, "pool_limit": 60, "rerank_limit": 60}, |
| 85 | } |
| 86 | |
| 87 | SEARCH_ALIAS = { |
| 88 | "hn": "hackernews", |
| 89 | "bsky": "bluesky", |
| 90 | "truth": "truthsocial", |
| 91 | "web": "grounding", |
| 92 | "xhs": "xiaohongshu", |
| 93 | "xquik": "x", # xquik is a backend of the single "x" source, not its own source |
| 94 | } |
| 95 | |
| 96 | # trustpilot is capped at 1: every subquery would use the identical company |
| 97 | # identifier, so N streams are pure redundancy -- and each extra stream risks |
| 98 | # its own WAF-cookie Chrome harvest. |
| 99 | MAX_SOURCE_FETCHES: dict[str, int] = {"x": 2, "jobs": 1, "linkedin": 1, "stocktwits": 1, "trustpilot": 1} |
| 100 | |
| 101 | |
| 102 | def _resolve_depth_settings(depth: str, config: dict[str, Any]) -> dict[str, int]: |
| 103 | """Depth profile with optional CLI cap overrides applied (issue #716). |
| 104 | |
| 105 | Returns a copy so the module-level DEPTH_SETTINGS is never mutated. Overrides |
| 106 | are set directly (not max()) so callers can also lower a cap. `--max-results` |
| 107 | raises the final ranked pool (pool_limit/rerank_limit); `--max-per-source` |
| 108 | raises the per-stream truncation applied before pooling. The per-source fetch |
| 109 | cap (`--max-source-fetches`) is applied separately at the fetch site. |
| 110 | """ |
| 111 | settings = dict(DEPTH_SETTINGS[depth]) |
| 112 | # `is not None` (not truthiness) so an explicit 0 is honored as a real lower |
| 113 | # bound rather than ignored as "unset" — matches how main() stashes these. |
| 114 | max_per_source = config.get("_max_per_source") |
| 115 | if max_per_source is not None: |
| 116 | settings["per_stream_limit"] = int(max_per_source) |
| 117 | max_results = config.get("_max_results") |
| 118 | if max_results is not None: |
| 119 | settings["pool_limit"] = int(max_results) |
| 120 | settings["rerank_limit"] = int(max_results) |
| 121 | return settings |
| 122 | |
| 123 | # Per-handle result caps for the X handle-search lanes. The FROM lane (the |
| 124 | # subject's own timeline) is the single best source for a person topic, so it |
| 125 | # gets the highest cap; the ABOUT (mention) and related-handle lanes stay |
| 126 | # modest so total volume and request budget don't balloon. |
| 127 | FROM_LANE_COUNT_PER = 8 |
| 128 | MENTION_LANE_COUNT_PER = 5 |
| 129 | RELATED_HANDLE_COUNT_PER = 3 |
| 130 | |
| 131 | |
| 132 | def _has_perplexity_provider(config: dict[str, Any]) -> bool: |
| 133 | return bool(config.get("PERPLEXITY_API_KEY") or config.get("OPENROUTER_API_KEY")) |
| 134 | |
| 135 | MOCK_AVAILABLE_SOURCES = [ |
| 136 | "reddit", |
| 137 | "x", |
| 138 | "youtube", |
| 139 | "tiktok", |
| 140 | "instagram", |
| 141 | "hackernews", |
| 142 | "bluesky", |
| 143 | "truthsocial", |
| 144 | "polymarket", |
| 145 | "grounding", |
| 146 | "xiaohongshu", |
| 147 | "github", |
| 148 | "perplexity", |
| 149 | "threads", |
| 150 | "pinterest", |
| 151 | "digg", |
| 152 | "arxiv", |
| 153 | "techmeme", |
| 154 | "trustpilot", |
| 155 | "jobs", |
| 156 | "linkedin", |
| 157 | "corpus", |
| 158 | "dripstack", |
| 159 | ] |
| 160 | |
| 161 | |
| 162 | def normalize_requested_sources(sources: list[str] | None) -> list[str] | None: |
| 163 | if not sources: |
| 164 | return None |
| 165 | normalized = [] |
| 166 | for source in sources: |
| 167 | key = SEARCH_ALIAS.get(source.lower(), source.lower()) |
| 168 | if key not in normalized: |
| 169 | normalized.append(key) |
| 170 | return normalized |
| 171 | |
| 172 | |
| 173 | def available_sources( |
| 174 | config: dict[str, Any], |
| 175 | requested_sources: list[str] | None = None, |
| 176 | *, |
| 177 | x_pending: bool | None = None, |
| 178 | local_only: bool = False, |
| 179 | ) -> list[str]: |
| 180 | """List the sources the next run can serve. |
| 181 | |
| 182 | ``local_only=True`` is the safe/diagnose flavor (doctor's permission |
| 183 | block): availability is answered from local evidence only, so the X |
| 184 | check never spawns xurl's live ``whoami`` network call. Research-time |
| 185 | callers keep the default live semantics. |
| 186 | """ |
| 187 | available: list[str] = [] |
| 188 | # reddit_public needs no API key - always available |
| 189 | available.append("reddit") |
| 190 | if corpus.resolve_directories( |
| 191 | config.get("_CORPUS_DIRS"), config.get("LAST30DAYS_CORPUS_DIRS") |
| 192 | ): |
| 193 | available.append("corpus") |
| 194 | if config.get("SCRAPECREATORS_API_KEY"): |
| 195 | available.extend(["tiktok", "instagram"]) |
| 196 | if env.get_x_source(config, local_only=local_only): |
| 197 | available.append("x") |
| 198 | else: |
| 199 | # Safe inspection (--diagnose/--preflight) skips browser-cookie |
| 200 | # extraction, so get_x_source is None even though a real run would |
| 201 | # authenticate X via FROM_BROWSER. Report it as available so consumers |
| 202 | # of available_sources (SKILL.md ACTIVE_SOURCES_LIST) don't under-report. |
| 203 | # diagnose() precomputes the predicate and passes it via x_pending to |
| 204 | # avoid evaluating it twice in one diagnose() call. |
| 205 | if x_pending is None: |
| 206 | x_pending = env.x_pending_browser_auth(config) |
| 207 | if x_pending: |
| 208 | available.append("x") |
| 209 | if which("yt-dlp") or env.is_youtube_sc_available(config): |
| 210 | available.append("youtube") |
| 211 | available.extend(["hackernews", "polymarket"]) |
| 212 | # StockTwits is gated to ticker/crypto topics only (flag set in run()). |
| 213 | if config.get("_financial_topic"): |
| 214 | available.append("stocktwits") |
| 215 | # GitHub is reachable via the unauthenticated REST tier too, so it is |
| 216 | # available even without a token/gh CLI (a token only raises rate limits). |
| 217 | available.append("github") |
| 218 | # DripStack is opt-in only (owner decision, #791): a commercial |
| 219 | # third-party API must never receive default-run traffic. Opt in per run |
| 220 | # (--search dripstack) or persistently (INCLUDE_SOURCES=dripstack in |
| 221 | # .env, the LinkedIn/Perplexity pattern); the search API is free and |
| 222 | # public (no key), so the opt-in itself is the gate. |
| 223 | include_sources = { |
| 224 | token.strip() |
| 225 | for token in (config.get("INCLUDE_SOURCES") or "").lower().split(",") |
| 226 | if token.strip() |
| 227 | } |
| 228 | if "dripstack" in include_sources or ( |
| 229 | requested_sources and "dripstack" in requested_sources |
| 230 | ): |
| 231 | available.append("dripstack") |
| 232 | if which("digg-pp-cli"): |
| 233 | available.append("digg") |
| 234 | # arXiv is default-on when its Printing Press CLI is installed (zero auth). |
| 235 | # The adapter relevance-and-recency gates so it stays quiet off-topic. |
| 236 | if which("arxiv-pp-cli"): |
| 237 | available.append("arxiv") |
| 238 | # Techmeme is default-on when its CLI is installed (zero auth; sub-second |
| 239 | # local sync before each run's first search). |
| 240 | if which("techmeme-pp-cli"): |
| 241 | available.append("techmeme") |
| 242 | if env.is_bluesky_available(config): |
| 243 | available.append("bluesky") |
| 244 | if env.is_truthsocial_available(config): |
| 245 | available.append("truthsocial") |
| 246 | # Grounding (general web) is available when a paid backend is configured OR |
| 247 | # the keyless floor is permitted (i.e. the host has no native search). On a |
| 248 | # native-search host with no paid key, keyless_web_allowed is False and the |
| 249 | # engine leaves general web to the model's own search. |
| 250 | if (config.get("BRAVE_API_KEY") or config.get("EXA_API_KEY") |
| 251 | or config.get("SERPER_API_KEY") or config.get("PARALLEL_API_KEY") |
| 252 | or env.keyless_web_allowed(config)): |
| 253 | available.append("grounding") |
| 254 | if requested_sources and "jobs" in requested_sources: |
| 255 | available.append("jobs") |
| 256 | # Perplexity Sonar: opt-in additive source via INCLUDE_SOURCES=perplexity |
| 257 | if _has_perplexity_provider(config) and ( |
| 258 | "perplexity" in include_sources or (requested_sources and "perplexity" in requested_sources) |
| 259 | ): |
| 260 | available.append("perplexity") |
| 261 | # LinkedIn: opt-in additive source via INCLUDE_SOURCES=linkedin (same |
| 262 | # consent pattern as Perplexity). Unlike tiktok/instagram, which are |
| 263 | # offered during SKILL.md Step 0 onboarding, LinkedIn is power-user-only |
| 264 | # and must not silently activate for existing SCRAPECREATORS_API_KEY |
| 265 | # holders. |
| 266 | if config.get("SCRAPECREATORS_API_KEY") and ( |
| 267 | "linkedin" in include_sources or (requested_sources and "linkedin" in requested_sources) |
| 268 | ): |
| 269 | available.append("linkedin") |
| 270 | # Trustpilot: opt-in additive source via INCLUDE_SOURCES=trustpilot (same |
| 271 | # consent pattern as Perplexity/LinkedIn). Off by default -- unlike arXiv and |
| 272 | # Techmeme, which are zero-auth, it can spawn a one-time headless-Chrome WAF |
| 273 | # cookie harvest on a brand topic, so activating it is the user's choice. |
| 274 | if which("trustpilot-pp-cli") and ( |
| 275 | "trustpilot" in include_sources or (requested_sources and "trustpilot" in requested_sources) |
| 276 | ): |
| 277 | available.append("trustpilot") |
| 278 | if ( |
| 279 | "xiaohongshu" in include_sources |
| 280 | or (requested_sources and "xiaohongshu" in requested_sources) |
| 281 | ) and env.is_xiaohongshu_available(config): |
| 282 | available.append("xiaohongshu") |
| 283 | # Threads: opt-in via INCLUDE_SOURCES (same pattern as perplexity/linkedin). |
| 284 | # Was auto-on with the key; gated so the onboarding "Everything" tier is a |
| 285 | # real choice vs the "Recommended" (TikTok/Instagram) tier. |
| 286 | if env.is_threads_available(config) and ( |
| 287 | "threads" in include_sources or (requested_sources and "threads" in requested_sources) |
| 288 | ): |
| 289 | available.append("threads") |
| 290 | # Pinterest: opt-in via INCLUDE_SOURCES. Previously read requested_sources |
| 291 | # only, so a persisted INCLUDE_SOURCES=pinterest never activated it; now it |
| 292 | # honors both the per-run --sources list and the saved config. |
| 293 | if env.is_pinterest_available(config) and ( |
| 294 | "pinterest" in include_sources or (requested_sources and "pinterest" in requested_sources) |
| 295 | ): |
| 296 | available.append("pinterest") |
| 297 | # xquik is a backend of the single "x" source (see env.x_backend_chain), |
| 298 | # not a separate parallel source — registered via the "x" entry above. |
| 299 | exclude = {s.strip().lower() for s in (config.get("EXCLUDE_SOURCES") or "").split(",") if s.strip()} |
| 300 | if exclude: |
| 301 | available = [s for s in available if s not in exclude] |
| 302 | return available |
| 303 | |
| 304 | |
| 305 | def _mock_discovery_items( |
| 306 | source: str, |
| 307 | domain: str, |
| 308 | to_date: str, |
| 309 | ) -> list[dict[str, Any]]: |
| 310 | """Deterministic listing fixtures for the public --mock CLI contract.""" |
| 311 | labels = [ |
| 312 | "Agent memory protocols", |
| 313 | "Browser-using agents", |
| 314 | "Local agent runtimes", |
| 315 | "Multi-agent orchestration", |
| 316 | "Agent security sandboxes", |
| 317 | "Voice agent latency", |
| 318 | ] |
| 319 | end = datetime.fromisoformat(to_date).date() |
| 320 | items: list[dict[str, Any]] = [] |
| 321 | for index, label in enumerate(labels, start=1): |
| 322 | published = (end - timedelta(days=index)).isoformat() |
| 323 | slug = re.sub(r"[^a-z0-9]+", "-", label.lower()).strip("-") |
| 324 | if source == "reddit": |
| 325 | items.append({ |
| 326 | "id": f"discovery-r-{index}", |
| 327 | "title": label, |
| 328 | "url": f"https://reddit.com/r/example/comments/{slug}", |
| 329 | "subreddit": "example", |
| 330 | "date": published, |
| 331 | "engagement": {"score": 180 - index * 10, "num_comments": 30 + index}, |
| 332 | "selftext": label, |
| 333 | "relevance": 0.9, |
| 334 | "why_relevant": "Mock discovery listing", |
| 335 | }) |
| 336 | elif source == "hackernews": |
| 337 | items.append({ |
| 338 | "id": f"discovery-hn-{index}", |
| 339 | "title": label, |
| 340 | "url": f"https://example.com/{slug}", |
| 341 | "hn_url": f"https://news.ycombinator.com/item?id={index}", |
| 342 | "author": f"example{index}", |
| 343 | "date": published, |
| 344 | "engagement": {"points": 120 - index * 8, "comments": 20 + index}, |
| 345 | "relevance": 0.88, |
| 346 | "why_relevant": "Mock HN discovery listing", |
| 347 | }) |
| 348 | elif source == "digg": |
| 349 | items.append({ |
| 350 | "id": f"discovery-d-{index}", |
| 351 | "title": label, |
| 352 | "url": f"https://di.gg/ai/{slug}", |
| 353 | "tldr": label, |
| 354 | "date": published, |
| 355 | "engagement": {"postCount": 30 - index, "uniqueAuthors": 12 - index}, |
| 356 | "relevance": 0.9, |
| 357 | "why_relevant": "Mock Digg discovery cluster", |
| 358 | }) |
| 359 | elif source == "x": |
| 360 | items.append({ |
| 361 | "id": f"discovery-x-{index}", |
| 362 | "text": label, |
| 363 | "url": f"https://x.com/example{index}/status/{index}", |
| 364 | "author_handle": f"example{index}", |
| 365 | "date": published, |
| 366 | "engagement": {"likes": 140 - index * 9, "reposts": 18 + index}, |
| 367 | "relevance": 0.9, |
| 368 | "why_relevant": "Mock X discovery activity", |
| 369 | }) |
| 370 | return items |
| 371 | |
| 372 | |
| 373 | def _matches_discovery_domain(domain: str, text: str) -> bool: |
| 374 | """Require a distinctive domain term, not a generic token such as ``AI``.""" |
| 375 | def terms(value: str) -> set[str]: |
| 376 | # Keep BOTH the surface form and the naive stem: replacing the token |
| 377 | # broke non-plurals ("bias" -> "bia", "crisis" -> "crisi") so in-domain |
| 378 | # listings stopped intersecting. The union preserves plural matching |
| 379 | # without corrupting the anchor. |
| 380 | words: set[str] = set() |
| 381 | for word in relevance.tokenize(value): |
| 382 | words.add(word) |
| 383 | if len(word) > 4 and word.endswith("s") and not word.endswith("ss"): |
| 384 | words.add(word[:-1]) |
| 385 | return words |
| 386 | |
| 387 | domain_terms = terms(domain) |
| 388 | anchors = domain_terms - _DISCOVERY_GENERIC_DOMAIN_TERMS |
| 389 | return bool((anchors or domain_terms) & terms(text)) |
| 390 | |
| 391 | |
| 392 | def _fetch_discovery_source( |
| 393 | source: str, |
| 394 | plan: schema.DiscoveryPlan, |
| 395 | *, |
| 396 | from_date: str, |
| 397 | to_date: str, |
| 398 | depth: str, |
| 399 | mock: bool, |
| 400 | config: dict[str, Any], |
| 401 | keyword_gate: bool = True, |
| 402 | ) -> tuple[list[dict[str, Any]], str | None]: |
| 403 | """Fetch one listing/river source for the nominate stage. |
| 404 | |
| 405 | ``keyword_gate`` controls whether items are filtered to the domain by |
| 406 | ``_matches_discovery_domain``. Domain-scoped discovery (``--discover X``) |
| 407 | keeps the gate on; global trending (``--discover`` with no domain) turns it |
| 408 | off, because there is no keyword to gate against - the river feeds ARE the |
| 409 | "what is hot right now" signal, and the confidence floor downstream is what |
| 410 | keeps junk out, not a keyword match. |
| 411 | """ |
| 412 | if mock: |
| 413 | return _mock_discovery_items(source, plan.domain, to_date), None |
| 414 | if source == "reddit": |
| 415 | result = reddit_listing.fetch_discovery_listings( |
| 416 | plan.subreddits, depth=depth, query=plan.domain, |
| 417 | ) |
| 418 | items = result.get("items") or [] |
| 419 | if keyword_gate: |
| 420 | items = [ |
| 421 | item for item in items |
| 422 | if _matches_discovery_domain( |
| 423 | plan.domain, |
| 424 | f"{item.get('title') or ''} {item.get('selftext') or ''}", |
| 425 | ) |
| 426 | ] |
| 427 | return items, "; ".join(result.get("errors") or []) or None |
| 428 | if source == "hackernews": |
| 429 | result = hackernews.fetch_discovery_listings(from_date, to_date, depth=depth) |
| 430 | items = result.get("items") or [] |
| 431 | for item in items: |
| 432 | item["relevance"] = relevance.token_overlap_relevance( |
| 433 | plan.domain, |
| 434 | str(item.get("title") or ""), |
| 435 | ) |
| 436 | # HN is a broad technology listing, so keep only domain-bearing stories |
| 437 | # when a domain is in play; global trending keeps the whole front page. |
| 438 | if keyword_gate: |
| 439 | items = [ |
| 440 | item for item in items |
| 441 | if _matches_discovery_domain(plan.domain, str(item.get("title") or "")) |
| 442 | ] |
| 443 | errors = result.get("errors") or [] |
| 444 | return items, "; ".join(errors) or None |
| 445 | if source == "digg": |
| 446 | result = digg.search_digg(plan.domain, from_date, to_date, depth=depth) |
| 447 | items = digg.parse_digg_response(result, query=plan.domain) |
| 448 | # Digg is an AI-focused broad listing, so keep only domain-bearing |
| 449 | # clusters when scoped; global trending keeps the whole feed. |
| 450 | if keyword_gate: |
| 451 | items = [ |
| 452 | item for item in items |
| 453 | if _matches_discovery_domain(plan.domain, str(item.get("title") or "")) |
| 454 | ] |
| 455 | return items, result.get("error") |
| 456 | if source == "x": |
| 457 | subquery = schema.SubQuery( |
| 458 | label="discovery-listings", |
| 459 | search_query=plan.domain, |
| 460 | ranking_query=f"What is accelerating in {plan.domain}?", |
| 461 | sources=["x"], |
| 462 | ) |
| 463 | last_error = "" |
| 464 | for backend in env.x_backend_chain(config): |
| 465 | items, error = _fetch_x_backend( |
| 466 | backend, subquery, from_date, to_date, depth, config, |
| 467 | ) |
| 468 | if items: |
| 469 | # Earlier failed-over backends' errors are observability, not |
| 470 | # degradation - but the producing backend's own error means |
| 471 | # these items are partial and must surface as such. |
| 472 | if last_error: |
| 473 | print(f"[x] earlier backend failed: {last_error}", file=sys.stderr) |
| 474 | return items, error or None |
| 475 | if error: |
| 476 | last_error = f"{backend}: {error}" |
| 477 | return [], last_error or None |
| 478 | raise ValueError(f"Unsupported discovery source: {source}") |
| 479 | |
| 480 | |
| 481 | def _discovery_engagement( |
| 482 | items: list[schema.SourceItem], |
| 483 | ) -> dict[str, dict[str, float | int]]: |
| 484 | totals: dict[str, dict[str, float | int]] = {} |
| 485 | for item in items: |
| 486 | bucket = totals.setdefault(item.source, {}) |
| 487 | for field, value in item.engagement.items(): |
| 488 | if not isinstance(value, (int, float)) or isinstance(value, bool): |
| 489 | continue |
| 490 | # Rank/score/reach metadata is not additive engagement: summing |
| 491 | # Digg ranks across items fabricates a metric (agent-export uses |
| 492 | # the same counter-field rule). |
| 493 | if not schema._is_counter_field(field): |
| 494 | continue |
| 495 | bucket[field] = bucket.get(field, 0) + value |
| 496 | return { |
| 497 | source: dict(sorted(metrics.items())) |
| 498 | for source, metrics in sorted(totals.items()) |
| 499 | } |
| 500 | |
| 501 | |
| 502 | def _discovery_momentum(items: list[schema.SourceItem], to_date: str) -> str: |
| 503 | as_of = datetime.fromisoformat(to_date).date() |
| 504 | ages: list[int] = [] |
| 505 | for item in items: |
| 506 | try: |
| 507 | published = datetime.fromisoformat((item.published_at or "").replace("Z", "+00:00")).date() |
| 508 | except (TypeError, ValueError): |
| 509 | continue |
| 510 | ages.append(max(0, (as_of - published).days)) |
| 511 | return "new-this-week" if ages and max(ages) < 7 else "building" |
| 512 | |
| 513 | |
| 514 | def nominate_candidates( |
| 515 | plan: schema.DiscoveryPlan, |
| 516 | *, |
| 517 | from_date: str, |
| 518 | to_date: str, |
| 519 | depth: str, |
| 520 | mock: bool, |
| 521 | config: dict[str, Any], |
| 522 | lookback_days: int, |
| 523 | keyword_gate: bool = True, |
| 524 | ) -> schema.RetrievalBundle: |
| 525 | """Stage 1 of discovery: fetch, normalize, and bundle candidate hot items |
| 526 | from the river/listing feeds. |
| 527 | |
| 528 | This is the topic-nomination pass. For domain discovery ``keyword_gate`` is |
| 529 | on and the feeds are filtered to the domain; for global trending it is off |
| 530 | and the feeds' own hot ranking IS the signal. The returned bundle feeds the |
| 531 | clustering + enrichment stages downstream. Every source's failure is |
| 532 | recorded on the bundle (never raised) so a single dead feed cannot sink the |
| 533 | run - the confidence floor decides whether the surviving evidence is enough. |
| 534 | """ |
| 535 | bundle = schema.RetrievalBundle() |
| 536 | with ThreadPoolExecutor(max_workers=max(1, len(plan.sources))) as executor: |
| 537 | futures = { |
| 538 | executor.submit( |
| 539 | _fetch_discovery_source, |
| 540 | source, |
| 541 | plan, |
| 542 | from_date=from_date, |
| 543 | to_date=to_date, |
| 544 | depth=depth, |
| 545 | mock=mock, |
| 546 | config=config, |
| 547 | keyword_gate=keyword_gate, |
| 548 | ): source |
| 549 | for source in plan.sources |
| 550 | } |
| 551 | for future in as_completed(futures): |
| 552 | source = futures[future] |
| 553 | bundle.mark_attempted(source) |
| 554 | try: |
| 555 | raw_items, partial_error = future.result() |
| 556 | normalized = normalize.normalize_source_items( |
| 557 | source, |
| 558 | raw_items, |
| 559 | from_date, |
| 560 | to_date, |
| 561 | freshness_mode="breaking", |
| 562 | ) |
| 563 | # Global trending has no domain; annotate against a neutral |
| 564 | # phrase so snippet extraction still works without biasing |
| 565 | # relevance toward any keyword. |
| 566 | prepared = relevance.PreparedQuery(plan.domain or "trending now") |
| 567 | normalized = signals.annotate_stream( |
| 568 | normalized, |
| 569 | prepared, |
| 570 | "breaking", |
| 571 | reference_date=to_date, |
| 572 | max_days=lookback_days, |
| 573 | ) |
| 574 | normalized = dedupe.dedupe_items(normalized) |
| 575 | for item in normalized: |
| 576 | item.snippet = snippet.extract_best_snippet(item, prepared) |
| 577 | bundle.add_items("discovery-listings", source, normalized) |
| 578 | if partial_error: |
| 579 | failure_state = ( |
| 580 | bird_x.classify_run_failure(partial_error) |
| 581 | if source == "x" and partial_error.startswith("bird:") |
| 582 | else http.classify_failure(message=partial_error) |
| 583 | ) |
| 584 | bundle.record_failure( |
| 585 | source, |
| 586 | failure_state, |
| 587 | partial_error, |
| 588 | ) |
| 589 | except Exception as exc: |
| 590 | state, attempted = _classify_source_failure(exc) |
| 591 | bundle.record_failure(source, state, str(exc), attempted=attempted) |
| 592 | return bundle |
| 593 | |
| 594 | |
| 595 | @dataclass(frozen=True) |
| 596 | class Nomination: |
| 597 | """A named candidate topic produced by the nominate stage. |
| 598 | |
| 599 | ``seed_score`` is the cheap pre-enrichment rank - seed velocity on the |
| 600 | nominate stage, blended with the HOST judge's content-worthiness on the |
| 601 | protocol resume leg (see ``rerank.judge_blended_score``). Enough to |
| 602 | decide WHICH candidates deserve a full pipeline pass, but not the final |
| 603 | ranking signal (that comes from enriched evidence downstream). |
| 604 | ``junk_shape`` flags help-me/beginner/musing shapes that should not |
| 605 | become content topics; ``worthiness`` is the host judge's 0-100 content |
| 606 | score, None on the heuristic path. |
| 607 | """ |
| 608 | |
| 609 | name: str |
| 610 | seed_score: float |
| 611 | items: list[schema.SourceItem] = field(default_factory=list) |
| 612 | summary: str = "" |
| 613 | junk_shape: bool = False |
| 614 | worthiness: float | None = None |
| 615 | |
| 616 | |
| 617 | def _cluster_entity_counts( |
| 618 | cluster: schema.Cluster, |
| 619 | candidate_map: dict[str, schema.Candidate], |
| 620 | ) -> Counter: |
| 621 | """Entity-token frequencies across a cluster's members (title + snippet).""" |
| 622 | counts: Counter = Counter() |
| 623 | for candidate_id in cluster.candidate_ids: |
| 624 | candidate = candidate_map.get(candidate_id) |
| 625 | if candidate: |
| 626 | counts.update(entity_extract.extract_text_entities( |
| 627 | f"{candidate.title} {candidate.snippet}" |
| 628 | )) |
| 629 | return counts |
| 630 | |
| 631 | |
| 632 | # Bound on how many distinguishing entity tokens a colliding cluster may try |
| 633 | # before it is treated as indistinguishable from the earlier story. Keeps a |
| 634 | # pathological cluster (dozens of unique tokens, every resulting name already |
| 635 | # taken) from scanning its whole vocabulary. |
| 636 | _DISAMBIGUATION_TOKEN_LIMIT = 5 |
| 637 | |
| 638 | |
| 639 | def _disambiguated_topic_name( |
| 640 | name: str, |
| 641 | cluster: schema.Cluster, |
| 642 | earlier_cluster: schema.Cluster, |
| 643 | candidate_map: dict[str, schema.Candidate], |
| 644 | entity_counts_cache: dict[str, Counter], |
| 645 | taken_names: dict[str, schema.Cluster], |
| 646 | ) -> str | None: |
| 647 | """Disambiguate a colliding topic name by appending the later cluster's |
| 648 | strongest entity token that the earlier cluster does not share. |
| 649 | |
| 650 | Distinguishing tokens are tried in descending strength order (bounded at |
| 651 | ``_DISAMBIGUATION_TOKEN_LIMIT``) and the first resulting name not already |
| 652 | present in ``taken_names`` (casefolded keys) wins: a first-choice suffix |
| 653 | colliding with an already-taken name must not drop a distinct story while |
| 654 | another distinguishing token remains. |
| 655 | |
| 656 | ``entity_counts_cache`` (keyed by cluster id, owned by the caller) memoizes |
| 657 | per-cluster entity counts so repeated collisions against the same cluster |
| 658 | never recompute them. |
| 659 | |
| 660 | Returns None when no distinguishing entity yields an unused name - the |
| 661 | clusters cannot be told apart by content, so the caller treats them as the |
| 662 | same story. |
| 663 | """ |
| 664 | def cached_counts(target: schema.Cluster) -> Counter: |
| 665 | counts = entity_counts_cache.get(target.cluster_id) |
| 666 | if counts is None: |
| 667 | counts = _cluster_entity_counts(target, candidate_map) |
| 668 | entity_counts_cache[target.cluster_id] = counts |
| 669 | return counts |
| 670 | |
| 671 | later_counts = cached_counts(cluster) |
| 672 | earlier_entities = set(cached_counts(earlier_cluster)) |
| 673 | name_tokens = {token.casefold() for token in name.split()} |
| 674 | choices = [ |
| 675 | (count, token) for token, count in later_counts.items() |
| 676 | if token not in earlier_entities and token.casefold() not in name_tokens |
| 677 | ] |
| 678 | # Strongest first = most frequent across the cluster; alphabetical |
| 679 | # tie-break keeps the result deterministic. |
| 680 | ranked = sorted(choices, key=lambda entry: (-entry[0], entry[1])) |
| 681 | for _, token in ranked[:_DISAMBIGUATION_TOKEN_LIMIT]: |
| 682 | display = token |
| 683 | for candidate_id in cluster.candidate_ids: |
| 684 | candidate = candidate_map.get(candidate_id) |
| 685 | if candidate is None: |
| 686 | continue |
| 687 | match = next( |
| 688 | ( |
| 689 | word.strip("\"'`()[]{}.,:;!?") |
| 690 | for word in f"{candidate.title} {candidate.snippet}".split() |
| 691 | if word.strip("\"'`()[]{}.,:;!?").lower() == token |
| 692 | ), |
| 693 | None, |
| 694 | ) |
| 695 | if match: |
| 696 | display = match |
| 697 | break |
| 698 | resolved = f"{name} {display}" |
| 699 | if resolved.casefold() not in taken_names: |
| 700 | return resolved |
| 701 | return None |
| 702 | |
| 703 | |
| 704 | def nominate_topic_pool( |
| 705 | bundle: schema.RetrievalBundle, |
| 706 | query_plan: schema.QueryPlan, |
| 707 | plan: schema.DiscoveryPlan, |
| 708 | *, |
| 709 | to_date: str, |
| 710 | limit: int, |
| 711 | ) -> list[tuple[Nomination, str]]: |
| 712 | """Stage 1b of discovery: cluster nominated items into named candidate |
| 713 | topics, rank them, and pair each with its source cluster id. |
| 714 | |
| 715 | This is the shared core behind ``nominate_topics`` (the one-shot path, |
| 716 | which drops the cluster ids) and the leg-1 nominate-only sweep (which |
| 717 | keys nominations-bundle rows on them, see ``run_discover_nominate``). |
| 718 | |
| 719 | Naming and junk classification are the deterministic ``topic_shape`` |
| 720 | heuristics and ranking is velocity-only - the engine runs no LLM here. |
| 721 | Reasoning-model judgment lives in the host-judged protocol: the host |
| 722 | renames, junk-filters, and worthiness-scores this pool from the leg-1 |
| 723 | bundle, and ``run_discover_resume`` applies those verdicts. The one-shot |
| 724 | path ships the heuristic names as-is. |
| 725 | |
| 726 | Casefold name collisions are disambiguated (the later cluster's strongest |
| 727 | non-shared entity token is appended, trying successive tokens when the |
| 728 | first-choice suffix is itself already taken) rather than blindly dropped: |
| 729 | short distilled names collide far more often than raw 96-char titles, and |
| 730 | a silent drop hides a distinct story. A colliding cluster is dropped only |
| 731 | when it shares a representative candidate with the earlier one (the same |
| 732 | story surfacing twice) or when no distinguishing entity token yields an |
| 733 | unused name. |
| 734 | |
| 735 | Returns at most ``limit`` ``(nomination, cluster_id)`` pairs, never |
| 736 | padded - fewer clusters than ``limit`` means a shorter list, and the |
| 737 | confidence floor downstream decides whether what survived is worth |
| 738 | showing. |
| 739 | """ |
| 740 | candidates = weighted_rrf(bundle.items_by_source_and_query, query_plan, pool_limit=80) |
| 741 | for candidate in candidates: |
| 742 | velocity = rerank.discovery_velocity_score(candidate.source_items, as_of_date=to_date) |
| 743 | candidate.final_score = min(100.0, 12.0 * math.log1p(velocity)) if velocity else 0.0 |
| 744 | candidates.sort(key=lambda candidate: (-candidate.final_score, candidate.title.lower())) |
| 745 | clusters = cluster_candidates(candidates, query_plan) |
| 746 | candidate_map = {candidate.candidate_id: candidate for candidate in candidates} |
| 747 | |
| 748 | ranked_clusters: list[tuple[float, schema.Cluster, list[schema.SourceItem]]] = [] |
| 749 | for cluster in clusters: |
| 750 | cluster_items: list[schema.SourceItem] = [] |
| 751 | for candidate_id in cluster.candidate_ids: |
| 752 | candidate = candidate_map.get(candidate_id) |
| 753 | if candidate: |
| 754 | cluster_items.extend(candidate.source_items) |
| 755 | score = rerank.discovery_velocity_score(cluster_items, as_of_date=to_date) |
| 756 | if score <= 0: |
| 757 | continue |
| 758 | ranked_clusters.append((score, cluster, cluster_items)) |
| 759 | ranked_clusters.sort(key=lambda entry: (-entry[0], entry[1].title.lower())) |
| 760 | |
| 761 | # Heuristic naming from each cluster's leader text (title + snippet). |
| 762 | named: list[tuple[float, schema.Cluster, list[schema.SourceItem], str, bool]] = [] |
| 763 | for score, cluster, cluster_items in ranked_clusters: |
| 764 | leader = candidate_map.get(cluster.representative_ids[0]) if cluster.representative_ids else None |
| 765 | title = (leader.title if leader else cluster.title) or "" |
| 766 | snip = (leader.snippet if leader else "") or "" |
| 767 | name = topic_shape.distill_topic_name(title, snip) or plan.domain or title |
| 768 | junk_shape = topic_shape.is_junk_shape(title, snip) |
| 769 | named.append((score, cluster, cluster_items, name, junk_shape)) |
| 770 | named.sort(key=lambda entry: (-entry[0], entry[3].lower())) |
| 771 | |
| 772 | pool: list[tuple[Nomination, str]] = [] |
| 773 | taken_names: dict[str, schema.Cluster] = {} |
| 774 | entity_counts_cache: dict[str, Counter] = {} |
| 775 | for score, cluster, cluster_items, name, junk_shape in named: |
| 776 | name_key = name.casefold() |
| 777 | if name_key in taken_names: |
| 778 | earlier_cluster = taken_names[name_key] |
| 779 | if set(cluster.representative_ids) & set(earlier_cluster.representative_ids): |
| 780 | continue # same story surfacing twice |
| 781 | resolved = _disambiguated_topic_name( |
| 782 | name, cluster, earlier_cluster, candidate_map, entity_counts_cache, |
| 783 | taken_names, |
| 784 | ) |
| 785 | if resolved is None: |
| 786 | continue # indistinguishable by content: treat as the same story |
| 787 | name = resolved |
| 788 | name_key = name.casefold() |
| 789 | taken_names[name_key] = cluster |
| 790 | leader = candidate_map.get(cluster.representative_ids[0]) if cluster.representative_ids else None |
| 791 | summary = (leader.snippet if leader else "") or (leader.title if leader else name) |
| 792 | pool.append((Nomination( |
| 793 | name=name, |
| 794 | seed_score=score, |
| 795 | items=cluster_items, |
| 796 | summary=summary, |
| 797 | junk_shape=junk_shape, |
| 798 | ), cluster.cluster_id)) |
| 799 | if len(pool) >= limit: |
| 800 | break |
| 801 | return pool |
| 802 | |
| 803 | |
| 804 | def nominate_topics( |
| 805 | bundle: schema.RetrievalBundle, |
| 806 | query_plan: schema.QueryPlan, |
| 807 | plan: schema.DiscoveryPlan, |
| 808 | *, |
| 809 | to_date: str, |
| 810 | limit: int, |
| 811 | ) -> list[Nomination]: |
| 812 | """``nominate_topic_pool`` without the cluster ids: the one-shot |
| 813 | discovery path's contract (see that function for the full semantics).""" |
| 814 | return [ |
| 815 | nomination |
| 816 | for nomination, _cluster_id in nominate_topic_pool( |
| 817 | bundle, query_plan, plan, to_date=to_date, limit=limit, |
| 818 | ) |
| 819 | ] |
| 820 | |
| 821 | |
| 822 | # Enrichment fan-out bounds. Sub-runs hit the same upstream APIs as a normal |
| 823 | # research pass, so parallelism stays low and the whole batch runs against a |
| 824 | # wall-clock budget - a slow topic is dropped, never fatal. |
| 825 | ENRICH_LIMIT = 6 |
| 826 | ENRICH_DEPTH = "quick" |
| 827 | ENRICH_MAX_WORKERS = 3 |
| 828 | ENRICH_BUDGET_SECONDS = 240.0 |
| 829 | |
| 830 | |
| 831 | @dataclass |
| 832 | class EnrichedTopic: |
| 833 | """A nomination plus the full-pipeline evidence gathered for it. |
| 834 | |
| 835 | ``report`` is None when enrichment for this topic failed or ran past the |
| 836 | batch budget - the topic survives as nomination-only and the confidence |
| 837 | floor downstream decides whether its seed evidence is enough to show. |
| 838 | """ |
| 839 | |
| 840 | nomination: Nomination |
| 841 | report: schema.Report | None = None |
| 842 | error: str | None = None |
| 843 | |
| 844 | |
| 845 | def enrich_nominations( |
| 846 | nominations: list[Nomination], |
| 847 | *, |
| 848 | config: dict[str, Any], |
| 849 | requested_sources: list[str] | None = None, |
| 850 | mock: bool = False, |
| 851 | depth: str = ENRICH_DEPTH, |
| 852 | lookback_days: int = 30, |
| 853 | as_of_date: str | None = None, |
| 854 | max_workers: int = ENRICH_MAX_WORKERS, |
| 855 | budget_seconds: float = ENRICH_BUDGET_SECONDS, |
| 856 | ) -> list[EnrichedTopic]: |
| 857 | """Stage 2 of discovery: run the real research pipeline on each nomination. |
| 858 | |
| 859 | Each nominated topic gets a full ``run()`` pass (``internal_subrun=True``, |
| 860 | same lane as comparison-mode sub-runs), which buys the whole multi-source |
| 861 | corpus - Reddit with comments, X, YouTube, Techmeme, arXiv, HN, Polymarket, |
| 862 | web - plus clustering and ranking, with zero bespoke fetch code. |
| 863 | |
| 864 | Failure containment: a topic whose sub-run raises is returned with |
| 865 | ``report=None`` and the error recorded; topics still unfinished when the |
| 866 | batch budget expires are likewise dropped to nomination-only. The batch |
| 867 | never raises and preserves nomination order. |
| 868 | """ |
| 869 | if not nominations: |
| 870 | return [] |
| 871 | |
| 872 | def _run_one(nomination: Nomination) -> schema.Report: |
| 873 | return run( |
| 874 | topic=nomination.name, |
| 875 | config=config, |
| 876 | depth=depth, |
| 877 | requested_sources=requested_sources, |
| 878 | mock=mock, |
| 879 | lookback_days=lookback_days, |
| 880 | as_of_date=as_of_date, |
| 881 | internal_subrun=True, |
| 882 | ) |
| 883 | |
| 884 | # Daemon threads + a semaphore instead of ThreadPoolExecutor: executor |
| 885 | # threads are non-daemon and joined at interpreter shutdown, so one hung |
| 886 | # sub-run could keep the whole process alive long after its topic was |
| 887 | # downgraded to nomination-only. Daemon workers make the wall-clock budget |
| 888 | # real - stragglers cannot delay process exit. Abandonment is safe because |
| 889 | # internal_subrun passes write nothing to disk (no save, no library sync, |
| 890 | # no store), and every fetch layer inside run() carries its own timeout. |
| 891 | youtube_yt.reset_search_cache() |
| 892 | enriched: dict[str, EnrichedTopic] = {} |
| 893 | results_queue: queue.Queue[tuple[Nomination, schema.Report | None, Exception | None]] = queue.Queue() |
| 894 | slots = threading.Semaphore(max(1, max_workers)) |
| 895 | |
| 896 | def _worker(nomination: Nomination) -> None: |
| 897 | with slots: |
| 898 | try: |
| 899 | results_queue.put((nomination, _run_one(nomination), None)) |
| 900 | except Exception as exc: # noqa: BLE001 - containment is the contract |
| 901 | results_queue.put((nomination, None, exc)) |
| 902 | |
| 903 | for nomination in nominations: |
| 904 | threading.Thread( |
| 905 | target=_worker, |
| 906 | args=(nomination,), |
| 907 | name=f"discover-enrich-{nomination.name[:32]}", |
| 908 | daemon=True, |
| 909 | ).start() |
| 910 | |
| 911 | deadline = time.monotonic() + max(1.0, budget_seconds) |
| 912 | pending = len(nominations) |
| 913 | while pending and (remaining := deadline - time.monotonic()) > 0: |
| 914 | try: |
| 915 | nomination, report, exc = results_queue.get(timeout=min(remaining, 0.5)) |
| 916 | except queue.Empty: |
| 917 | continue |
| 918 | pending -= 1 |
| 919 | if exc is None: |
| 920 | enriched[nomination.name] = EnrichedTopic( |
| 921 | nomination=nomination, report=report, |
| 922 | ) |
| 923 | else: |
| 924 | enriched[nomination.name] = EnrichedTopic( |
| 925 | nomination=nomination, |
| 926 | error=f"{type(exc).__name__}: {exc}", |
| 927 | ) |
| 928 | print( |
| 929 | f"[Discover] enrichment failed for {nomination.name!r}: " |
| 930 | f"{type(exc).__name__}: {exc}", |
| 931 | file=sys.stderr, |
| 932 | ) |
| 933 | # Budget expired (or all done): unfinished topics fall through below as |
| 934 | # nomination-only; their daemon workers are abandoned and cannot block exit. |
| 935 | |
| 936 | results: list[EnrichedTopic] = [] |
| 937 | for nomination in nominations: |
| 938 | entry = enriched.get(nomination.name) |
| 939 | if entry is None: |
| 940 | entry = EnrichedTopic( |
| 941 | nomination=nomination, |
| 942 | error="enrichment budget exhausted", |
| 943 | ) |
| 944 | print( |
| 945 | f"[Discover] enrichment budget exhausted before {nomination.name!r} " |
| 946 | "finished; keeping nomination-only evidence", |
| 947 | file=sys.stderr, |
| 948 | ) |
| 949 | results.append(entry) |
| 950 | return results |
| 951 | |
| 952 | |
| 953 | def _enriched_evidence_items(entry: EnrichedTopic) -> list[schema.SourceItem]: |
| 954 | """The items a topic is judged on: the enriched corpus when the pipeline |
| 955 | pass succeeded, the nomination's seed items otherwise.""" |
| 956 | if entry.report is not None: |
| 957 | flattened: list[schema.SourceItem] = [] |
| 958 | for source_items in entry.report.items_by_source.values(): |
| 959 | flattened.extend(source_items) |
| 960 | if flattened: |
| 961 | return flattened |
| 962 | return entry.nomination.items |
| 963 | |
| 964 | |
| 965 | def _best_community_comment(items: list[schema.SourceItem]) -> str | None: |
| 966 | """The strongest verbatim community comment across a topic's evidence, |
| 967 | formatted with attribution - the voice-of-the-people line on a trend card. |
| 968 | |
| 969 | Vote strength is per-platform-normalized (signals.normalized_comment_vote) |
| 970 | so one viral platform's counts don't drown out the rest. |
| 971 | """ |
| 972 | best: tuple[float, str, str | None, float | int | None] | None = None |
| 973 | for item in items: |
| 974 | comments = item.metadata.get("top_comments") or [] |
| 975 | for comment in comments: |
| 976 | if not isinstance(comment, dict): |
| 977 | continue |
| 978 | body = (comment.get("excerpt") or comment.get("text") or comment.get("body") or "").strip() |
| 979 | if len(body) < 12: |
| 980 | continue |
| 981 | strength = signals.normalized_comment_vote(item.source, comment.get("score")) |
| 982 | if best is None or strength > best[0]: |
| 983 | best = (strength, body, comment.get("author"), comment.get("score")) |
| 984 | if best is None: |
| 985 | return None |
| 986 | _, body, author, score = best |
| 987 | # Comment bodies that themselves start/end with quote characters would |
| 988 | # render as doubled quotes inside our wrapping quotes. |
| 989 | body = body.strip('"“”‘’\'').strip() |
| 990 | if len(body) > 200: |
| 991 | body = body[:197].rsplit(" ", 1)[0] + "..." |
| 992 | attribution = f" - {author}" if author else "" |
| 993 | votes = ( |
| 994 | f" ({int(score):,} votes)" |
| 995 | if isinstance(score, (int, float)) and not isinstance(score, bool) and score > 0 |
| 996 | else "" |
| 997 | ) |
| 998 | return f'"{body}"{attribution}{votes}' |
| 999 | |
| 1000 | |
| 1001 | @dataclass(frozen=True) |
| 1002 | class _DiscoverySweep: |
| 1003 | """The shared front half of both discovery entry points: the resolved |
| 1004 | plan and window, the swept listing bundle, and finalized per-source |
| 1005 | status. Everything downstream (judging, enrichment, floor, queue) |
| 1006 | belongs to the caller's leg.""" |
| 1007 | |
| 1008 | plan: schema.DiscoveryPlan |
| 1009 | query_plan: schema.QueryPlan |
| 1010 | from_date: str |
| 1011 | to_date: str |
| 1012 | bundle: schema.RetrievalBundle |
| 1013 | source_status: dict[str, schema.SourceOutcome] |
| 1014 | |
| 1015 | |
| 1016 | def _discovery_sweep( |
| 1017 | *, |
| 1018 | domain: str, |
| 1019 | config: dict[str, Any], |
| 1020 | depth: str, |
| 1021 | requested_sources: list[str] | None, |
| 1022 | mock: bool, |
| 1023 | subreddits: list[str] | None, |
| 1024 | lookback_days: int, |
| 1025 | as_of_date: str | None, |
| 1026 | ) -> _DiscoverySweep: |
| 1027 | """Resolve the momentum window, validate/bound the listing sources, build |
| 1028 | the discovery plan, sweep the river feeds, and finalize source status. |
| 1029 | |
| 1030 | Shared verbatim by ``run_discover`` (one-shot) and |
| 1031 | ``run_discover_nominate`` (protocol leg 1) so the two paths can never |
| 1032 | drift on what a sweep means.""" |
| 1033 | from_date, to_date = dates.get_date_range(lookback_days, as_of_date=as_of_date) |
| 1034 | requested = normalize_requested_sources(requested_sources) |
| 1035 | unsupported = sorted(set(requested or []) - set(DISCOVERY_SOURCES)) |
| 1036 | if unsupported: |
| 1037 | raise ValueError( |
| 1038 | "Discovery supports listing sources only: reddit, hackernews, digg " |
| 1039 | f"(unsupported: {', '.join(unsupported)})" |
| 1040 | ) |
| 1041 | available = list(DISCOVERY_SOURCES) if mock else [ |
| 1042 | source for source in available_sources(config, requested, x_pending=False) |
| 1043 | if source in DISCOVERY_SOURCES |
| 1044 | ] |
| 1045 | if requested: |
| 1046 | available = [source for source in available if source in requested] |
| 1047 | plan = planner.build_discovery_plan( |
| 1048 | domain, |
| 1049 | available_sources=available, |
| 1050 | subreddits=subreddits, |
| 1051 | ) |
| 1052 | |
| 1053 | global_mode = not plan.domain |
| 1054 | domain_label = plan.domain or "everything" |
| 1055 | query_plan = schema.QueryPlan( |
| 1056 | intent="breaking_news", |
| 1057 | freshness_mode="breaking", |
| 1058 | cluster_mode="story", |
| 1059 | raw_topic=plan.domain, |
| 1060 | subqueries=[schema.SubQuery( |
| 1061 | label="discovery-listings", |
| 1062 | search_query=plan.domain, |
| 1063 | ranking_query=f"What is accelerating in {domain_label}?", |
| 1064 | sources=list(plan.sources), |
| 1065 | )], |
| 1066 | source_weights={source: 1.0 for source in plan.sources}, |
| 1067 | notes=["discover-mode", "listing-sweep"], |
| 1068 | ) |
| 1069 | |
| 1070 | bundle = nominate_candidates( |
| 1071 | plan, |
| 1072 | from_date=from_date, |
| 1073 | to_date=to_date, |
| 1074 | depth=depth, |
| 1075 | mock=mock, |
| 1076 | config=config, |
| 1077 | lookback_days=lookback_days, |
| 1078 | # Global trending has no keyword to gate against - the river feeds' own |
| 1079 | # hot ranking is the signal and the confidence floor culls the junk. |
| 1080 | keyword_gate=not global_mode, |
| 1081 | ) |
| 1082 | |
| 1083 | source_status: dict[str, schema.SourceOutcome] = {} |
| 1084 | for source in DISCOVERY_SOURCES: |
| 1085 | if source in bundle.source_status: |
| 1086 | continue |
| 1087 | detail = ( |
| 1088 | "Source is not configured for discovery." |
| 1089 | ) |
| 1090 | source_status[source] = schema.SourceOutcome( |
| 1091 | source=source, |
| 1092 | state=schema.SKIPPED_UNCONFIGURED, |
| 1093 | attempted=False, |
| 1094 | detail=detail, |
| 1095 | fix_hint="doctor", |
| 1096 | ) |
| 1097 | source_status.update(_finalize_source_status(bundle.source_status, bundle.items_by_source)) |
| 1098 | return _DiscoverySweep( |
| 1099 | plan=plan, |
| 1100 | query_plan=query_plan, |
| 1101 | from_date=from_date, |
| 1102 | to_date=to_date, |
| 1103 | bundle=bundle, |
| 1104 | source_status=source_status, |
| 1105 | ) |
| 1106 | |
| 1107 | |
| 1108 | def _degraded_discovery_sources( |
| 1109 | source_status: dict[str, schema.SourceOutcome], |
| 1110 | ) -> list[str]: |
| 1111 | """Sources whose outcome is neither clean nor an expected skip.""" |
| 1112 | return [ |
| 1113 | source for source, outcome_state in source_status.items() |
| 1114 | if outcome_state.state not in {health.OK, schema.NO_RESULTS, schema.SKIPPED_UNCONFIGURED} |
| 1115 | ] |
| 1116 | |
| 1117 | |
| 1118 | @dataclass(frozen=True) |
| 1119 | class DiscoverNominateResult: |
| 1120 | """Leg 1 output of the host-judged discovery protocol: the ranked judge |
| 1121 | pool as ``(nomination, cluster_id)`` pairs plus the sweep context the CLI |
| 1122 | needs to write the nominations bundle - or to render the nothing-solid |
| 1123 | brief when the pool is empty.""" |
| 1124 | |
| 1125 | plan: schema.DiscoveryPlan |
| 1126 | from_date: str |
| 1127 | to_date: str |
| 1128 | source_status: dict[str, schema.SourceOutcome] |
| 1129 | pool: list[tuple[Nomination, str]] |
| 1130 | |
| 1131 | |
| 1132 | def run_discover_nominate( |
| 1133 | *, |
| 1134 | domain: str, |
| 1135 | config: dict[str, Any], |
| 1136 | depth: str = "default", |
| 1137 | requested_sources: list[str] | None = None, |
| 1138 | mock: bool = False, |
| 1139 | subreddits: list[str] | None = None, |
| 1140 | lookback_days: int = 30, |
| 1141 | as_of_date: str | None = None, |
| 1142 | ) -> DiscoverNominateResult: |
| 1143 | """Protocol leg 1: sweep the listings and build the FULL judge pool. |
| 1144 | |
| 1145 | Same sweep and clustering as ``run_discover``, but the pool is cut at |
| 1146 | ``rerank.JUDGE_POOL_LIMIT`` (not the enrichment limit). Like every |
| 1147 | discovery path it is deterministic-heuristic: no provider is ever |
| 1148 | resolved, so names and junk flags are the ``topic_shape`` baselines the |
| 1149 | host judges against. No enrichment, no confidence floor, no queue |
| 1150 | writes - those belong to legs 2 and 3. |
| 1151 | """ |
| 1152 | sweep = _discovery_sweep( |
| 1153 | domain=domain, |
| 1154 | config=config, |
| 1155 | depth=depth, |
| 1156 | requested_sources=requested_sources, |
| 1157 | mock=mock, |
| 1158 | subreddits=subreddits, |
| 1159 | lookback_days=lookback_days, |
| 1160 | as_of_date=as_of_date, |
| 1161 | ) |
| 1162 | pool = nominate_topic_pool( |
| 1163 | sweep.bundle, sweep.query_plan, sweep.plan, |
| 1164 | to_date=sweep.to_date, |
| 1165 | limit=rerank.JUDGE_POOL_LIMIT, |
| 1166 | ) |
| 1167 | return DiscoverNominateResult( |
| 1168 | plan=sweep.plan, |
| 1169 | from_date=sweep.from_date, |
| 1170 | to_date=sweep.to_date, |
| 1171 | source_status=sweep.source_status, |
| 1172 | pool=pool, |
| 1173 | ) |
| 1174 | |
| 1175 | |
| 1176 | def nominate_nothing_solid_report(result: DiscoverNominateResult) -> schema.DiscoveryReport: |
| 1177 | """The honest-empty leg-1 report: a zero-nomination sweep renders the |
| 1178 | same nothing-solid brief a one-shot run would (and writes no bundle).""" |
| 1179 | warnings = [ |
| 1180 | "The listing sweep nominated no topics this window; reporting " |
| 1181 | "nothing solid instead of ranked noise." |
| 1182 | ] |
| 1183 | failed = _degraded_discovery_sources(result.source_status) |
| 1184 | if failed: |
| 1185 | warnings.append(f"Some discovery sources degraded: {', '.join(sorted(failed))}.") |
| 1186 | return schema.DiscoveryReport( |
| 1187 | domain=result.plan.domain, |
| 1188 | range_from=result.from_date, |
| 1189 | range_to=result.to_date, |
| 1190 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 1191 | plan=result.plan, |
| 1192 | topics=[], |
| 1193 | source_status=result.source_status, |
| 1194 | warnings=warnings, |
| 1195 | outcome="nothing-solid", |
| 1196 | weak_signal=None, |
| 1197 | ) |
| 1198 | |
| 1199 | |
| 1200 | def _floor_survivor_records( |
| 1201 | enriched_entries: list[EnrichedTopic], |
| 1202 | *, |
| 1203 | to_date: str, |
| 1204 | topic_limit: int, |
| 1205 | ) -> tuple[ |
| 1206 | list[dict[str, Any]], |
| 1207 | tuple[float, str] | None, |
| 1208 | tuple[float, str] | None, |
| 1209 | ]: |
| 1210 | """Apply the discovery confidence floor to enriched entries in order, |
| 1211 | returning the survivor records plus the strongest non-junk and junk weak |
| 1212 | signals among the failures. |
| 1213 | |
| 1214 | Shared verbatim by ``run_discover`` (one-shot) and ``run_discover_resume`` |
| 1215 | (protocol leg 2) so floor semantics can never drift between the paths. |
| 1216 | """ |
| 1217 | survivors: list[dict[str, Any]] = [] |
| 1218 | weak_signal: tuple[float, str] | None = None |
| 1219 | junk_weak_signal: tuple[float, str] | None = None |
| 1220 | for entry in enriched_entries: |
| 1221 | nomination = entry.nomination |
| 1222 | evidence_items = _enriched_evidence_items(entry) |
| 1223 | sources = sorted({item.source for item in evidence_items}) |
| 1224 | native_total = sum( |
| 1225 | rerank.discovery_engagement_total(item) for item in evidence_items |
| 1226 | ) |
| 1227 | score = rerank.discovery_velocity_score(evidence_items, as_of_date=to_date) |
| 1228 | if not rerank.passes_discovery_floor( |
| 1229 | source_count=len(sources), |
| 1230 | engagement_total=native_total, |
| 1231 | item_count=len(evidence_items), |
| 1232 | junk_shape=nomination.junk_shape, |
| 1233 | # Junk corroboration counts distinct SEED listing sources, never |
| 1234 | # the enriched corpus - a successful enrichment pass is |
| 1235 | # multi-source for almost any topic, so it would never bind. |
| 1236 | seed_source_count=len({item.source for item in nomination.items}), |
| 1237 | ): |
| 1238 | # Sub-floor evidence never ranks; remember what came closest so a |
| 1239 | # nothing-solid brief can still name the strongest weak signal. |
| 1240 | # Junk-shaped failures are tracked separately: the brief prefers |
| 1241 | # the strongest NON-junk failure and names a junk one only when |
| 1242 | # every failure is junk-shaped (never empty when failures exist). |
| 1243 | if nomination.junk_shape: |
| 1244 | if junk_weak_signal is None or score > junk_weak_signal[0]: |
| 1245 | junk_weak_signal = (score, nomination.name) |
| 1246 | elif weak_signal is None or score > weak_signal[0]: |
| 1247 | weak_signal = (score, nomination.name) |
| 1248 | continue |
| 1249 | if len(survivors) >= topic_limit: |
| 1250 | break |
| 1251 | source_phrase = ", ".join(sources[:-1]) + ( |
| 1252 | f" and {sources[-1]}" if len(sources) > 1 else (sources[0] if sources else "the listings") |
| 1253 | ) |
| 1254 | noun = "evidence item" if entry.report is not None else "listing item" |
| 1255 | why = ( |
| 1256 | f"{len(evidence_items)} {noun}{'s' if len(evidence_items) != 1 else ''} on " |
| 1257 | f"{source_phrase} generated {native_total:,.0f} native interactions. " |
| 1258 | f"{nomination.summary[:220]}" |
| 1259 | ) |
| 1260 | top_comment = _best_community_comment(evidence_items) if entry.report is not None else None |
| 1261 | # Stage-2 angle input: the survivor's strongest evidence, enriched |
| 1262 | # corpus when the pipeline pass succeeded, seed items otherwise |
| 1263 | # (evidence_items already resolves that). |
| 1264 | top_titles = [ |
| 1265 | item.title.strip() |
| 1266 | for item in sorted( |
| 1267 | evidence_items, |
| 1268 | key=rerank.discovery_engagement_total, |
| 1269 | reverse=True, |
| 1270 | ) |
| 1271 | if item.title and item.title.strip() |
| 1272 | ][:3] |
| 1273 | survivors.append({ |
| 1274 | "name": nomination.name, |
| 1275 | "why": why, |
| 1276 | "momentum": _discovery_momentum(evidence_items, to_date), |
| 1277 | "velocity_score": round(score, 2), |
| 1278 | "sources": sources, |
| 1279 | "engagement_by_source": _discovery_engagement(evidence_items), |
| 1280 | "evidence_urls": list(dict.fromkeys(item.url for item in evidence_items if item.url))[:5], |
| 1281 | "top_comment": top_comment, |
| 1282 | "titles": "; ".join(top_titles), |
| 1283 | "engagement_phrase": f"{native_total:,.0f} native interactions across {source_phrase}", |
| 1284 | }) |
| 1285 | return survivors, weak_signal, junk_weak_signal |
| 1286 | |
| 1287 | |
| 1288 | def _fold_same_story_records(survivors: list[dict[str, Any]]) -> list[dict[str, Any]]: |
| 1289 | """Same-story fold + velocity ordering over floor-survivor records. |
| 1290 | |
| 1291 | Floor survivors that share enriched evidence are the SAME story wearing |
| 1292 | two judged names (the real-run failure: two topics quoting the identical |
| 1293 | 1,635-vote comment). Duplicates = identical non-None top comment OR >= 2 |
| 1294 | shared evidence URLs; the lower-velocity twin is dropped, and a winning |
| 1295 | replacement re-scans the kept list to a fixpoint so chained overlap |
| 1296 | (A~C~B) still collapses to one survivor. Selection stays seed-ordered |
| 1297 | upstream; this only prunes, then sorts by displayed velocity (stable) so |
| 1298 | rank 1 is the highest velocity_score. |
| 1299 | """ |
| 1300 | def _same_story(a: dict[str, Any], b: dict[str, Any]) -> bool: |
| 1301 | if a["top_comment"] is not None and a["top_comment"] == b["top_comment"]: |
| 1302 | return True |
| 1303 | return len(set(a["evidence_urls"]) & set(b["evidence_urls"])) >= 2 |
| 1304 | |
| 1305 | folded: list[dict[str, Any]] = [] |
| 1306 | for record in survivors: |
| 1307 | # Fold to a fixpoint: when the incoming record REPLACES a kept one, |
| 1308 | # the replacement may share evidence with entries the dropped record |
| 1309 | # never matched (three-way chains: A kept, C shares the comment with |
| 1310 | # A and URLs with B). The winner re-scans the remaining kept entries |
| 1311 | # until nothing matches, so one story always yields one survivor. |
| 1312 | incoming: dict[str, Any] | None = record |
| 1313 | while incoming is not None: |
| 1314 | dup_index = next( |
| 1315 | (index for index, kept in enumerate(folded) if _same_story(incoming, kept)), |
| 1316 | None, |
| 1317 | ) |
| 1318 | if dup_index is None: |
| 1319 | folded.append(incoming) |
| 1320 | break |
| 1321 | kept = folded[dup_index] |
| 1322 | if incoming["velocity_score"] > kept["velocity_score"]: |
| 1323 | folded.pop(dup_index) |
| 1324 | dropped_name, kept_name = kept["name"], incoming["name"] |
| 1325 | else: |
| 1326 | dropped_name, kept_name = incoming["name"], kept["name"] |
| 1327 | incoming = None # dropped; the kept entry stays in place |
| 1328 | log.source_log( |
| 1329 | "Discover", |
| 1330 | f"folded duplicate story {dropped_name!r} into {kept_name!r} (shared evidence)", |
| 1331 | tty_only=False, |
| 1332 | ) |
| 1333 | |
| 1334 | folded.sort(key=lambda record: record["velocity_score"], reverse=True) |
| 1335 | return folded |
| 1336 | |
| 1337 | |
| 1338 | def _records_to_discovery_topics( |
| 1339 | folded: list[dict[str, Any]], |
| 1340 | ) -> list[schema.DiscoveryTopic]: |
| 1341 | """Folded survivor records to ranked topics (ranks = 1-based positions).""" |
| 1342 | return [ |
| 1343 | schema.DiscoveryTopic( |
| 1344 | rank=position, |
| 1345 | name=record["name"], |
| 1346 | why_spiking=record["why"], |
| 1347 | momentum=record["momentum"], |
| 1348 | velocity_score=record["velocity_score"], |
| 1349 | sources=record["sources"], |
| 1350 | engagement_by_source=record["engagement_by_source"], |
| 1351 | command=f'/last30days "{record["name"].replace(chr(34), chr(39))}"', |
| 1352 | evidence_urls=record["evidence_urls"], |
| 1353 | top_comment=record["top_comment"], |
| 1354 | corroboration_count=len(record["sources"]), |
| 1355 | ) |
| 1356 | for position, record in enumerate(folded, start=1) |
| 1357 | ] |
| 1358 | |
| 1359 | |
| 1360 | def _discovery_report_warnings( |
| 1361 | topics: list[schema.DiscoveryTopic], |
| 1362 | outcome: str, |
| 1363 | source_status: dict[str, schema.SourceOutcome], |
| 1364 | ) -> list[str]: |
| 1365 | """Coverage warnings shared by the one-shot and resume discovery paths. |
| 1366 | The resume leg never re-sweeps: it passes the bundle's RESTORED leg-1 |
| 1367 | sweep status, so a degraded feed from the sweep still reaches the leg-2 |
| 1368 | report exactly as the one-shot reports it.""" |
| 1369 | warnings: list[str] = [] |
| 1370 | if outcome == "nothing-solid": |
| 1371 | warnings.append( |
| 1372 | "No topic cleared the discovery confidence floor this window; " |
| 1373 | "reporting nothing solid instead of ranked noise." |
| 1374 | ) |
| 1375 | elif len(topics) < 5: |
| 1376 | warnings.append("Fewer than five topic clusters cleared the confidence floor this window.") |
| 1377 | if topics and all(len(topic.sources) == 1 for topic in topics): |
| 1378 | warnings.append("Discovery evidence is single-source; configure Digg for broader confirmation.") |
| 1379 | failed = _degraded_discovery_sources(source_status) |
| 1380 | if failed: |
| 1381 | warnings.append(f"Some discovery sources degraded: {', '.join(sorted(failed))}.") |
| 1382 | return warnings |
| 1383 | |
| 1384 | |
| 1385 | def run_discover( |
| 1386 | *, |
| 1387 | domain: str, |
| 1388 | config: dict[str, Any], |
| 1389 | depth: str = "default", |
| 1390 | requested_sources: list[str] | None = None, |
| 1391 | mock: bool = False, |
| 1392 | subreddits: list[str] | None = None, |
| 1393 | lookback_days: int = 30, |
| 1394 | as_of_date: str | None = None, |
| 1395 | limit: int = 10, |
| 1396 | enrich: bool = False, |
| 1397 | enrich_requested_sources: list[str] | None = None, |
| 1398 | ) -> schema.DiscoveryReport: |
| 1399 | """Sweep category listings and rank the topics gaining velocity. |
| 1400 | |
| 1401 | ``requested_sources`` bounds the listing sweep (discovery-capable feeds |
| 1402 | only). ``enrich_requested_sources`` bounds the per-topic research passes: |
| 1403 | None means every available source - which is what lets Techmeme, arXiv, |
| 1404 | YouTube, Polymarket, and community comments reach discovery despite having |
| 1405 | no river feed of their own. Pass the user's original --search list here so |
| 1406 | an explicit source boundary holds through enrichment too. |
| 1407 | """ |
| 1408 | sweep = _discovery_sweep( |
| 1409 | domain=domain, |
| 1410 | config=config, |
| 1411 | depth=depth, |
| 1412 | requested_sources=requested_sources, |
| 1413 | mock=mock, |
| 1414 | subreddits=subreddits, |
| 1415 | lookback_days=lookback_days, |
| 1416 | as_of_date=as_of_date, |
| 1417 | ) |
| 1418 | plan = sweep.plan |
| 1419 | from_date, to_date = sweep.from_date, sweep.to_date |
| 1420 | source_status = sweep.source_status |
| 1421 | |
| 1422 | # The engine never names or angles topics with an LLM: the one-shot path |
| 1423 | # is deterministic-heuristic by design, and reasoning-model judgment |
| 1424 | # lives in the host-judged SKILL.md protocol. Say so loudly once per live |
| 1425 | # run; --mock stays silent (a deliberate mock run is not a degraded run). |
| 1426 | if not mock: |
| 1427 | log.source_log( |
| 1428 | "Discover", |
| 1429 | "one-shot run: topic names use deterministic heuristics and no " |
| 1430 | "content angles are generated - a reasoning-model host running " |
| 1431 | "the SKILL.md discovery protocol gets host-judged names, junk " |
| 1432 | "filtering, and podcast/X angles", |
| 1433 | tty_only=False, |
| 1434 | ) |
| 1435 | |
| 1436 | topic_limit = max(5, min(10, limit)) |
| 1437 | nominations = nominate_topics( |
| 1438 | sweep.bundle, sweep.query_plan, plan, |
| 1439 | to_date=to_date, |
| 1440 | limit=ENRICH_LIMIT if enrich else topic_limit, |
| 1441 | ) |
| 1442 | |
| 1443 | if enrich and nominations: |
| 1444 | enriched_entries = enrich_nominations( |
| 1445 | nominations, |
| 1446 | config=config, |
| 1447 | requested_sources=enrich_requested_sources, |
| 1448 | mock=mock, |
| 1449 | lookback_days=lookback_days, |
| 1450 | as_of_date=as_of_date, |
| 1451 | ) |
| 1452 | else: |
| 1453 | enriched_entries = [ |
| 1454 | EnrichedTopic(nomination=nomination) for nomination in nominations |
| 1455 | ] |
| 1456 | |
| 1457 | survivors, weak_signal, junk_weak_signal = _floor_survivor_records( |
| 1458 | enriched_entries, to_date=to_date, topic_limit=topic_limit, |
| 1459 | ) |
| 1460 | folded = _fold_same_story_records(survivors) |
| 1461 | # One-shot topics ship without angles (podcast_angle / x_article_angle |
| 1462 | # stay None and the renderer omits those lines): content angles are a |
| 1463 | # host-judged protocol deliverable, written on the finalize leg. |
| 1464 | topics = _records_to_discovery_topics(folded) |
| 1465 | |
| 1466 | if weak_signal is None: |
| 1467 | weak_signal = junk_weak_signal |
| 1468 | |
| 1469 | outcome = "ok" if topics else "nothing-solid" |
| 1470 | |
| 1471 | return schema.DiscoveryReport( |
| 1472 | domain=plan.domain, |
| 1473 | range_from=from_date, |
| 1474 | range_to=to_date, |
| 1475 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 1476 | plan=plan, |
| 1477 | topics=topics, |
| 1478 | source_status=source_status, |
| 1479 | warnings=_discovery_report_warnings(topics, outcome, source_status), |
| 1480 | outcome=outcome, |
| 1481 | weak_signal=weak_signal[1] if weak_signal and not topics else None, |
| 1482 | ) |
| 1483 | |
| 1484 | |
| 1485 | # Protocol leg 2 (resume) deep-tier enrichment bounds. The module-level |
| 1486 | # ENRICH_* constants above stay the one-shot --discover contract (quick depth, |
| 1487 | # 240s budget, 3 workers); a deep-tier bundle upgrades its per-topic sub-runs |
| 1488 | # to the default research depth with a wider wall-clock budget and one more |
| 1489 | # worker, because leg 2 is the protocol's only research pass. Shallow-tier |
| 1490 | # bundles keep the one-shot quick constants. Both tiers flow through |
| 1491 | # enrich_nominations' PARAMETERS - the constants themselves are never edited, |
| 1492 | # so neither tier can leak into the other path. |
| 1493 | RESUME_DEEP_ENRICH_DEPTH = "default" |
| 1494 | RESUME_DEEP_ENRICH_MAX_WORKERS = 4 |
| 1495 | RESUME_DEEP_ENRICH_BUDGET_SECONDS = 450.0 |
| 1496 | |
| 1497 | |
| 1498 | def _resume_enrich_budget_seconds(config: dict[str, Any]) -> float: |
| 1499 | """Deep-tier batch budget: LAST30DAYS_ENRICH_BUDGET_SECONDS from the |
| 1500 | RESOLVED config dict only (env.get_config already layers the process env |
| 1501 | over the .env files) - never read from bare os.environ. Blank, |
| 1502 | non-numeric, or non-positive values fall back to the 450s default.""" |
| 1503 | raw = config.get("LAST30DAYS_ENRICH_BUDGET_SECONDS") |
| 1504 | if raw is None or str(raw).strip() == "": |
| 1505 | return RESUME_DEEP_ENRICH_BUDGET_SECONDS |
| 1506 | try: |
| 1507 | value = float(raw) |
| 1508 | except (TypeError, ValueError): |
| 1509 | return RESUME_DEEP_ENRICH_BUDGET_SECONDS |
| 1510 | return value if value > 0 else RESUME_DEEP_ENRICH_BUDGET_SECONDS |
| 1511 | |
| 1512 | |
| 1513 | @dataclass(frozen=True) |
| 1514 | class DiscoverResumeResult: |
| 1515 | """Leg 2 output of the host-judged discovery protocol: the floored, |
| 1516 | folded, velocity-ranked report plus the per-topic angle inputs (keyed by |
| 1517 | surviving nomination id) that the host writes leg-3 angles from. |
| 1518 | ``report.source_status`` is the bundle's restored leg-1 sweep status - |
| 1519 | leg 2 never re-sweeps the listing feeds, so the sweep's degraded-coverage |
| 1520 | signal must survive the handoff instead of reading as clean.""" |
| 1521 | |
| 1522 | report: schema.DiscoveryReport |
| 1523 | angle_inputs: dict[str, dict[str, str]] |
| 1524 | |
| 1525 | |
| 1526 | def run_discover_resume( |
| 1527 | bundle: Any, |
| 1528 | judgments: dict[str, Any], |
| 1529 | *, |
| 1530 | config: dict[str, Any], |
| 1531 | mock: bool = False, |
| 1532 | ) -> DiscoverResumeResult: |
| 1533 | """Protocol leg 2: apply host judgments to the leg-1 bundle, enrich the |
| 1534 | slot winners, and floor/fold/rank on the same code path as the one-shot |
| 1535 | run. |
| 1536 | |
| 1537 | ``bundle`` is a ``discovery_handoff.NominationsBundle`` and ``judgments`` |
| 1538 | the mapping ``discovery_handoff.read_judgments`` returns (annotated |
| 1539 | loosely because discovery_handoff imports this module at load time). |
| 1540 | |
| 1541 | Judgment application is per field: an absent host name falls back to the |
| 1542 | bundle's heuristic name, an absent junk flag to the heuristic junk flag, |
| 1543 | and absent worthiness to the neutral blend default (None -> 50 inside |
| 1544 | ``rerank.judge_blended_score`` - the same treatment the judge-absent path |
| 1545 | always used). Applied names are collision-resolved over the whole pool |
| 1546 | before anything keys on them, and the applied name IS the enrichment |
| 1547 | sub-run topic. |
| 1548 | |
| 1549 | Slot selection: host-junk rows never contend for enrichment slots, and a |
| 1550 | heuristic-junk fallback row with fewer than ``rerank.FLOOR_MIN_SOURCES`` |
| 1551 | distinct seed sources is skipped pre-enrichment (it structurally cannot |
| 1552 | pass the floor's seed-corroboration rule). Both stay eligible to be the |
| 1553 | junk-tracked weak signal of a nothing-solid brief, and the brief prefers |
| 1554 | a non-junk weak signal exactly like the one-shot path. At the floor, |
| 1555 | host-judged rows pass ``junk_shape=False`` (host-junk never earned a |
| 1556 | slot) while heuristic-fallback rows keep their heuristic flag with the |
| 1557 | existing seed-source corroboration. |
| 1558 | |
| 1559 | Velocity, momentum, and the enrichment window all score against the |
| 1560 | bundle's momentum window (from_date/to_date), never the resume-time |
| 1561 | clock: the host may judge up to the handoff TTL after the sweep, and the |
| 1562 | numbers must describe the window the sweep captured. |
| 1563 | """ |
| 1564 | # Runtime-only import: discovery_handoff imports pipeline at module load, |
| 1565 | # so the reverse import must happen at call time (no import-time cycle). |
| 1566 | from . import discovery_handoff |
| 1567 | |
| 1568 | to_date = bundle.to_date |
| 1569 | verdicts = [ |
| 1570 | discovery_handoff.judgment_for(judgments, entry.nomination_id) |
| 1571 | for entry in bundle.nominations |
| 1572 | ] |
| 1573 | applied_names = discovery_handoff.resolve_name_collisions([ |
| 1574 | ( |
| 1575 | entry.nomination, |
| 1576 | verdict.name or entry.heuristic_name or entry.nomination.name, |
| 1577 | ) |
| 1578 | for entry, verdict in zip(bundle.nominations, verdicts) |
| 1579 | ]) |
| 1580 | |
| 1581 | ranked: list[tuple[float, str, Nomination]] = [] |
| 1582 | junk_weak_signal: tuple[float, str] | None = None |
| 1583 | for entry, verdict, name in zip(bundle.nominations, verdicts, applied_names): |
| 1584 | items = entry.nomination.items |
| 1585 | velocity = rerank.discovery_velocity_score(items, as_of_date=to_date) |
| 1586 | seed_source_count = len({item.source for item in items}) |
| 1587 | host_junk = verdict.junk is True |
| 1588 | fallback_junk = verdict.junk is None and entry.heuristic_junk |
| 1589 | if host_junk or ( |
| 1590 | fallback_junk and seed_source_count < rerank.FLOOR_MIN_SOURCES |
| 1591 | ): |
| 1592 | if junk_weak_signal is None or velocity > junk_weak_signal[0]: |
| 1593 | junk_weak_signal = (velocity, name) |
| 1594 | continue |
| 1595 | worthiness = ( |
| 1596 | float(verdict.worthiness) if verdict.worthiness is not None else None |
| 1597 | ) |
| 1598 | blended = rerank.judge_blended_score(velocity, worthiness) |
| 1599 | ranked.append(( |
| 1600 | blended, |
| 1601 | entry.nomination_id, |
| 1602 | replace( |
| 1603 | entry.nomination, |
| 1604 | name=name, |
| 1605 | seed_score=blended, |
| 1606 | junk_shape=( |
| 1607 | False if verdict.junk is not None else entry.heuristic_junk |
| 1608 | ), |
| 1609 | worthiness=worthiness, |
| 1610 | ), |
| 1611 | )) |
| 1612 | |
| 1613 | ranked.sort(key=lambda row: (-row[0], row[2].name.lower())) |
| 1614 | selected = ranked[:ENRICH_LIMIT] |
| 1615 | nominations = [nomination for _blended, _nomination_id, nomination in selected] |
| 1616 | |
| 1617 | if bundle.tier == "shallow": |
| 1618 | depth, max_workers, budget_seconds = ( |
| 1619 | ENRICH_DEPTH, ENRICH_MAX_WORKERS, ENRICH_BUDGET_SECONDS, |
| 1620 | ) |
| 1621 | else: |
| 1622 | depth = RESUME_DEEP_ENRICH_DEPTH |
| 1623 | max_workers = RESUME_DEEP_ENRICH_MAX_WORKERS |
| 1624 | budget_seconds = _resume_enrich_budget_seconds(config) |
| 1625 | |
| 1626 | enriched_entries = enrich_nominations( |
| 1627 | nominations, |
| 1628 | config=config, |
| 1629 | requested_sources=bundle.enrichment_source_boundary, |
| 1630 | mock=mock, |
| 1631 | depth=depth, |
| 1632 | lookback_days=bundle.lookback_days, |
| 1633 | as_of_date=to_date, |
| 1634 | max_workers=max_workers, |
| 1635 | budget_seconds=budget_seconds, |
| 1636 | ) if nominations else [] |
| 1637 | |
| 1638 | # topic_limit mirrors the one-shot default cap (limit=10); the slot cut |
| 1639 | # above already bounds the pool at ENRICH_LIMIT. |
| 1640 | survivors, weak_signal, floor_junk_weak_signal = _floor_survivor_records( |
| 1641 | enriched_entries, to_date=to_date, topic_limit=10, |
| 1642 | ) |
| 1643 | if floor_junk_weak_signal is not None and ( |
| 1644 | junk_weak_signal is None |
| 1645 | or floor_junk_weak_signal[0] > junk_weak_signal[0] |
| 1646 | ): |
| 1647 | junk_weak_signal = floor_junk_weak_signal |
| 1648 | folded = _fold_same_story_records(survivors) |
| 1649 | topics = _records_to_discovery_topics(folded) |
| 1650 | |
| 1651 | nomination_id_by_name = { |
| 1652 | nomination.name: nomination_id |
| 1653 | for _blended, nomination_id, nomination in selected |
| 1654 | } |
| 1655 | angle_inputs = { |
| 1656 | nomination_id_by_name[record["name"]]: { |
| 1657 | "name": record["name"], |
| 1658 | "titles": record["titles"], |
| 1659 | "top_comment": record["top_comment"] or "", |
| 1660 | "engagement": record["engagement_phrase"], |
| 1661 | } |
| 1662 | for record in folded |
| 1663 | } |
| 1664 | |
| 1665 | if weak_signal is None: |
| 1666 | weak_signal = junk_weak_signal |
| 1667 | outcome = "ok" if topics else "nothing-solid" |
| 1668 | plan = schema.DiscoveryPlan( |
| 1669 | domain=bundle.domain, |
| 1670 | category=None, |
| 1671 | subreddits=[], |
| 1672 | sources=( |
| 1673 | list(bundle.requested_sources) |
| 1674 | if bundle.requested_sources |
| 1675 | else sorted({ |
| 1676 | item.source |
| 1677 | for entry in bundle.nominations |
| 1678 | for item in entry.nomination.items |
| 1679 | }) |
| 1680 | ), |
| 1681 | ) |
| 1682 | # The bundle's restored leg-1 sweep status (empty for pre-field bundles): |
| 1683 | # degraded sweep coverage must reach this report's status map and its |
| 1684 | # degraded-sources warning exactly as the one-shot reports it. |
| 1685 | source_status = dict(getattr(bundle, "source_status", None) or {}) |
| 1686 | report = schema.DiscoveryReport( |
| 1687 | domain=bundle.domain, |
| 1688 | range_from=bundle.from_date, |
| 1689 | range_to=to_date, |
| 1690 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 1691 | plan=plan, |
| 1692 | topics=topics, |
| 1693 | source_status=source_status, |
| 1694 | warnings=_discovery_report_warnings(topics, outcome, source_status), |
| 1695 | outcome=outcome, |
| 1696 | weak_signal=weak_signal[1] if weak_signal and not topics else None, |
| 1697 | ) |
| 1698 | return DiscoverResumeResult(report=report, angle_inputs=angle_inputs) |
| 1699 | |
| 1700 | |
| 1701 | def diagnose( |
| 1702 | config: dict[str, Any], |
| 1703 | requested_sources: list[str] | None = None, |
| 1704 | *, |
| 1705 | safe: bool = False, |
| 1706 | ) -> dict[str, Any]: |
| 1707 | requested_sources = normalize_requested_sources(requested_sources) |
| 1708 | google_key = _google_key(config) |
| 1709 | x_status = env.get_x_source_status(config, probe=not safe) |
| 1710 | # Compute once and reuse for both the diag flag and available_sources below. |
| 1711 | # safe=True (doctor/--diagnose/--preflight) must stay network-free. |
| 1712 | x_pending = env.x_pending_browser_auth(config, local_only=safe) |
| 1713 | native_web_backend = None |
| 1714 | if config.get("BRAVE_API_KEY"): |
| 1715 | native_web_backend = "brave" |
| 1716 | elif config.get("EXA_API_KEY"): |
| 1717 | native_web_backend = "exa" |
| 1718 | elif config.get("SERPER_API_KEY"): |
| 1719 | native_web_backend = "serper" |
| 1720 | elif config.get("PARALLEL_API_KEY"): |
| 1721 | native_web_backend = "parallel" |
| 1722 | providers_status = { |
| 1723 | "google": bool(google_key), |
| 1724 | "openai": bool(config.get("OPENAI_API_KEY")) and config.get("OPENAI_AUTH_STATUS") == env.AUTH_STATUS_OK, |
| 1725 | "xai": bool(config.get("XAI_API_KEY")), |
| 1726 | "openrouter": bool(config.get("OPENROUTER_API_KEY")), |
| 1727 | "perplexity": bool(config.get("PERPLEXITY_API_KEY")), |
| 1728 | } |
| 1729 | reasoning_provider_available = any( |
| 1730 | providers_status[name] for name in ("google", "openai", "xai", "openrouter") |
| 1731 | ) |
| 1732 | external_commands = { |
| 1733 | "yt-dlp": bool(which("yt-dlp")), |
| 1734 | "digg-pp-cli": bool(which("digg-pp-cli")), |
| 1735 | "arxiv-pp-cli": bool(which("arxiv-pp-cli")), |
| 1736 | "techmeme-pp-cli": bool(which("techmeme-pp-cli")), |
| 1737 | "trustpilot-pp-cli": bool(which("trustpilot-pp-cli")), |
| 1738 | "gh": bool(which("gh")), |
| 1739 | } |
| 1740 | credential_destinations = { |
| 1741 | "global_env": str(env.CONFIG_FILE) if env.CONFIG_FILE else None, |
| 1742 | } |
| 1743 | browser_cookies = { |
| 1744 | "mode": config.get("_BROWSER_COOKIE_MODE", "off"), |
| 1745 | "browsers": list(config.get("_BROWSER_COOKIE_BROWSERS") or []), |
| 1746 | "reads_values": False if safe else config.get("_BROWSER_COOKIE_MODE") == "read", |
| 1747 | } |
| 1748 | ignored_project_keys = list(config.get("_IGNORED_PROJECT_CONFIG_KEYS") or []) |
| 1749 | ignored_endpoint_overrides = [ |
| 1750 | key for key in ignored_project_keys if key in permission_preflight.ENDPOINT_OVERRIDE_KEYS |
| 1751 | ] |
| 1752 | local_writes: list[dict[str, str]] = [] |
| 1753 | if config.get("LAST30DAYS_MEMORY_DIR"): |
| 1754 | local_writes.append({"kind": "report", "path": str(config.get("LAST30DAYS_MEMORY_DIR"))}) |
| 1755 | diag = { |
| 1756 | "providers": providers_status, |
| 1757 | "local_mode": not reasoning_provider_available, |
| 1758 | "reasoning_provider": (config.get("LAST30DAYS_REASONING_PROVIDER") or "auto").lower(), |
| 1759 | "x_backend": x_status["source"], |
| 1760 | "bird_installed": x_status["bird_installed"], |
| 1761 | "bird_authenticated": x_status["bird_authenticated"], |
| 1762 | "bird_username": x_status["bird_username"], |
| 1763 | "x_pending_browser_auth": x_pending, |
| 1764 | "xquik_available": x_status.get("xquik_available", False), |
| 1765 | "xquik_working": x_status.get("xquik_working"), |
| 1766 | "xquik_status": x_status.get("xquik_status", ""), |
| 1767 | "native_web_backend": native_web_backend, |
| 1768 | "native_search": env.is_native_search(config), |
| 1769 | "has_scrapecreators": bool(config.get("SCRAPECREATORS_API_KEY")), |
| 1770 | "has_github": bool(config.get("GITHUB_TOKEN") or which("gh")), |
| 1771 | # safe=True (doctor/--diagnose/--preflight) must stay network-free: |
| 1772 | # answer X availability from local evidence only. x_pending is |
| 1773 | # precomputed by diagnose() to avoid double evaluation. |
| 1774 | "available_sources": available_sources( |
| 1775 | config, requested_sources, x_pending=x_pending, local_only=safe |
| 1776 | ), |
| 1777 | "safe": safe, |
| 1778 | "config_source": config.get("_CONFIG_SOURCE"), |
| 1779 | "ignored_project_config": config.get("_IGNORED_PROJECT_CONFIG"), |
| 1780 | "ignored_project_config_keys": ignored_project_keys, |
| 1781 | "ignored_endpoint_overrides": ignored_endpoint_overrides, |
| 1782 | "browser_cookies": browser_cookies, |
| 1783 | "external_commands": external_commands, |
| 1784 | "credential_destinations": credential_destinations, |
| 1785 | "local_writes": local_writes, |
| 1786 | } |
| 1787 | diag["permission_preflight"] = permission_preflight.build(config, diag) |
| 1788 | return diag |
| 1789 | |
| 1790 | |
| 1791 | def _inner_max_workers(stream_count: int, *, internal_subrun: bool) -> int: |
| 1792 | """Worker-pool size for the per-stream fanout inside a single pipeline run. |
| 1793 | |
| 1794 | Top-level runs use up to 16 workers. Subruns of ``run_competitor_fanout`` |
| 1795 | cap the inner pool to 4 so a six-way competitor fan-out stays below |
| 1796 | roughly 30 worker threads in aggregate instead of ~96. |
| 1797 | """ |
| 1798 | if internal_subrun: |
| 1799 | return max(2, min(4, stream_count or 1)) |
| 1800 | return max(4, min(16, stream_count or 1)) |
| 1801 | |
| 1802 | |
| 1803 | def _load_library_context( |
| 1804 | *, |
| 1805 | topic: str, |
| 1806 | config: dict[str, Any], |
| 1807 | mock: bool, |
| 1808 | internal_subrun: bool, |
| 1809 | x_handle: str | None, |
| 1810 | github_user: str | None, |
| 1811 | github_repos: list[str] | None, |
| 1812 | save_dir: Path | str | None = None, |
| 1813 | ) -> tuple[list[schema.LibraryContext], str | None]: |
| 1814 | """Resolve compact prior-run context without making a research run depend on it.""" |
| 1815 | setting = str(config.get("LAST30DAYS_LIBRARY_CONTEXT") or "off").strip().lower() |
| 1816 | if mock or internal_subrun or setting in {"0", "false", "no", "off"}: |
| 1817 | return [], None |
| 1818 | if save_dir == "": |
| 1819 | return [], None |
| 1820 | |
| 1821 | memory_dir = ( |
| 1822 | save_dir |
| 1823 | if save_dir is not None |
| 1824 | else config.get("LAST30DAYS_MEMORY_DIR") or library.DEFAULT_MEMORY_DIR |
| 1825 | ) |
| 1826 | briefs_dir = config.get("_LAST30DAYS_LIBRARY_BRIEFS_DIR") or ( |
| 1827 | Path(memory_dir).expanduser() / "briefings" |
| 1828 | if save_dir is not None |
| 1829 | else library.DEFAULT_BRIEFS_DIR |
| 1830 | ) |
| 1831 | db_path = config.get("_LAST30DAYS_LIBRARY_DB") |
| 1832 | if not db_path: |
| 1833 | db_path = ( |
| 1834 | Path(memory_dir).expanduser().resolve() / ".last30days-library.db" |
| 1835 | if save_dir is not None |
| 1836 | else library_index.DEFAULT_LIBRARY_DB |
| 1837 | ) |
| 1838 | store_db = config.get("_LAST30DAYS_STORE_DB") |
| 1839 | if not store_db: |
| 1840 | # Scoped runs read only a store inside the save dir (usually absent); |
| 1841 | # the shared store would leak other scopes' sightings into this one. |
| 1842 | store_db = ( |
| 1843 | Path(memory_dir).expanduser().resolve() / "research.db" |
| 1844 | if save_dir is not None |
| 1845 | else library_index.DEFAULT_STORE_DB |
| 1846 | ) |
| 1847 | queries = [topic, x_handle or "", github_user or "", *(github_repos or [])] |
| 1848 | queries = list(dict.fromkeys(value.strip() for value in queries if value and value.strip())) |
| 1849 | try: |
| 1850 | library_index.sync_library(memory_dir, briefs_dir, db_path=db_path) |
| 1851 | matches: list[library_index.LibrarySearchMatch] = [] |
| 1852 | for query_text in queries: |
| 1853 | matches.extend( |
| 1854 | library_index.search( |
| 1855 | query_text, |
| 1856 | limit=6, |
| 1857 | db_path=db_path, |
| 1858 | store_db_path=store_db, |
| 1859 | ) |
| 1860 | ) |
| 1861 | except (library_index.LibrarySearchUnavailable, OSError, sqlite3.DatabaseError) as exc: |
| 1862 | return [], f"Library context unavailable: {exc}" |
| 1863 | |
| 1864 | contexts: list[schema.LibraryContext] = [] |
| 1865 | seen_runs: set[tuple[str, date]] = set() |
| 1866 | for match in sorted( |
| 1867 | matches, |
| 1868 | key=lambda item: (-item.published_date.toordinal(), item.rank, item.topic.casefold()), |
| 1869 | ): |
| 1870 | if match.run_key in seen_runs: |
| 1871 | continue |
| 1872 | seen_runs.add(match.run_key) |
| 1873 | contexts.append( |
| 1874 | schema.LibraryContext( |
| 1875 | topic=match.topic, |
| 1876 | published_date=match.published_date.isoformat(), |
| 1877 | headline=match.headline, |
| 1878 | summary=match.snippet or match.headline, |
| 1879 | source_kind=match.source_kind, |
| 1880 | ) |
| 1881 | ) |
| 1882 | if len(contexts) == 3: |
| 1883 | break |
| 1884 | return contexts, None |
| 1885 | |
| 1886 | |
| 1887 | def run( |
| 1888 | *, |
| 1889 | topic: str, |
| 1890 | config: dict[str, Any], |
| 1891 | depth: str, |
| 1892 | requested_sources: list[str] | None = None, |
| 1893 | mock: bool = False, |
| 1894 | x_handle: str | None = None, |
| 1895 | x_related: list[str] | None = None, |
| 1896 | web_backend: str = "auto", |
| 1897 | external_plan: dict | None = None, |
| 1898 | subreddits: list[str] | None = None, |
| 1899 | tiktok_hashtags: list[str] | None = None, |
| 1900 | tiktok_creators: list[str] | None = None, |
| 1901 | ig_creators: list[str] | None = None, |
| 1902 | lookback_days: int = 30, |
| 1903 | as_of_date: str | None = None, |
| 1904 | github_user: str | None = None, |
| 1905 | github_repos: list[str] | None = None, |
| 1906 | trustpilot_domain: str | None = None, |
| 1907 | trustpilot_domain_is_hint: bool = False, |
| 1908 | hiring_signals_mode: bool = False, |
| 1909 | internal_subrun: bool = False, |
| 1910 | save_dir: Path | str | None = None, |
| 1911 | corpus_dirs: list[str] | None = None, |
| 1912 | corpus_all_time: bool = False, |
| 1913 | ) -> schema.Report: |
| 1914 | # Standalone runs (not competitor/discover sub-runs) own the YouTube |
| 1915 | # search-cache lifecycle. Comparison fan-out clears once before submit so |
| 1916 | # parallel entity sub-runs can still share in-run hits. |
| 1917 | if not internal_subrun: |
| 1918 | youtube_yt.reset_search_cache() |
| 1919 | settings = _resolve_depth_settings(depth, config) |
| 1920 | requested_sources = normalize_requested_sources(requested_sources) |
| 1921 | from_date, to_date = dates.get_date_range(lookback_days, as_of_date=as_of_date) |
| 1922 | resolved_corpus_dirs = corpus.resolve_directories( |
| 1923 | corpus_dirs or config.get("_CORPUS_DIRS"), |
| 1924 | config.get("LAST30DAYS_CORPUS_DIRS"), |
| 1925 | ) |
| 1926 | excluded_sources = { |
| 1927 | source.strip().lower() |
| 1928 | for source in str(config.get("EXCLUDE_SOURCES") or "").split(",") |
| 1929 | if source.strip() |
| 1930 | } |
| 1931 | corpus_enabled = bool(resolved_corpus_dirs) and "corpus" not in excluded_sources |
| 1932 | corpus_requested = bool(requested_sources and "corpus" in requested_sources) |
| 1933 | if corpus_enabled and requested_sources and "corpus" not in requested_sources: |
| 1934 | requested_sources = [*requested_sources, "corpus"] |
| 1935 | |
| 1936 | # Gate StockTwits to ticker/crypto topics. Single chokepoint: when False, |
| 1937 | # available_sources() never registers stocktwits, so the planner can't |
| 1938 | # assign it (eligible_sources = available ∩ capabilities). |
| 1939 | config["_financial_topic"] = stocktwits.is_financial_topic(topic) |
| 1940 | |
| 1941 | if mock: |
| 1942 | runtime = providers.mock_runtime(config, depth) |
| 1943 | reasoning_provider = None |
| 1944 | available = list(requested_sources or MOCK_AVAILABLE_SOURCES) |
| 1945 | if corpus_enabled and "corpus" not in available: |
| 1946 | available.append("corpus") |
| 1947 | if not corpus_enabled and not corpus_requested: |
| 1948 | available = [source for source in available if source != "corpus"] |
| 1949 | if not requested_sources and not hiring_signals_mode and not _company_topic_likely(topic): |
| 1950 | available = [source for source in available if source != "jobs"] |
| 1951 | else: |
| 1952 | runtime, reasoning_provider = providers.resolve_runtime(config, depth) |
| 1953 | available = available_sources(config, requested_sources) |
| 1954 | if requested_sources: |
| 1955 | available = [source for source in available if source in requested_sources] |
| 1956 | # Keep an explicitly requested but unconfigured corpus in the plan long |
| 1957 | # enough to record its skipped-unconfigured source outcome. It is never |
| 1958 | # submitted to the network executor below. |
| 1959 | if corpus_requested and "corpus" not in excluded_sources and "corpus" not in available: |
| 1960 | available.append("corpus") |
| 1961 | if web_backend == "none": |
| 1962 | available = [s for s in available if s != "grounding"] |
| 1963 | elif web_backend in ("brave", "exa", "serper", "parallel", "keyless") and "grounding" not in available: |
| 1964 | available.append("grounding") |
| 1965 | if ( |
| 1966 | hiring_signals_mode |
| 1967 | or (not requested_sources and _company_topic_likely(topic)) |
| 1968 | ) and "jobs" not in available: |
| 1969 | available.append("jobs") |
| 1970 | if hiring_signals_mode: |
| 1971 | config = dict(config) |
| 1972 | config["_hiring_signals_mode"] = True |
| 1973 | if not requested_sources: |
| 1974 | available = ["jobs"] |
| 1975 | if not available: |
| 1976 | raise RuntimeError("No sources are available for this run.") |
| 1977 | |
| 1978 | planner_requested_sources = requested_sources |
| 1979 | if hiring_signals_mode and not planner_requested_sources: |
| 1980 | planner_requested_sources = ["jobs"] |
| 1981 | |
| 1982 | if external_plan is not None: |
| 1983 | # External plan provided (e.g., from Claude Code via --plan flag). |
| 1984 | # Explicit input is a contract: validate it before permissive sanitization. |
| 1985 | planner.validate_external_plan(external_plan) |
| 1986 | plan = planner._sanitize_plan( |
| 1987 | external_plan, topic, available, planner_requested_sources, depth, |
| 1988 | ) |
| 1989 | plan_source = "external" |
| 1990 | else: |
| 1991 | plan = planner.plan_query( |
| 1992 | topic=topic, |
| 1993 | available_sources=available, |
| 1994 | requested_sources=planner_requested_sources, |
| 1995 | depth=depth, |
| 1996 | provider=None if mock else reasoning_provider, |
| 1997 | model=None if mock else runtime.planner_model, |
| 1998 | context=config.get("_auto_resolve_context", ""), |
| 1999 | internal_subrun=internal_subrun, |
| 2000 | ) |
| 2001 | # Source labelling: the fallback path annotates notes with "fallback-plan" |
| 2002 | # or "deterministic-comparison-plan"; anything else came from the LLM. |
| 2003 | if any("fallback" in note or "deterministic" in note for note in (plan.notes or [])): |
| 2004 | plan_source = "deterministic" |
| 2005 | elif not mock and reasoning_provider and runtime.planner_model: |
| 2006 | plan_source = "llm" |
| 2007 | else: |
| 2008 | plan_source = "deterministic" |
| 2009 | |
| 2010 | # Safety net: ensure grounding appears in all subqueries even if the planner |
| 2011 | # omits it. This is redundant when the planner includes grounding via |
| 2012 | # SOURCE_CAPABILITIES, but kept as a fallback. |
| 2013 | if ( |
| 2014 | web_backend != "none" |
| 2015 | and "grounding" in available |
| 2016 | and "drill-mode" not in plan.notes |
| 2017 | ): |
| 2018 | for sq in plan.subqueries: |
| 2019 | if "grounding" not in sq.sources: |
| 2020 | sq.sources.append("grounding") |
| 2021 | if "drill-mode" not in plan.notes: |
| 2022 | # Drill plans re-fetch only the sources that contributed to the matched |
| 2023 | # cluster; the company-topic jobs injection must not widen that set. |
| 2024 | _ensure_jobs_in_plan(plan, available, explicit=hiring_signals_mode, topic=topic) |
| 2025 | if "corpus" in available and plan.subqueries: |
| 2026 | # Corpus is deterministic and user-registered, so it always gets one |
| 2027 | # bounded stream even when a quick/LLM plan omits it. Reuse the primary |
| 2028 | # subquery instead of multiplying local scans across every subquery. |
| 2029 | if "corpus" not in plan.subqueries[0].sources: |
| 2030 | plan.subqueries[0].sources.append("corpus") |
| 2031 | if "corpus" not in plan.source_weights: |
| 2032 | plan.source_weights["corpus"] = 1.0 |
| 2033 | plan.source_weights = planner._normalize_weights(plan.source_weights) |
| 2034 | |
| 2035 | # Always-on planner trace. Emits one summary line plus one per subquery |
| 2036 | # so retrieval-breadth failures like the 2026-04-19 Hermes Agent Use Cases |
| 2037 | # disaster are visible without --debug. Stderr only; does not leak into |
| 2038 | # the user-facing stdout synthesis. |
| 2039 | print( |
| 2040 | f"[Planner] Plan: intent={plan.intent}, freshness={plan.freshness_mode}, " |
| 2041 | f"cluster_mode={plan.cluster_mode}, subqueries={len(plan.subqueries)}, " |
| 2042 | f"source={plan_source}", |
| 2043 | file=sys.stderr, |
| 2044 | ) |
| 2045 | if plan.subqueries: |
| 2046 | for index, sq in enumerate(plan.subqueries, start=1): |
| 2047 | sources_str = ",".join(sq.sources) if sq.sources else "(none)" |
| 2048 | print( |
| 2049 | f"[Planner] sq{index} label={sq.label} " |
| 2050 | f'search="{sq.search_query}" sources=[{sources_str}]', |
| 2051 | file=sys.stderr, |
| 2052 | ) |
| 2053 | else: |
| 2054 | print("[Planner] (no subqueries in plan)", file=sys.stderr) |
| 2055 | |
| 2056 | bundle = schema.RetrievalBundle(artifacts={"grounding": []}) |
| 2057 | for source in (requested_sources or []): |
| 2058 | if source not in available: |
| 2059 | bundle.record_failure( |
| 2060 | source, |
| 2061 | schema.SKIPPED_UNCONFIGURED, |
| 2062 | "Source was requested but is not configured for this run.", |
| 2063 | attempted=False, |
| 2064 | ) |
| 2065 | if corpus_requested and not corpus_enabled: |
| 2066 | bundle.record_failure( |
| 2067 | "corpus", |
| 2068 | schema.SKIPPED_UNCONFIGURED, |
| 2069 | "Corpus was requested but no readable directory was configured.", |
| 2070 | attempted=False, |
| 2071 | ) |
| 2072 | # Expose plan_source to the renderer so render_compact can emit the |
| 2073 | # DEGRADED RUN banner when a named-entity topic was invoked bare |
| 2074 | # (source=deterministic AND no pre-research flags). LAW 7 backstop. |
| 2075 | bundle.artifacts["plan_source"] = plan_source |
| 2076 | bundle.artifacts["corpus_in_export"] = bool(config.get("_CORPUS_IN_EXPORT")) |
| 2077 | # Hiring-signals is deliberately jobs-only with no multi-source --plan, so |
| 2078 | # the LAW 7 degraded-run and Step 0.55 pre-research banners do not apply - |
| 2079 | # they would contradict the documented jobs-scoped flow. Suppress them. |
| 2080 | bundle.artifacts["hiring_signals_mode"] = hiring_signals_mode |
| 2081 | |
| 2082 | # Project-mode or person-mode GitHub: run once before the main subquery loop |
| 2083 | _github_custom_done = False |
| 2084 | _github_enriched_repos: set[str] = set() |
| 2085 | |
| 2086 | # Project mode takes priority over person mode |
| 2087 | if github_repos and "github" in available: |
| 2088 | bundle.mark_attempted("github") |
| 2089 | try: |
| 2090 | project_items = github.search_github_project( |
| 2091 | github_repos, from_date, to_date, |
| 2092 | depth=depth, token=config.get("GITHUB_TOKEN"), |
| 2093 | ) |
| 2094 | if project_items: |
| 2095 | normalized = _normalize_score_dedupe( |
| 2096 | "github", project_items, from_date, to_date, |
| 2097 | freshness_mode=plan.freshness_mode, |
| 2098 | ranking_query=f"What are {', '.join(github_repos)} doing on GitHub?", |
| 2099 | ) |
| 2100 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 2101 | bundle.add_items(primary_label, "github", normalized) |
| 2102 | _github_custom_done = True |
| 2103 | _github_enriched_repos = {r.lower() for r in github_repos} |
| 2104 | except Exception as exc: |
| 2105 | bundle.errors_by_source["github"] = f"Project-mode failed: {exc}" |
| 2106 | state, attempted = _classify_source_failure(exc) |
| 2107 | bundle.record_failure("github", state, str(exc), attempted=attempted) |
| 2108 | |
| 2109 | _github_person_done = False |
| 2110 | if github_user and "github" in available and not _github_custom_done: |
| 2111 | bundle.mark_attempted("github") |
| 2112 | _github_person_done = True |
| 2113 | try: |
| 2114 | person_items = github.search_github_person( |
| 2115 | github_user, from_date, to_date, |
| 2116 | depth=depth, token=config.get("GITHUB_TOKEN"), |
| 2117 | ) |
| 2118 | if person_items: |
| 2119 | normalized = _normalize_score_dedupe( |
| 2120 | "github", person_items, from_date, to_date, |
| 2121 | freshness_mode=plan.freshness_mode, |
| 2122 | ranking_query=f"What is @{github_user} doing on GitHub?", |
| 2123 | ) |
| 2124 | # Use the first subquery's label so RRF can look up the weight |
| 2125 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 2126 | bundle.add_items(primary_label, "github", normalized) |
| 2127 | else: |
| 2128 | # A pinned --github-user that yields nothing must not be |
| 2129 | # silently backfilled by generic keyword search: the report |
| 2130 | # would then present unrelated repos as this person's work. |
| 2131 | bundle.record_failure( |
| 2132 | "github", |
| 2133 | "no-results", |
| 2134 | f"Person mode found no activity for @{github_user} in the window", |
| 2135 | ) |
| 2136 | except Exception as exc: |
| 2137 | bundle.errors_by_source["github"] = f"Person-mode failed: {exc}" |
| 2138 | state, attempted = _classify_source_failure(exc) |
| 2139 | bundle.record_failure("github", state, str(exc), attempted=attempted) |
| 2140 | |
| 2141 | # Trustpilot session warm-up happens inside search_trustpilot at the |
| 2142 | # first (capped, single) fetch -- lazily, so it never delays the other |
| 2143 | # sources' streams and never fires for runs whose plan fetches no |
| 2144 | # Trustpilot. The module-level lock in lib/trustpilot.py serializes |
| 2145 | # concurrent vs-mode sub-runs so they never race Chrome harvests. |
| 2146 | |
| 2147 | # Thread-safe set prevents redundant fetches after a source returns 429 |
| 2148 | rate_limited_sources: set[str] = set() |
| 2149 | rate_limit_lock = threading.Lock() |
| 2150 | |
| 2151 | # Local corpus retrieval is intentionally outside the network executor and |
| 2152 | # retry budget. One bounded stream participates in the same signal scoring, |
| 2153 | # fusion, reranking, and per-source result cap as remote sources. |
| 2154 | if corpus_enabled and plan.subqueries: |
| 2155 | primary = plan.subqueries[0] |
| 2156 | bundle.mark_attempted("corpus") |
| 2157 | result = corpus.search( |
| 2158 | topic, |
| 2159 | resolved_corpus_dirs, |
| 2160 | from_date=from_date, |
| 2161 | to_date=to_date, |
| 2162 | all_time=corpus_all_time, |
| 2163 | limit=settings["per_stream_limit"], |
| 2164 | cache_dir=env.CONFIG_DIR, |
| 2165 | ) |
| 2166 | prepared_query = relevance.PreparedQuery(primary.ranking_query) |
| 2167 | lookback_window_days = ( |
| 2168 | datetime.strptime(to_date, "%Y-%m-%d").date() |
| 2169 | - datetime.strptime(from_date, "%Y-%m-%d").date() |
| 2170 | ).days |
| 2171 | corpus_items = signals.annotate_stream( |
| 2172 | result.items, |
| 2173 | prepared_query, |
| 2174 | plan.freshness_mode, |
| 2175 | reference_date=to_date, |
| 2176 | max_days=lookback_window_days, |
| 2177 | ) |
| 2178 | corpus_items = signals.prune_low_relevance(corpus_items) |
| 2179 | corpus_items = dedupe.dedupe_items(corpus_items) |
| 2180 | for item in corpus_items: |
| 2181 | item.snippet = snippet.extract_best_snippet(item, prepared_query) |
| 2182 | bundle.add_items(primary.label, "corpus", corpus_items) |
| 2183 | if result.notes: |
| 2184 | outcome = bundle.source_status["corpus"] |
| 2185 | bundle.source_status["corpus"] = schema.SourceOutcome( |
| 2186 | source="corpus", |
| 2187 | state=outcome.state, |
| 2188 | items_returned=outcome.items_returned, |
| 2189 | attempted=True, |
| 2190 | detail="; ".join(result.notes), |
| 2191 | ) |
| 2192 | bundle.artifacts["corpus"] = { |
| 2193 | "files_scanned": result.files_scanned, |
| 2194 | "cache_hits": result.cache_hits, |
| 2195 | "all_time": corpus_all_time, |
| 2196 | } |
| 2197 | |
| 2198 | futures = {} |
| 2199 | # Per-source fetch budget prevents redundant API calls |
| 2200 | source_fetch_count: dict[str, int] = {} |
| 2201 | stream_count = sum( |
| 2202 | 1 |
| 2203 | for subquery in plan.subqueries |
| 2204 | for source in subquery.sources |
| 2205 | if source in available and source != "corpus" |
| 2206 | ) |
| 2207 | max_workers = _inner_max_workers(stream_count, internal_subrun=internal_subrun) |
| 2208 | with ThreadPoolExecutor(max_workers=max_workers) as executor: |
| 2209 | for subquery in plan.subqueries: |
| 2210 | for source in subquery.sources: |
| 2211 | if source not in available: |
| 2212 | continue |
| 2213 | if source == "corpus": |
| 2214 | continue |
| 2215 | # Skip GitHub keyword search if person-mode already ran |
| 2216 | if source == "github" and (_github_person_done or _github_custom_done): |
| 2217 | continue |
| 2218 | # Enforce per-source fetch cap. A CLI override (issue #716) raises |
| 2219 | # the cap for capped sources so every X subquery in a multi-angle |
| 2220 | # --plan fetches, instead of only the first two. |
| 2221 | cap = MAX_SOURCE_FETCHES.get(source) |
| 2222 | _cap_override = config.get("_max_source_fetches") |
| 2223 | if cap is not None and _cap_override is not None: |
| 2224 | # `is not None` so --max-source-fetches 0 (disable fetching a |
| 2225 | # capped source) is honored instead of falling back to default. |
| 2226 | cap = int(_cap_override) |
| 2227 | if cap is not None: |
| 2228 | current = source_fetch_count.get(source, 0) |
| 2229 | if current >= cap: |
| 2230 | continue |
| 2231 | source_fetch_count[source] = current + 1 |
| 2232 | bundle.mark_attempted(source) |
| 2233 | futures[ |
| 2234 | executor.submit( |
| 2235 | _retrieve_stream, |
| 2236 | topic=topic, |
| 2237 | subquery=subquery, |
| 2238 | source=source, |
| 2239 | config=config, |
| 2240 | depth=depth, |
| 2241 | date_range=(from_date, to_date), |
| 2242 | runtime=runtime, |
| 2243 | mock=mock, |
| 2244 | rate_limited_sources=rate_limited_sources, |
| 2245 | rate_limit_lock=rate_limit_lock, |
| 2246 | web_backend=web_backend, |
| 2247 | raw_topic=topic, |
| 2248 | subreddits=subreddits, |
| 2249 | tiktok_hashtags=tiktok_hashtags, |
| 2250 | tiktok_creators=tiktok_creators, |
| 2251 | ig_creators=ig_creators, |
| 2252 | trustpilot_domain=trustpilot_domain, |
| 2253 | trustpilot_domain_is_hint=trustpilot_domain_is_hint, |
| 2254 | ) |
| 2255 | ] = (subquery, source) |
| 2256 | |
| 2257 | for future in as_completed(futures): |
| 2258 | subquery, source = futures[future] |
| 2259 | try: |
| 2260 | raw_items, artifact = future.result() |
| 2261 | except Exception as exc: |
| 2262 | # Share 429 signal so pending futures skip this source |
| 2263 | if _is_rate_limit_error(exc): |
| 2264 | with rate_limit_lock: |
| 2265 | rate_limited_sources.add(source) |
| 2266 | bundle.errors_by_source[source] = str(exc) |
| 2267 | state, attempted = _classify_source_failure(exc) |
| 2268 | bundle.record_failure(source, state, str(exc), attempted=attempted) |
| 2269 | continue |
| 2270 | # Retry once for transient 5xx errors |
| 2271 | if _is_transient_error(exc): |
| 2272 | time.sleep(3) |
| 2273 | try: |
| 2274 | raw_items, artifact = _retrieve_stream( |
| 2275 | topic=topic, subquery=subquery, source=source, |
| 2276 | config=config, depth=depth, date_range=(from_date, to_date), |
| 2277 | runtime=runtime, mock=mock, |
| 2278 | rate_limited_sources=rate_limited_sources, |
| 2279 | rate_limit_lock=rate_limit_lock, |
| 2280 | web_backend=web_backend, |
| 2281 | raw_topic=topic, |
| 2282 | subreddits=subreddits, |
| 2283 | tiktok_hashtags=tiktok_hashtags, |
| 2284 | tiktok_creators=tiktok_creators, |
| 2285 | ig_creators=ig_creators, |
| 2286 | trustpilot_domain=trustpilot_domain, |
| 2287 | trustpilot_domain_is_hint=trustpilot_domain_is_hint, |
| 2288 | ) |
| 2289 | except Exception as retry_exc: |
| 2290 | detail = f"{exc} (retried once, still failed: {retry_exc})" |
| 2291 | bundle.errors_by_source[source] = detail |
| 2292 | state, attempted = _classify_source_failure(retry_exc) |
| 2293 | bundle.record_failure(source, state, detail, attempted=attempted) |
| 2294 | continue |
| 2295 | else: |
| 2296 | bundle.errors_by_source[source] = str(exc) |
| 2297 | state, attempted = _classify_source_failure(exc) |
| 2298 | bundle.record_failure(source, state, str(exc), attempted=attempted) |
| 2299 | continue |
| 2300 | outcome_note = None |
| 2301 | if isinstance(artifact, dict) and artifact.get("_source_outcome"): |
| 2302 | artifact = dict(artifact) |
| 2303 | outcome_note = artifact.pop("_source_outcome") |
| 2304 | bundle.record_failure( |
| 2305 | source, |
| 2306 | outcome_note["state"], |
| 2307 | outcome_note["detail"], |
| 2308 | attempted=outcome_note.get("attempted", True), |
| 2309 | ) |
| 2310 | normalized = _normalize_score_dedupe( |
| 2311 | source, raw_items, from_date, to_date, |
| 2312 | freshness_mode=plan.freshness_mode, |
| 2313 | ranking_query=subquery.ranking_query, |
| 2314 | ) |
| 2315 | # Jobs is exempt from per_stream_limit: a careers board is a complete |
| 2316 | # snapshot of open roles, and truncating it to the default 12 drops |
| 2317 | # strategic postings (the whole point of hiring-signals coverage). |
| 2318 | if source != "jobs": |
| 2319 | normalized = normalized[: settings["per_stream_limit"]] |
| 2320 | bundle.add_items(subquery.label, source, normalized) |
| 2321 | if artifact: |
| 2322 | bundle.artifacts.setdefault("grounding", []).append(artifact) |
| 2323 | |
| 2324 | # Phase 2: supplemental entity-based searches |
| 2325 | _run_supplemental_searches( |
| 2326 | topic=topic, |
| 2327 | bundle=bundle, |
| 2328 | plan=plan, |
| 2329 | config=config, |
| 2330 | depth=depth, |
| 2331 | date_range=(from_date, to_date), |
| 2332 | runtime=runtime, |
| 2333 | mock=mock, |
| 2334 | rate_limited_sources=rate_limited_sources, |
| 2335 | rate_limit_lock=rate_limit_lock, |
| 2336 | x_handle=x_handle, |
| 2337 | x_related=x_related, |
| 2338 | ) |
| 2339 | |
| 2340 | # Phase 2b: retry thin sources with simplified query |
| 2341 | # Note: _github_skip_sources tells the retry to not re-run GitHub keyword search |
| 2342 | # when project-mode or person-mode already provided authoritative data. |
| 2343 | _github_skip_retry = {"corpus"} |
| 2344 | if _github_person_done or _github_custom_done: |
| 2345 | _github_skip_retry.add("github") |
| 2346 | _retry_thin_sources( |
| 2347 | topic=topic, |
| 2348 | bundle=bundle, |
| 2349 | plan=plan, |
| 2350 | config=config, |
| 2351 | depth=depth, |
| 2352 | date_range=(from_date, to_date), |
| 2353 | runtime=runtime, |
| 2354 | mock=mock, |
| 2355 | rate_limited_sources=rate_limited_sources, |
| 2356 | rate_limit_lock=rate_limit_lock, |
| 2357 | settings=settings, |
| 2358 | web_backend=web_backend, |
| 2359 | skip_sources=_github_skip_retry, |
| 2360 | subreddits=subreddits, |
| 2361 | tiktok_hashtags=tiktok_hashtags, |
| 2362 | tiktok_creators=tiktok_creators, |
| 2363 | ig_creators=ig_creators, |
| 2364 | ) |
| 2365 | |
| 2366 | # Reclassify partial failures as DEGRADED instead of silently dropping them. |
| 2367 | # A source that 429'd on one subquery but succeeded on another is not a hard |
| 2368 | # failure, but it is not healthy either: it likely returned fewer results |
| 2369 | # than it should have. Move it out of errors_by_source (so it isn't reported |
| 2370 | # as "failed") and into degraded_by_source (so it survives into warnings), |
| 2371 | # rather than deleting the signal outright as the engine used to. |
| 2372 | degraded_by_source: dict[str, str] = {} |
| 2373 | for source in list(bundle.errors_by_source): |
| 2374 | if bundle.items_by_source.get(source): |
| 2375 | degraded_by_source[source] = bundle.errors_by_source[source] |
| 2376 | del bundle.errors_by_source[source] |
| 2377 | |
| 2378 | hiring_summary = _apply_hiring_signal_gate( |
| 2379 | bundle, |
| 2380 | explicit=hiring_signals_mode, |
| 2381 | topic=topic, |
| 2382 | ) |
| 2383 | if hiring_summary: |
| 2384 | bundle.artifacts["hiring_signals"] = hiring_summary |
| 2385 | |
| 2386 | items_by_source = _finalize_items_by_source( |
| 2387 | bundle.items_by_source, topic=topic, config=config, depth=depth, mock=mock, |
| 2388 | ) |
| 2389 | source_status = _finalize_source_status(bundle.source_status, items_by_source) |
| 2390 | candidates = weighted_rrf(bundle.items_by_source_and_query, plan, pool_limit=settings["pool_limit"]) |
| 2391 | # Normalized set of handles this run resolved for the topic. A candidate |
| 2392 | # authored by one of these is first-party and is exempted from the |
| 2393 | # entity-miss demotion in rerank (a post never repeats its own author's |
| 2394 | # name, so the body-text grounding check would otherwise zero out the |
| 2395 | # subject's own highest-signal posts). |
| 2396 | resolved_handles = { |
| 2397 | h.lstrip("@").strip().lower() |
| 2398 | for h in ([x_handle, github_user, *(x_related or [])]) |
| 2399 | if h and h.strip() |
| 2400 | } |
| 2401 | private_candidates = [ |
| 2402 | candidate |
| 2403 | for candidate in candidates |
| 2404 | if candidate.source == "corpus" |
| 2405 | or any(item.source == "corpus" for item in candidate.source_items) |
| 2406 | ] |
| 2407 | private_candidate_ids = {id(candidate) for candidate in private_candidates} |
| 2408 | public_candidates = [ |
| 2409 | candidate for candidate in candidates if id(candidate) not in private_candidate_ids |
| 2410 | ] |
| 2411 | ranked_public = rerank.rerank_candidates( |
| 2412 | topic=topic, |
| 2413 | plan=plan, |
| 2414 | candidates=public_candidates, |
| 2415 | provider=None if mock else reasoning_provider, |
| 2416 | model=None if mock else runtime.rerank_model, |
| 2417 | shortlist_size=settings["rerank_limit"], |
| 2418 | resolved_handles=resolved_handles, |
| 2419 | ) |
| 2420 | # Corpus titles/snippets must never enter a hosted reasoning prompt. Score |
| 2421 | # every candidate carrying corpus evidence with the deterministic fallback, |
| 2422 | # even when the rest of the run uses a remote reranker. |
| 2423 | ranked_private = rerank.rerank_candidates( |
| 2424 | topic=topic, |
| 2425 | plan=plan, |
| 2426 | candidates=private_candidates, |
| 2427 | provider=None, |
| 2428 | model=None, |
| 2429 | shortlist_size=settings["rerank_limit"], |
| 2430 | resolved_handles=resolved_handles, |
| 2431 | ) |
| 2432 | ranked_public = rerank.prune_fallback_entity_misses(ranked_public, topic=topic) |
| 2433 | # Private corpus already cleared a body-aware retrieval floor; do not apply |
| 2434 | # the public title/snippet visibility gate (filenames often omit the head |
| 2435 | # token even when the document body matched). |
| 2436 | ranked_candidates = sorted( |
| 2437 | [*ranked_public, *ranked_private], |
| 2438 | key=lambda candidate: ( |
| 2439 | -candidate.final_score, |
| 2440 | -(candidate.engagement or -1), |
| 2441 | min(candidate.native_ranks.values(), default=999), |
| 2442 | candidate.title, |
| 2443 | ), |
| 2444 | ) |
| 2445 | rerank.score_fun( |
| 2446 | topic=topic, |
| 2447 | candidates=ranked_public, |
| 2448 | provider=None if mock else reasoning_provider, |
| 2449 | model=None if mock else runtime.rerank_model, |
| 2450 | ) |
| 2451 | rerank.score_fun( |
| 2452 | topic=topic, |
| 2453 | candidates=ranked_private, |
| 2454 | provider=None, |
| 2455 | model=None, |
| 2456 | ) |
| 2457 | |
| 2458 | # Phase 3: post-rerank GitHub star enrichment. Record/replay-aware so the |
| 2459 | # eval harness stays fully offline: this path calls the GitHub API (and the |
| 2460 | # gh-credential fallback) outside the _retrieve_stream seam, so it gets its |
| 2461 | # own fixture exchange keyed by phase. |
| 2462 | if "github" in available and not mock: |
| 2463 | star_request = { |
| 2464 | "source": "github", |
| 2465 | "phase": "post_rerank_star_enrichment", |
| 2466 | "topic": topic, |
| 2467 | "depth": depth, |
| 2468 | } |
| 2469 | star_matched, star_replayed = http.fixture_source_replay(star_request) |
| 2470 | if star_matched: |
| 2471 | star_map = star_replayed if isinstance(star_replayed, dict) else {} |
| 2472 | github.apply_star_map(ranked_candidates, star_map) |
| 2473 | else: |
| 2474 | collected_star_map: dict[str, int] = {} |
| 2475 | github.enrich_candidates_with_stars( |
| 2476 | ranked_candidates, |
| 2477 | token=config.get("GITHUB_TOKEN"), |
| 2478 | already_enriched=_github_enriched_repos, |
| 2479 | collect_map=collected_star_map, |
| 2480 | ) |
| 2481 | http.fixture_source_record(star_request, collected_star_map) |
| 2482 | |
| 2483 | clusters = cluster_candidates(ranked_candidates, plan) |
| 2484 | warnings = _warnings(items_by_source, ranked_candidates, bundle.errors_by_source, degraded_by_source) |
| 2485 | library_context, library_warning = _load_library_context( |
| 2486 | topic=topic, |
| 2487 | config=config, |
| 2488 | mock=mock, |
| 2489 | internal_subrun=internal_subrun, |
| 2490 | x_handle=x_handle, |
| 2491 | github_user=github_user, |
| 2492 | github_repos=github_repos, |
| 2493 | save_dir=save_dir, |
| 2494 | ) |
| 2495 | if library_warning: |
| 2496 | warnings.append(library_warning) |
| 2497 | |
| 2498 | return schema.Report( |
| 2499 | topic=topic, |
| 2500 | range_from=from_date, |
| 2501 | range_to=to_date, |
| 2502 | generated_at=datetime.now(timezone.utc).isoformat(), |
| 2503 | provider_runtime=runtime, |
| 2504 | query_plan=plan, |
| 2505 | clusters=clusters, |
| 2506 | ranked_candidates=ranked_candidates, |
| 2507 | items_by_source=items_by_source, |
| 2508 | errors_by_source=bundle.errors_by_source, |
| 2509 | source_status=source_status, |
| 2510 | warnings=warnings, |
| 2511 | artifacts=bundle.artifacts, |
| 2512 | library_context=library_context, |
| 2513 | ) |
| 2514 | |
| 2515 | |
| 2516 | def _candidate_is_duplicate( |
| 2517 | candidate: schema.Candidate, |
| 2518 | kept: list[schema.Candidate], |
| 2519 | ) -> bool: |
| 2520 | if any(existing.candidate_id == candidate.candidate_id for existing in kept): |
| 2521 | return True |
| 2522 | if candidate.url and any(existing.url == candidate.url for existing in kept): |
| 2523 | return True |
| 2524 | candidate_text = " ".join((candidate.title, candidate.snippet)).strip() |
| 2525 | return bool(candidate_text) and any( |
| 2526 | dedupe.hybrid_similarity( |
| 2527 | candidate_text, |
| 2528 | " ".join((existing.title, existing.snippet)).strip(), |
| 2529 | ) >= 0.7 |
| 2530 | for existing in kept |
| 2531 | ) |
| 2532 | |
| 2533 | |
| 2534 | def merge_drill_report( |
| 2535 | report: schema.Report, |
| 2536 | drill_report: schema.Report, |
| 2537 | matched_clusters: list[schema.Cluster], |
| 2538 | *, |
| 2539 | target: str, |
| 2540 | ) -> schema.Report: |
| 2541 | """Merge a narrow follow-up into its cached report while preserving other clusters.""" |
| 2542 | merged = copy.deepcopy(report) |
| 2543 | selected_cluster_ids = {cluster.cluster_id for cluster in matched_clusters} |
| 2544 | selected_candidate_ids = { |
| 2545 | candidate_id |
| 2546 | for cluster in matched_clusters |
| 2547 | for candidate_id in cluster.candidate_ids |
| 2548 | } |
| 2549 | original_candidates = { |
| 2550 | candidate.candidate_id: candidate for candidate in merged.ranked_candidates |
| 2551 | } |
| 2552 | unrelated_candidates = [ |
| 2553 | candidate for candidate in merged.ranked_candidates |
| 2554 | if candidate.candidate_id not in selected_candidate_ids |
| 2555 | ] |
| 2556 | original_summary = "" |
| 2557 | for cluster in matched_clusters: |
| 2558 | for candidate_id in cluster.representative_ids: |
| 2559 | candidate = original_candidates.get(candidate_id) |
| 2560 | if candidate: |
| 2561 | original_summary = candidate.snippet or candidate.explanation or candidate.title |
| 2562 | if original_summary: |
| 2563 | break |
| 2564 | if original_summary: |
| 2565 | break |
| 2566 | |
| 2567 | unrelated_candidate_indexes = { |
| 2568 | candidate.candidate_id: index |
| 2569 | for index, candidate in enumerate(unrelated_candidates) |
| 2570 | } |
| 2571 | focused_candidates: list[schema.Candidate] = [] |
| 2572 | for candidate in [ |
| 2573 | *copy.deepcopy(drill_report.ranked_candidates), |
| 2574 | *[ |
| 2575 | copy.deepcopy(candidate) |
| 2576 | for candidate in merged.ranked_candidates |
| 2577 | if candidate.candidate_id in selected_candidate_ids |
| 2578 | ], |
| 2579 | ]: |
| 2580 | unrelated_index = unrelated_candidate_indexes.get(candidate.candidate_id) |
| 2581 | if unrelated_index is not None: |
| 2582 | candidate.cluster_id = unrelated_candidates[unrelated_index].cluster_id |
| 2583 | unrelated_candidates[unrelated_index] = candidate |
| 2584 | continue |
| 2585 | if not _candidate_is_duplicate(candidate, focused_candidates): |
| 2586 | focused_candidates.append(candidate) |
| 2587 | |
| 2588 | primary_cluster = matched_clusters[0] |
| 2589 | for candidate in focused_candidates: |
| 2590 | candidate.cluster_id = primary_cluster.cluster_id |
| 2591 | focused_ids = [candidate.candidate_id for candidate in focused_candidates] |
| 2592 | focused_sources = sorted({ |
| 2593 | source |
| 2594 | for candidate in focused_candidates |
| 2595 | for source in schema.candidate_sources(candidate) |
| 2596 | }) |
| 2597 | replacement_cluster = schema.Cluster( |
| 2598 | cluster_id=primary_cluster.cluster_id, |
| 2599 | title=primary_cluster.title, |
| 2600 | candidate_ids=focused_ids, |
| 2601 | representative_ids=focused_ids[:3], |
| 2602 | sources=focused_sources, |
| 2603 | score=max((candidate.final_score for candidate in focused_candidates), default=0.0), |
| 2604 | uncertainty="single-source" if len(focused_sources) == 1 else None, |
| 2605 | ) |
| 2606 | |
| 2607 | first_selected_index = min( |
| 2608 | index |
| 2609 | for index, cluster in enumerate(merged.clusters) |
| 2610 | if cluster.cluster_id in selected_cluster_ids |
| 2611 | ) |
| 2612 | remaining_clusters = [ |
| 2613 | cluster for cluster in merged.clusters |
| 2614 | if cluster.cluster_id not in selected_cluster_ids |
| 2615 | ] |
| 2616 | remaining_clusters.insert(first_selected_index, replacement_cluster) |
| 2617 | merged.clusters = remaining_clusters |
| 2618 | |
| 2619 | merged.ranked_candidates = focused_candidates + unrelated_candidates |
| 2620 | |
| 2621 | all_sources = set(merged.items_by_source) | set(drill_report.items_by_source) |
| 2622 | new_item_count = 0 |
| 2623 | merged_items: dict[str, list[schema.SourceItem]] = {} |
| 2624 | for source in sorted(all_sources): |
| 2625 | old_items = merged.items_by_source.get(source, []) |
| 2626 | new_items = drill_report.items_by_source.get(source, []) |
| 2627 | # Collapse exact URL matches first, preferring the drill's copy (it |
| 2628 | # carries fresh transcripts/comments); fuzzy dedupe alone keeps both |
| 2629 | # when enrichment changed the text substantially. |
| 2630 | new_urls = {item.url for item in new_items if item.url} |
| 2631 | kept_old = [item for item in old_items if not (item.url and item.url in new_urls)] |
| 2632 | combined = dedupe.dedupe_items([*copy.deepcopy(new_items), *kept_old]) |
| 2633 | old_unique = dedupe.dedupe_items(old_items) |
| 2634 | new_item_count += max(0, len(combined) - len(old_unique)) |
| 2635 | merged_items[source] = combined |
| 2636 | merged.items_by_source = merged_items |
| 2637 | |
| 2638 | merged.generated_at = drill_report.generated_at |
| 2639 | merged.query_plan = drill_report.query_plan |
| 2640 | # The drill's retrieval window is the report's window now (a --days/--as-of |
| 2641 | # override on the drill must not be mislabeled with the cached range). |
| 2642 | merged.range_from = drill_report.range_from |
| 2643 | merged.range_to = drill_report.range_to |
| 2644 | attempted_sources = { |
| 2645 | source |
| 2646 | for source, outcome in drill_report.source_status.items() |
| 2647 | if outcome.attempted or outcome.state == schema.SKIPPED_UNCONFIGURED |
| 2648 | } |
| 2649 | for source in attempted_sources: |
| 2650 | if source in drill_report.errors_by_source: |
| 2651 | merged.errors_by_source[source] = drill_report.errors_by_source[source] |
| 2652 | else: |
| 2653 | merged.errors_by_source.pop(source, None) |
| 2654 | merged.source_status[source] = drill_report.source_status[source] |
| 2655 | merged.source_status = _finalize_source_status( |
| 2656 | merged.source_status, |
| 2657 | merged.items_by_source, |
| 2658 | ) |
| 2659 | degraded_by_source = { |
| 2660 | source: outcome.detail or "partial results" |
| 2661 | for source, outcome in merged.source_status.items() |
| 2662 | if outcome.state == schema.PARTIAL |
| 2663 | } |
| 2664 | merged.warnings = _warnings( |
| 2665 | merged.items_by_source, |
| 2666 | merged.ranked_candidates, |
| 2667 | merged.errors_by_source, |
| 2668 | degraded_by_source, |
| 2669 | ) |
| 2670 | merged.artifacts.update(copy.deepcopy(drill_report.artifacts)) |
| 2671 | history = list(merged.artifacts.get("drill_history") or []) |
| 2672 | history.append({ |
| 2673 | "target": target, |
| 2674 | "clusters": [cluster.title for cluster in matched_clusters], |
| 2675 | "new_items": new_item_count, |
| 2676 | "generated_at": drill_report.generated_at, |
| 2677 | }) |
| 2678 | merged.artifacts["drill_history"] = history |
| 2679 | merged.artifacts["drill_context"] = { |
| 2680 | "target": target, |
| 2681 | "cluster_titles": [cluster.title for cluster in matched_clusters], |
| 2682 | "original_summary": original_summary, |
| 2683 | "new_items": new_item_count, |
| 2684 | "sources": focused_sources, |
| 2685 | } |
| 2686 | merged.drill_of = primary_cluster.title |
| 2687 | return merged |
| 2688 | |
| 2689 | |
| 2690 | def _normalize_score_dedupe( |
| 2691 | source: str, |
| 2692 | raw_items: list[dict], |
| 2693 | from_date: str, |
| 2694 | to_date: str, |
| 2695 | freshness_mode: str, |
| 2696 | ranking_query: str, |
| 2697 | ) -> list[schema.SourceItem]: |
| 2698 | """Normalize, annotate, prune, dedupe, and extract snippets for a batch of raw items.""" |
| 2699 | normalized = normalize.normalize_source_items( |
| 2700 | source, raw_items, from_date, to_date, |
| 2701 | freshness_mode=freshness_mode, |
| 2702 | ) |
| 2703 | prepared_query = relevance.PreparedQuery(ranking_query) |
| 2704 | lookback_window_days = ( |
| 2705 | datetime.strptime(to_date, "%Y-%m-%d").date() |
| 2706 | - datetime.strptime(from_date, "%Y-%m-%d").date() |
| 2707 | ).days |
| 2708 | normalized = signals.annotate_stream( |
| 2709 | normalized, |
| 2710 | prepared_query, |
| 2711 | freshness_mode, |
| 2712 | reference_date=to_date, |
| 2713 | max_days=lookback_window_days, |
| 2714 | ) |
| 2715 | if source != "jobs": |
| 2716 | normalized = signals.prune_low_relevance(normalized) |
| 2717 | normalized = dedupe.dedupe_items(normalized) |
| 2718 | for item in normalized: |
| 2719 | item.snippet = snippet.extract_best_snippet(item, prepared_query) |
| 2720 | return normalized |
| 2721 | |
| 2722 | |
| 2723 | def _finalize_items_by_source( |
| 2724 | items_by_source_raw: dict[str, list[schema.SourceItem]], |
| 2725 | topic: str = "", |
| 2726 | config: dict | None = None, |
| 2727 | depth: str = "default", |
| 2728 | mock: bool = False, |
| 2729 | ) -> dict[str, list[schema.SourceItem]]: |
| 2730 | finalized = {} |
| 2731 | for source, items in items_by_source_raw.items(): |
| 2732 | items = sorted(items, key=lambda item: item.local_rank_score or 0.0, reverse=True) |
| 2733 | items = dedupe.dedupe_items(items) |
| 2734 | enrichment_request = { |
| 2735 | "source": source, |
| 2736 | "phase": "post_ranking_enrichment", |
| 2737 | "topic": topic, |
| 2738 | "depth": depth, |
| 2739 | } |
| 2740 | if source == "youtube" and items and not mock: |
| 2741 | # Same budget-at-the-survivors principle as the digg branch |
| 2742 | # below: retrieval-time transcripts go to each search's |
| 2743 | # top-by-views candidates, while final selection ranks by |
| 2744 | # relevance. Backfill survivors that arrived without one so the |
| 2745 | # transcript budget lands on videos the brief actually shows |
| 2746 | # (#542). |
| 2747 | matched, replayed = http.fixture_source_replay(enrichment_request) |
| 2748 | if matched: |
| 2749 | items = _merge_replayed_enrichment(items, replayed) |
| 2750 | else: |
| 2751 | sc_token = ( |
| 2752 | config.get("SCRAPECREATORS_API_KEY") |
| 2753 | if config and env.is_youtube_sc_available(config) else None |
| 2754 | ) |
| 2755 | youtube_yt.backfill_transcripts( |
| 2756 | items, topic=topic, depth=depth, token=sc_token, |
| 2757 | ) |
| 2758 | http.fixture_source_record(enrichment_request, schema.to_dict(items)) |
| 2759 | # Post-merge topic-relevance filter for Polymarket: comparison queries |
| 2760 | # fan out into per-entity subqueries ("Hermes", "OpenClaw") whose topic |
| 2761 | # is too narrow for Gamma API to filter meaningfully. Re-validating the |
| 2762 | # merged list against the full original topic drops off-topic markets |
| 2763 | # (e.g., WTI crude oil, Elon tweet counts) before footer emission. |
| 2764 | if source == "polymarket" and topic: |
| 2765 | items = polymarket.filter_items_against_topic(topic, items) |
| 2766 | # --polymarket-keywords (via config): additional keyword filter |
| 2767 | # for ambiguous single-token topics (e.g., "Warriors" → nba,gsw). |
| 2768 | keywords = config.get("_polymarket_keywords") if isinstance(config, dict) else None |
| 2769 | if keywords: |
| 2770 | items = polymarket.filter_items_against_keywords(items, keywords) |
| 2771 | if source == "digg" and items: |
| 2772 | # Pull top-ranked X posts only for the survivors that will appear |
| 2773 | # in the brief. Spending the enrichment budget here (rather than |
| 2774 | # at retrieval time) keeps the inline 'via Digg' quotes |
| 2775 | # paired with the clusters dedupe actually kept. |
| 2776 | matched, replayed = http.fixture_source_replay(enrichment_request) |
| 2777 | if matched: |
| 2778 | items = _merge_replayed_enrichment(items, replayed) |
| 2779 | else: |
| 2780 | digg.enrich_source_items(items, top_k=3) |
| 2781 | http.fixture_source_record(enrichment_request, schema.to_dict(items)) |
| 2782 | finalized[source] = items |
| 2783 | return finalized |
| 2784 | |
| 2785 | |
| 2786 | def _merge_replayed_enrichment( |
| 2787 | items: list[schema.SourceItem], |
| 2788 | replayed: list[dict], |
| 2789 | ) -> list[schema.SourceItem]: |
| 2790 | """Apply recorded post-ranking enrichment onto freshly computed items. |
| 2791 | |
| 2792 | Enrichment (transcripts, Digg posts) only mutates ``metadata``. Merging by |
| 2793 | item_id instead of replacing the list keeps normalization, scoring, and |
| 2794 | dedupe regressions visible to the eval - fixture state must not overwrite |
| 2795 | what the current pipeline computed. |
| 2796 | """ |
| 2797 | replayed_by_id = { |
| 2798 | entry.get("item_id"): entry for entry in replayed if isinstance(entry, dict) |
| 2799 | } |
| 2800 | for item in items: |
| 2801 | record = replayed_by_id.get(item.item_id) |
| 2802 | if record and record.get("metadata"): |
| 2803 | item.metadata.update(record["metadata"]) |
| 2804 | return items |
| 2805 | |
| 2806 | |
| 2807 | def _apply_hiring_signal_gate( |
| 2808 | bundle: schema.RetrievalBundle, |
| 2809 | *, |
| 2810 | explicit: bool, |
| 2811 | topic: str, |
| 2812 | ) -> dict[str, Any] | None: |
| 2813 | jobs_items = bundle.items_by_source.get("jobs") or [] |
| 2814 | if not jobs_items: |
| 2815 | if explicit: |
| 2816 | return hiring_signals.analyze([], explicit=True, topic=topic) |
| 2817 | return None |
| 2818 | |
| 2819 | summary = hiring_signals.analyze(jobs_items, explicit=explicit, topic=topic) |
| 2820 | if not explicit and not summary.get("include"): |
| 2821 | bundle.items_by_source.pop("jobs", None) |
| 2822 | for key in list(bundle.items_by_source_and_query): |
| 2823 | if key[1] == "jobs": |
| 2824 | del bundle.items_by_source_and_query[key] |
| 2825 | return summary |
| 2826 | |
| 2827 | |
| 2828 | def _ensure_jobs_in_plan( |
| 2829 | plan: schema.QueryPlan, |
| 2830 | available: list[str], |
| 2831 | *, |
| 2832 | explicit: bool, |
| 2833 | topic: str, |
| 2834 | ) -> None: |
| 2835 | if "jobs" not in available: |
| 2836 | return |
| 2837 | if not (explicit or _company_topic_likely(topic)): |
| 2838 | return |
| 2839 | if "jobs" not in plan.source_weights: |
| 2840 | plan.source_weights["jobs"] = 1.0 |
| 2841 | for subquery in plan.subqueries: |
| 2842 | if "jobs" not in subquery.sources: |
| 2843 | subquery.sources.append("jobs") |
| 2844 | |
| 2845 | |
| 2846 | def _company_topic_likely(topic: str) -> bool: |
| 2847 | text = topic.strip() |
| 2848 | if not text: |
| 2849 | return False |
| 2850 | lower = text.lower() |
| 2851 | if "?" in text or len(text.split()) > 4: |
| 2852 | return False |
| 2853 | generic = { |
| 2854 | "how", "what", "why", "best", "top", "tutorial", "guide", "prompts", |
| 2855 | "news", "latest", "ideas", "examples", |
| 2856 | } |
| 2857 | if any(word in generic for word in lower.split()): |
| 2858 | return False |
| 2859 | known_single_word_companies = { |
| 2860 | "apple", "uber", "google", "microsoft", "amazon", "meta", "netflix", |
| 2861 | "openai", "anthropic", "qualtrics", "stripe", "brex", |
| 2862 | } |
| 2863 | if " vs " in lower or " versus " in lower: |
| 2864 | parts = re.split(r"\s+(?:vs|versus)\s+", text, maxsplit=1, flags=re.IGNORECASE) |
| 2865 | if len(parts) != 2: |
| 2866 | return False |
| 2867 | return _comparison_side_company_like(parts[0], known_single_word_companies) or _comparison_side_company_like( |
| 2868 | parts[1], known_single_word_companies |
| 2869 | ) |
| 2870 | return bool(text[:1].isupper() or lower in known_single_word_companies) |
| 2871 | |
| 2872 | |
| 2873 | def _comparison_side_company_like(side: str, known_companies: set[str]) -> bool: |
| 2874 | token = re.sub(r"[^\w.+#-]", "", side.strip().split()[0] if side.strip() else "") |
| 2875 | if not token: |
| 2876 | return False |
| 2877 | lower = token.lower() |
| 2878 | common_tech_terms = { |
| 2879 | "python", "ruby", "javascript", "typescript", "java", "go", "golang", |
| 2880 | "rust", "php", "swift", "kotlin", "scala", "clojure", "elixir", |
| 2881 | "react", "vue", "angular", "svelte", "node", "django", "rails", |
| 2882 | "postgres", "mysql", "redis", "kubernetes", "docker", |
| 2883 | } |
| 2884 | if lower in common_tech_terms: |
| 2885 | return False |
| 2886 | return bool(token[:1].isupper() or lower in known_companies) |
| 2887 | |
| 2888 | |
| 2889 | def _warnings( |
| 2890 | items_by_source: dict[str, list[schema.SourceItem]], |
| 2891 | candidates: list[schema.Candidate], |
| 2892 | errors_by_source: dict[str, str], |
| 2893 | degraded_by_source: dict[str, str] | None = None, |
| 2894 | ) -> list[str]: |
| 2895 | warnings: list[str] = [] |
| 2896 | if not candidates: |
| 2897 | warnings.append("No candidates survived retrieval and ranking.") |
| 2898 | if len(candidates) < 5: |
| 2899 | warnings.append("Evidence is thin for this topic.") |
| 2900 | top_sources = { |
| 2901 | source |
| 2902 | for candidate in candidates[:5] |
| 2903 | for source in schema.candidate_sources(candidate) |
| 2904 | } |
| 2905 | if len(top_sources) <= 1 and len(candidates) >= 3: |
| 2906 | warnings.append("Top evidence is highly concentrated in one source.") |
| 2907 | if errors_by_source: |
| 2908 | warnings.append(f"Some sources failed: {', '.join(sorted(errors_by_source))}") |
| 2909 | if degraded_by_source: |
| 2910 | # Partial failures: the source returned some items but errored/timed out |
| 2911 | # on at least one subquery, so its coverage is likely incomplete. Kept |
| 2912 | # distinct from hard failures so the signal is not silently dropped. |
| 2913 | warnings.append( |
| 2914 | f"Some sources returned partial results (degraded): {', '.join(sorted(degraded_by_source))}" |
| 2915 | ) |
| 2916 | if not items_by_source: |
| 2917 | warnings.append("No source returned usable items.") |
| 2918 | return warnings |
| 2919 | |
| 2920 | |
| 2921 | def _is_rate_limit_error(exc: Exception) -> bool: |
| 2922 | """Detect 429 rate-limit errors by status code or message text.""" |
| 2923 | if hasattr(exc, "status_code") and getattr(exc, "status_code", None) == 429: |
| 2924 | return True |
| 2925 | return "429" in str(exc) |
| 2926 | |
| 2927 | |
| 2928 | class SourceRunError(RuntimeError): |
| 2929 | """Source-specific failure that survived a module's fallback logic.""" |
| 2930 | |
| 2931 | def __init__(self, message: str, state: schema.RunOutcomeState | None = None): |
| 2932 | super().__init__(message) |
| 2933 | self.outcome_state = state or http.classify_failure(message=message) |
| 2934 | |
| 2935 | |
| 2936 | def _classify_source_failure(exc: Exception) -> tuple[schema.RunOutcomeState, bool]: |
| 2937 | """Classify HTTP, subprocess, and module-specific failures consistently.""" |
| 2938 | detail = str(exc) |
| 2939 | lowered = detail.lower() |
| 2940 | if any(marker in lowered for marker in ("not configured", "no api key", "not installed")): |
| 2941 | return schema.SKIPPED_UNCONFIGURED, False |
| 2942 | if any( |
| 2943 | marker in lowered |
| 2944 | for marker in ("cookie expired", "expired cookie", "login required", "not logged in") |
| 2945 | ): |
| 2946 | return schema.AUTH_FAILED, True |
| 2947 | state = getattr(exc, "outcome_state", None) or http.classify_failure( |
| 2948 | status_code=getattr(exc, "status_code", None), |
| 2949 | message=detail, |
| 2950 | ) |
| 2951 | return state, True |
| 2952 | |
| 2953 | |
| 2954 | def _outcome_artifact( |
| 2955 | state: schema.RunOutcomeState, |
| 2956 | detail: str, |
| 2957 | *, |
| 2958 | attempted: bool = True, |
| 2959 | ) -> dict[str, Any]: |
| 2960 | return { |
| 2961 | "_source_outcome": { |
| 2962 | "state": state, |
| 2963 | "detail": detail, |
| 2964 | "attempted": attempted, |
| 2965 | } |
| 2966 | } |
| 2967 | |
| 2968 | |
| 2969 | def _result_outcome_artifact(source: str, result: Any) -> dict[str, Any]: |
| 2970 | """Convert a legacy ``{"error": ...}`` source result into typed status.""" |
| 2971 | if not isinstance(result, dict) or not result.get("error"): |
| 2972 | return {} |
| 2973 | detail = str(result["error"]) |
| 2974 | if source == "reddit": |
| 2975 | state = reddit.classify_run_failure(detail) |
| 2976 | attempted = True |
| 2977 | elif source == "youtube": |
| 2978 | state = youtube_yt.classify_run_failure(detail) |
| 2979 | attempted = state != schema.SKIPPED_UNCONFIGURED |
| 2980 | elif source == "x": |
| 2981 | state = bird_x.classify_run_failure(detail) |
| 2982 | attempted = True |
| 2983 | elif source == "truthsocial" and detail == "Truth Social token expired": |
| 2984 | state = schema.AUTH_FAILED |
| 2985 | attempted = True |
| 2986 | elif source == "bluesky" and "network-level block" in detail.lower(): |
| 2987 | state = schema.UNREACHABLE |
| 2988 | attempted = True |
| 2989 | else: |
| 2990 | state, attempted = _classify_source_failure(SourceRunError(detail)) |
| 2991 | return _outcome_artifact(state, detail, attempted=attempted) |
| 2992 | |
| 2993 | |
| 2994 | def _legacy_artifact_outcome( |
| 2995 | source: str, |
| 2996 | artifact: Any, |
| 2997 | ) -> dict[str, Any] | None: |
| 2998 | """Map known pre-outcome artifact contracts to a typed outcome note.""" |
| 2999 | if not isinstance(artifact, dict): |
| 3000 | return None |
| 3001 | explicit = artifact.get("_source_outcome") |
| 3002 | if isinstance(explicit, dict): |
| 3003 | return explicit |
| 3004 | if source == "perplexity" and artifact.get("error"): |
| 3005 | error = str(artifact["error"]) |
| 3006 | detail = str( |
| 3007 | artifact.get("asyncErrorMessage") |
| 3008 | or artifact.get("message") |
| 3009 | or error |
| 3010 | ) |
| 3011 | state = ( |
| 3012 | health.TIMEOUT |
| 3013 | if error.lower() == "timeout" |
| 3014 | else http.classify_failure( |
| 3015 | status_code=artifact.get("statusCode"), |
| 3016 | message=f"{error}: {detail}", |
| 3017 | ) |
| 3018 | ) |
| 3019 | return _outcome_artifact(state, detail)["_source_outcome"] |
| 3020 | if ( |
| 3021 | source == "grounding" |
| 3022 | and artifact.get("reason") == "keyless-search-unavailable" |
| 3023 | ): |
| 3024 | return _outcome_artifact( |
| 3025 | schema.UNREACHABLE, |
| 3026 | "Keyless web search unavailable", |
| 3027 | )["_source_outcome"] |
| 3028 | return None |
| 3029 | |
| 3030 | |
| 3031 | def _resolve_stream_outcome( |
| 3032 | source: str, |
| 3033 | artifact: Any, |
| 3034 | failures: list[http.HTTPError], |
| 3035 | ) -> dict[str, Any] | None: |
| 3036 | """Choose the most specific artifact or captured HTTP outcome.""" |
| 3037 | artifact_outcome = _legacy_artifact_outcome(source, artifact) |
| 3038 | if not failures: |
| 3039 | return artifact_outcome |
| 3040 | # Pick the most specific failure rather than the last-appended one: |
| 3041 | # parallel workers append in nondeterministic order, and an auth failure |
| 3042 | # must not be masked by a later 429 (wrong doctor prescription). |
| 3043 | _FAILURE_SPECIFICITY = { |
| 3044 | health.AUTH_FAILED: 0, |
| 3045 | health.RATE_LIMITED: 1, |
| 3046 | health.SCHEMA_DRIFT: 2, |
| 3047 | health.TIMEOUT: 3, |
| 3048 | health.UNREACHABLE: 4, |
| 3049 | health.ERROR: 5, |
| 3050 | } |
| 3051 | failure = min( |
| 3052 | failures, |
| 3053 | key=lambda f: _FAILURE_SPECIFICITY.get(f.outcome_state, 9), |
| 3054 | ) |
| 3055 | captured_outcome = _outcome_artifact( |
| 3056 | failure.outcome_state, |
| 3057 | str(failure), |
| 3058 | )["_source_outcome"] |
| 3059 | if artifact_outcome is None: |
| 3060 | return captured_outcome |
| 3061 | if ( |
| 3062 | artifact_outcome.get("state") == health.ERROR |
| 3063 | and failure.outcome_state != health.ERROR |
| 3064 | ): |
| 3065 | return captured_outcome |
| 3066 | return artifact_outcome |
| 3067 | |
| 3068 | |
| 3069 | def _finalize_source_status( |
| 3070 | outcomes: dict[str, schema.SourceOutcome], |
| 3071 | items_by_source: dict[str, list[schema.SourceItem]], |
| 3072 | ) -> dict[str, schema.SourceOutcome]: |
| 3073 | """Sync outcome counts to the final post-filter evidence set.""" |
| 3074 | finalized: dict[str, schema.SourceOutcome] = {} |
| 3075 | for source, outcome in outcomes.items(): |
| 3076 | count = len(items_by_source.get(source, [])) |
| 3077 | state = outcome.state |
| 3078 | detail = outcome.detail |
| 3079 | fix_hint = outcome.fix_hint |
| 3080 | if state == schema.NO_RESULTS and count: |
| 3081 | state = health.OK |
| 3082 | detail = None |
| 3083 | fix_hint = None |
| 3084 | elif state == health.OK and not count: |
| 3085 | state = schema.NO_RESULTS |
| 3086 | elif state == schema.PARTIAL and not count: |
| 3087 | state = http.classify_failure(message=detail or "") |
| 3088 | finalized[source] = schema.SourceOutcome( |
| 3089 | source=source, |
| 3090 | state=state, |
| 3091 | items_returned=count, |
| 3092 | attempted=outcome.attempted, |
| 3093 | detail=detail, |
| 3094 | at=outcome.at, |
| 3095 | fix_hint=fix_hint, |
| 3096 | ) |
| 3097 | return finalized |
| 3098 | |
| 3099 | |
| 3100 | def _is_transient_error(exc: Exception) -> bool: |
| 3101 | """Detect 5xx server errors that are worth retrying.""" |
| 3102 | status = getattr(exc, "status_code", None) |
| 3103 | if isinstance(status, int) and 500 <= status < 600: |
| 3104 | return True |
| 3105 | msg = str(exc) |
| 3106 | return any(code in msg for code in ("500", "502", "503", "504")) |
| 3107 | |
| 3108 | |
| 3109 | def _run_supplemental_searches( |
| 3110 | *, |
| 3111 | topic: str, |
| 3112 | bundle: schema.RetrievalBundle, |
| 3113 | plan: schema.QueryPlan, |
| 3114 | config: dict[str, Any], |
| 3115 | depth: str, |
| 3116 | date_range: tuple[str, str], |
| 3117 | runtime: schema.ProviderRuntime, |
| 3118 | mock: bool, |
| 3119 | rate_limited_sources: set[str], |
| 3120 | rate_limit_lock: threading.Lock, |
| 3121 | x_handle: str | None = None, |
| 3122 | x_related: list[str] | None = None, |
| 3123 | ) -> None: |
| 3124 | """Phase 2: extract entities from Phase 1 results, run targeted supplemental searches.""" |
| 3125 | if depth == "quick" or mock: |
| 3126 | return |
| 3127 | |
| 3128 | from_date, to_date = date_range |
| 3129 | |
| 3130 | # Convert SourceItems to dicts for entity_extract. All X items (whatever |
| 3131 | # backend fetched them — bird, xai, xurl, xquik) land under the single "x" |
| 3132 | # slug, so this reads the whole X corpus. |
| 3133 | x_dicts = [ |
| 3134 | {"author_handle": item.author or "", "text": item.body or ""} |
| 3135 | for item in bundle.items_by_source.get("x", []) |
| 3136 | ] |
| 3137 | reddit_dicts = [ |
| 3138 | { |
| 3139 | "subreddit": item.container or "", |
| 3140 | "comment_insights": item.metadata.get("comment_insights", []), |
| 3141 | "top_comments": [ |
| 3142 | {"excerpt": c.get("excerpt", c.get("text", ""))} |
| 3143 | for c in (item.metadata.get("top_comments") or []) |
| 3144 | if isinstance(c, dict) |
| 3145 | ], |
| 3146 | } |
| 3147 | for item in bundle.items_by_source.get("reddit", []) |
| 3148 | ] |
| 3149 | |
| 3150 | if not x_dicts and not reddit_dicts and not x_handle and not x_related: |
| 3151 | return |
| 3152 | |
| 3153 | entities = entity_extract.extract_entities( |
| 3154 | reddit_dicts, x_dicts, |
| 3155 | max_handles=3, max_subreddits=3, |
| 3156 | ) |
| 3157 | |
| 3158 | handles = entities.get("x_handles", []) |
| 3159 | |
| 3160 | # Add explicit --x-handle if provided |
| 3161 | if x_handle: |
| 3162 | handle_clean = x_handle.lstrip("@").lower() |
| 3163 | if handle_clean not in [h.lower() for h in handles]: |
| 3164 | handles.insert(0, handle_clean) |
| 3165 | |
| 3166 | # Collect related handles (searched separately with lower weight) |
| 3167 | related_handles = [] |
| 3168 | if x_related: |
| 3169 | primary_lower = x_handle.lstrip("@").lower() if x_handle else "" |
| 3170 | for rh in x_related: |
| 3171 | rh_clean = rh.lstrip("@").lower().strip() |
| 3172 | if rh_clean and rh_clean != primary_lower and rh_clean not in [h.lower() for h in handles]: |
| 3173 | related_handles.append(rh_clean) |
| 3174 | |
| 3175 | if not handles and not related_handles: |
| 3176 | return |
| 3177 | |
| 3178 | # Pick the X handle-search backend: the first handle-capable backend in the |
| 3179 | # chain (bird or xquik). These supplemental from:/mentions lanes are |
| 3180 | # complementary to the topic search, so when the topic primary can't run |
| 3181 | # them (xai/xurl have no handle-lane implementation) but a capable backend |
| 3182 | # is available, use it rather than skipping Phase 2. bird scrapes X GraphQL |
| 3183 | # with the user's browser cookies; xquik runs the same lanes over its REST |
| 3184 | # API. All items land under the single "x" slug. |
| 3185 | x_slug = "x" |
| 3186 | chain = env.x_backend_chain(config) |
| 3187 | # Trust an explicit runtime backend as the head of the chain. |
| 3188 | pinned = runtime.x_search_backend |
| 3189 | if pinned: |
| 3190 | chain = [pinned] + [b for b in chain if b != pinned] |
| 3191 | primary = next((b for b in chain if b in ("bird", "xquik")), None) |
| 3192 | |
| 3193 | if primary == "bird": |
| 3194 | def _from_lane(hs: list, count: int) -> list: |
| 3195 | return bird_x.search_handles(hs, topic, from_date, count_per=count) |
| 3196 | |
| 3197 | def _about_lane(hs: list, count: int) -> list: |
| 3198 | return bird_x.search_mentions(hs, from_date, count_per=count) |
| 3199 | elif primary == "xquik": |
| 3200 | xquik_token = env.get_xquik_token(config) |
| 3201 | |
| 3202 | def _from_lane(hs: list, count: int) -> list: |
| 3203 | return xquik.search_handles(hs, topic, from_date, to_date, count_per=count, token=xquik_token) |
| 3204 | |
| 3205 | def _about_lane(hs: list, count: int) -> list: |
| 3206 | return xquik.search_mentions(hs, from_date, to_date, topic=topic, count_per=count, token=xquik_token) |
| 3207 | else: |
| 3208 | return # primary X backend has no handle-lane support (xai/xurl) or none configured |
| 3209 | |
| 3210 | # Skip if the X source is rate-limited. |
| 3211 | if x_slug in rate_limited_sources: |
| 3212 | return |
| 3213 | |
| 3214 | # Collect existing URLs for deduplication |
| 3215 | existing_urls = { |
| 3216 | item.url |
| 3217 | for items in bundle.items_by_source.values() |
| 3218 | for item in items |
| 3219 | if item.url |
| 3220 | } |
| 3221 | |
| 3222 | ranking_query = plan.subqueries[0].ranking_query if plan.subqueries else topic |
| 3223 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 3224 | |
| 3225 | # Search primary handles (full weight): FROM lane (their own tweets) + |
| 3226 | # ABOUT lane (tweets mentioning them). Both engagement-weighted and deduped |
| 3227 | # by URL at normalize time. |
| 3228 | if handles: |
| 3229 | # Independent try/except per lane so a failure in one does not discard |
| 3230 | # the other's already-computed results. |
| 3231 | from_items: list = [] |
| 3232 | about_items: list = [] |
| 3233 | try: |
| 3234 | from_items = _from_lane(handles, FROM_LANE_COUNT_PER) |
| 3235 | except Exception as exc: |
| 3236 | print(f"[Pipeline] Phase 2 FROM-lane search failed: {exc}", file=sys.stderr) |
| 3237 | state, attempted = _classify_source_failure(exc) |
| 3238 | bundle.record_failure( |
| 3239 | x_slug, |
| 3240 | state, |
| 3241 | f"Phase 2 FROM-lane: {exc}", |
| 3242 | attempted=attempted, |
| 3243 | ) |
| 3244 | if not bundle.items_by_source.get(x_slug): |
| 3245 | bundle.errors_by_source[x_slug] = f"Phase 2 FROM-lane: {exc}" |
| 3246 | try: |
| 3247 | about_items = _about_lane(handles, MENTION_LANE_COUNT_PER) |
| 3248 | except Exception as exc: |
| 3249 | print(f"[Pipeline] Phase 2 ABOUT-lane search failed: {exc}", file=sys.stderr) |
| 3250 | state, attempted = _classify_source_failure(exc) |
| 3251 | bundle.record_failure( |
| 3252 | x_slug, |
| 3253 | state, |
| 3254 | f"Phase 2 ABOUT-lane: {exc}", |
| 3255 | attempted=attempted, |
| 3256 | ) |
| 3257 | raw_items = from_items + about_items |
| 3258 | |
| 3259 | if raw_items: |
| 3260 | normalized = _normalize_score_dedupe( |
| 3261 | x_slug, raw_items, from_date, to_date, |
| 3262 | freshness_mode=plan.freshness_mode, |
| 3263 | ranking_query=ranking_query, |
| 3264 | ) |
| 3265 | # Deduplicate against Phase 1 URLs |
| 3266 | normalized = [item for item in normalized if item.url not in existing_urls] |
| 3267 | if normalized: |
| 3268 | bundle.add_items(primary_label, x_slug, normalized) |
| 3269 | # Update existing URLs for related-handle dedup |
| 3270 | for item in normalized: |
| 3271 | if item.url: |
| 3272 | existing_urls.add(item.url) |
| 3273 | |
| 3274 | # Search related handles with lower weight (0.3) |
| 3275 | if related_handles: |
| 3276 | try: |
| 3277 | raw_items = _from_lane(related_handles, RELATED_HANDLE_COUNT_PER) |
| 3278 | except Exception as exc: |
| 3279 | print(f"[Pipeline] Phase 2 related handle search failed: {exc}", file=sys.stderr) |
| 3280 | state, attempted = _classify_source_failure(exc) |
| 3281 | bundle.record_failure( |
| 3282 | x_slug, |
| 3283 | state, |
| 3284 | f"Phase 2 related handle search: {exc}", |
| 3285 | attempted=attempted, |
| 3286 | ) |
| 3287 | raw_items = [] |
| 3288 | |
| 3289 | if raw_items: |
| 3290 | normalized = _normalize_score_dedupe( |
| 3291 | x_slug, raw_items, from_date, to_date, |
| 3292 | freshness_mode=plan.freshness_mode, |
| 3293 | ranking_query=ranking_query, |
| 3294 | ) |
| 3295 | # Deduplicate against all existing URLs (Phase 1 + primary handles) |
| 3296 | normalized = [item for item in normalized if item.url not in existing_urls] |
| 3297 | if normalized: |
| 3298 | # Use a separate subquery label with lower weight so RRF |
| 3299 | # scores related-handle results below primary results. |
| 3300 | bundle.add_items("supplemental-related", x_slug, normalized) |
| 3301 | # Register the supplemental-related label in the plan for fusion |
| 3302 | if not any(sq.label == "supplemental-related" for sq in plan.subqueries): |
| 3303 | plan.subqueries.append( |
| 3304 | schema.SubQuery( |
| 3305 | label="supplemental-related", |
| 3306 | search_query=", ".join(related_handles), |
| 3307 | ranking_query=ranking_query, |
| 3308 | sources=[x_slug], |
| 3309 | weight=0.3, |
| 3310 | ) |
| 3311 | ) |
| 3312 | |
| 3313 | |
| 3314 | def _retry_thin_sources( |
| 3315 | *, |
| 3316 | topic: str, |
| 3317 | bundle: schema.RetrievalBundle, |
| 3318 | plan: schema.QueryPlan, |
| 3319 | config: dict[str, Any], |
| 3320 | depth: str, |
| 3321 | date_range: tuple[str, str], |
| 3322 | runtime: schema.ProviderRuntime, |
| 3323 | mock: bool, |
| 3324 | rate_limited_sources: set[str], |
| 3325 | rate_limit_lock: threading.Lock, |
| 3326 | settings: dict[str, Any], |
| 3327 | web_backend: str = "auto", |
| 3328 | skip_sources: set[str] | None = None, |
| 3329 | subreddits: list[str] | None = None, |
| 3330 | tiktok_hashtags: list[str] | None = None, |
| 3331 | tiktok_creators: list[str] | None = None, |
| 3332 | ig_creators: list[str] | None = None, |
| 3333 | ) -> None: |
| 3334 | """Retry sources with thin results using simplified core subject query.""" |
| 3335 | if depth == "quick": |
| 3336 | return |
| 3337 | |
| 3338 | planned_sources: list[str] = [] |
| 3339 | for subquery in plan.subqueries: |
| 3340 | for source in subquery.sources: |
| 3341 | if source not in planned_sources: |
| 3342 | planned_sources.append(source) |
| 3343 | # trustpilot returns at most ONE item by design, so the "<3 items" rule |
| 3344 | # would re-fetch it after every successful lookup -- bypassing |
| 3345 | # MAX_SOURCE_FETCHES and re-resolving WITHOUT the caller's |
| 3346 | # --trustpilot-domain (a lookalike-misattribution path). Its thin result |
| 3347 | # is its normal success state; never retry it here. |
| 3348 | _skip = (skip_sources or set()) | {"trustpilot"} |
| 3349 | thin_sources = [ |
| 3350 | source |
| 3351 | for source in planned_sources |
| 3352 | if len(bundle.items_by_source.get(source, [])) < 3 |
| 3353 | and source not in bundle.errors_by_source |
| 3354 | and source not in _skip |
| 3355 | ] |
| 3356 | |
| 3357 | if not thin_sources: |
| 3358 | return |
| 3359 | |
| 3360 | core = query.extract_core_subject(topic, max_words=3) |
| 3361 | if not core: |
| 3362 | return |
| 3363 | # Note: we intentionally do NOT skip when core == topic. For short topics |
| 3364 | # like "Kanye West", the 3-word core IS the topic — but the planner may |
| 3365 | # have sent a different (worse) query to the source. Retrying with the |
| 3366 | # raw core subject is still valuable. |
| 3367 | |
| 3368 | from_date, to_date = date_range |
| 3369 | |
| 3370 | # Create a retry subquery with the simplified core subject |
| 3371 | retry_subquery = schema.SubQuery( |
| 3372 | label="retry", |
| 3373 | search_query=core, |
| 3374 | ranking_query=f"What recent evidence from the last 30 days matters for {core}?", |
| 3375 | sources=thin_sources, |
| 3376 | weight=0.3, |
| 3377 | ) |
| 3378 | |
| 3379 | def _retry_one_source( |
| 3380 | source: str, |
| 3381 | ) -> tuple[str, list[schema.SourceItem], dict[str, Any] | None]: |
| 3382 | raw_items, artifact = _retrieve_stream( |
| 3383 | topic=topic, |
| 3384 | subquery=retry_subquery, |
| 3385 | source=source, |
| 3386 | config=config, |
| 3387 | depth=depth, |
| 3388 | date_range=date_range, |
| 3389 | runtime=runtime, |
| 3390 | mock=mock, |
| 3391 | rate_limited_sources=rate_limited_sources, |
| 3392 | rate_limit_lock=rate_limit_lock, |
| 3393 | web_backend=web_backend, |
| 3394 | raw_topic=topic, |
| 3395 | subreddits=subreddits, |
| 3396 | tiktok_hashtags=tiktok_hashtags, |
| 3397 | tiktok_creators=tiktok_creators, |
| 3398 | ig_creators=ig_creators, |
| 3399 | ) |
| 3400 | outcome_note = artifact.get("_source_outcome") if isinstance(artifact, dict) else None |
| 3401 | normalized = _normalize_score_dedupe( |
| 3402 | source, |
| 3403 | raw_items, |
| 3404 | from_date, |
| 3405 | to_date, |
| 3406 | freshness_mode=plan.freshness_mode, |
| 3407 | ranking_query=retry_subquery.ranking_query, |
| 3408 | ) |
| 3409 | if source == "jobs": |
| 3410 | return source, normalized, outcome_note |
| 3411 | return source, normalized[:settings["per_stream_limit"]], outcome_note |
| 3412 | |
| 3413 | retryable = [s for s in thin_sources if s not in rate_limited_sources] |
| 3414 | |
| 3415 | from concurrent.futures import ThreadPoolExecutor, as_completed |
| 3416 | with ThreadPoolExecutor(max_workers=min(4, len(retryable) or 1)) as executor: |
| 3417 | futures = {executor.submit(_retry_one_source, s): s for s in retryable} |
| 3418 | for future in as_completed(futures): |
| 3419 | source = futures[future] |
| 3420 | try: |
| 3421 | source, normalized, outcome_note = future.result() |
| 3422 | if outcome_note: |
| 3423 | bundle.record_failure( |
| 3424 | source, |
| 3425 | outcome_note["state"], |
| 3426 | outcome_note["detail"], |
| 3427 | attempted=outcome_note.get("attempted", True), |
| 3428 | ) |
| 3429 | existing_urls = {item.url for item in bundle.items_by_source.get(source, []) if item.url} |
| 3430 | new_items = [item for item in normalized if item.url not in existing_urls] |
| 3431 | |
| 3432 | if new_items: |
| 3433 | primary_label = plan.subqueries[0].label if plan.subqueries else "primary" |
| 3434 | bundle.add_items(primary_label, source, new_items) |
| 3435 | except Exception as exc: |
| 3436 | print(f"[Pipeline] Retry failed for {source}: {type(exc).__name__}: {exc}", file=sys.stderr) |
| 3437 | state, attempted = _classify_source_failure(exc) |
| 3438 | bundle.record_failure( |
| 3439 | source, |
| 3440 | state, |
| 3441 | f"Simplified-query retry failed: {exc}", |
| 3442 | attempted=attempted, |
| 3443 | ) |
| 3444 | |
| 3445 | |
| 3446 | def _fetch_x_backend(backend, subquery, from_date, to_date, depth, config): |
| 3447 | """Fetch X items from a single backend. Returns (items, error_str). |
| 3448 | |
| 3449 | Backends are tried in priority order by the caller (env.x_backend_chain); |
| 3450 | a non-empty error_str signals a hard failure (auth/payment/etc.) so the |
| 3451 | caller can fail over to the next backend or surface the error honestly. |
| 3452 | """ |
| 3453 | query = subquery.search_query |
| 3454 | if backend == "bird": |
| 3455 | result = bird_x.search_x(query, from_date, to_date, depth=depth) |
| 3456 | items = bird_x.parse_bird_response(result, query=query) |
| 3457 | elif backend == "xai": |
| 3458 | model = config.get("LAST30DAYS_X_MODEL") or config.get("XAI_MODEL_PIN") or providers.XAI_DEFAULT |
| 3459 | result = xai_x.search_x(config["XAI_API_KEY"], model, query, from_date, to_date, depth=depth) |
| 3460 | items = xai_x.parse_x_response(result) |
| 3461 | elif backend == "xurl": |
| 3462 | result = xurl_x.search_x(query, depth=depth) |
| 3463 | items = xurl_x.parse_x_response(result, topic=query) |
| 3464 | elif backend == "xquik": |
| 3465 | result = xquik.search_xquik(query, from_date, to_date, depth=depth, token=env.get_xquik_token(config)) |
| 3466 | items = xquik.parse_xquik_response(result) |
| 3467 | else: |
| 3468 | return [], f"unknown X backend: {backend}" |
| 3469 | err = result.get("error") if isinstance(result, dict) else "" |
| 3470 | return items, (err or "") |
| 3471 | |
| 3472 | |
| 3473 | def _reddit_post_key(item: dict) -> str: |
| 3474 | """Stable per-thread dedupe key (base36 post id from the url/permalink).""" |
| 3475 | url = item.get("url") or item.get("permalink") or "" |
| 3476 | m = re.search(r"/comments/([A-Za-z0-9]+)", url) |
| 3477 | return m.group(1) if m else url |
| 3478 | |
| 3479 | |
| 3480 | def _merge_reddit_items(free: list[dict], sc: list[dict]) -> list[dict]: |
| 3481 | """Merge free + ScrapeCreators Reddit items, free first, deduped by post id. |
| 3482 | |
| 3483 | Used when the thinness-floor trigger backfills a thin free run with SC, so a |
| 3484 | thread present in both is never double-listed. |
| 3485 | """ |
| 3486 | merged = list(free) |
| 3487 | seen = {_reddit_post_key(it) for it in free} |
| 3488 | for it in sc: |
| 3489 | key = _reddit_post_key(it) |
| 3490 | if key and key not in seen: |
| 3491 | seen.add(key) |
| 3492 | merged.append(it) |
| 3493 | return merged |
| 3494 | |
| 3495 | |
| 3496 | def _retrieve_stream(*args, **kwargs) -> tuple[list[dict], dict]: |
| 3497 | """Run one stream and retain HTTP failures swallowed by source adapters.""" |
| 3498 | source = str(kwargs.get("source") or "") |
| 3499 | fixture_request = { |
| 3500 | "source": source, |
| 3501 | "topic": kwargs.get("topic") or "", |
| 3502 | "search_query": getattr(kwargs.get("subquery"), "search_query", ""), |
| 3503 | "date_range": list(kwargs.get("date_range") or ()), |
| 3504 | "depth": kwargs.get("depth") or "", |
| 3505 | } |
| 3506 | module_backed = source in { |
| 3507 | "reddit", |
| 3508 | "x", |
| 3509 | "youtube", |
| 3510 | "stocktwits", |
| 3511 | "digg", |
| 3512 | "arxiv", |
| 3513 | "techmeme", |
| 3514 | "trustpilot", |
| 3515 | "github", |
| 3516 | } |
| 3517 | if module_backed: |
| 3518 | matched, replayed = http.fixture_source_replay(fixture_request) |
| 3519 | if matched: |
| 3520 | return replayed[0], replayed[1] |
| 3521 | try: |
| 3522 | with http.capture_failures() as failures, \ |
| 3523 | http.fixture_module_capture(module_backed): |
| 3524 | items, artifact = _retrieve_stream_impl(*args, **kwargs) |
| 3525 | except Exception as exc: |
| 3526 | recorded_exc = exc |
| 3527 | if failures and not getattr(exc, "outcome_state", None): |
| 3528 | failure = failures[-1] |
| 3529 | recorded_exc = SourceRunError(str(exc), failure.outcome_state) |
| 3530 | if module_backed: |
| 3531 | http.fixture_source_record_error(fixture_request, recorded_exc) |
| 3532 | if recorded_exc is not exc: |
| 3533 | raise recorded_exc from exc |
| 3534 | raise |
| 3535 | outcome_note = _resolve_stream_outcome( |
| 3536 | str(kwargs.get("source") or ""), |
| 3537 | artifact, |
| 3538 | failures, |
| 3539 | ) |
| 3540 | if outcome_note: |
| 3541 | artifact = dict(artifact or {}) |
| 3542 | artifact["_source_outcome"] = outcome_note |
| 3543 | if module_backed: |
| 3544 | http.fixture_source_record(fixture_request, [items, artifact]) |
| 3545 | return items, artifact |
| 3546 | |
| 3547 | |
| 3548 | def _retrieve_stream_impl( |
| 3549 | *, |
| 3550 | topic: str, |
| 3551 | subquery: schema.SubQuery, |
| 3552 | source: str, |
| 3553 | config: dict[str, Any], |
| 3554 | depth: str, |
| 3555 | date_range: tuple[str, str], |
| 3556 | runtime: schema.ProviderRuntime, |
| 3557 | mock: bool, |
| 3558 | rate_limited_sources: set[str] | None = None, |
| 3559 | rate_limit_lock: threading.Lock | None = None, |
| 3560 | web_backend: str = "auto", |
| 3561 | raw_topic: str = "", |
| 3562 | subreddits: list[str] | None = None, |
| 3563 | tiktok_hashtags: list[str] | None = None, |
| 3564 | tiktok_creators: list[str] | None = None, |
| 3565 | ig_creators: list[str] | None = None, |
| 3566 | trustpilot_domain: str | None = None, |
| 3567 | trustpilot_domain_is_hint: bool = False, |
| 3568 | ) -> tuple[list[dict], dict]: |
| 3569 | # Early exit if source was rate-limited by a sibling future |
| 3570 | if rate_limited_sources is not None and source in rate_limited_sources: |
| 3571 | return [], {} |
| 3572 | from_date, to_date = date_range |
| 3573 | if mock: |
| 3574 | return _mock_stream_results(source, subquery) |
| 3575 | if source == "grounding": |
| 3576 | return grounding.web_search( |
| 3577 | subquery.search_query, date_range, config, backend=web_backend) |
| 3578 | if source == "jobs": |
| 3579 | return jobs.search_jobs( |
| 3580 | raw_topic or topic or subquery.search_query, |
| 3581 | date_range, |
| 3582 | config, |
| 3583 | depth=depth, |
| 3584 | web_backend=web_backend, |
| 3585 | explicit=bool(config.get("_hiring_signals_mode")), |
| 3586 | ) |
| 3587 | if source == "reddit": |
| 3588 | # Use raw_topic so expand_reddit_queries() generates diverse variants |
| 3589 | # from the original user topic, not the planner's narrowed search_query. |
| 3590 | reddit_query = raw_topic or subquery.search_query |
| 3591 | dedicated_subreddits = config.get("_dedicated_subreddits") or None |
| 3592 | has_sc_key = bool(config.get("SCRAPECREATORS_API_KEY")) |
| 3593 | sc_first = ( |
| 3594 | has_sc_key |
| 3595 | and (config.get(env.REDDIT_BACKEND_PIN_VAR) or "").lower() |
| 3596 | == "scrapecreators" |
| 3597 | ) |
| 3598 | if sc_first: |
| 3599 | # env.REDDIT_BACKEND_PIN_VAR=scrapecreators: SC primary, public fallback |
| 3600 | primary_failure: Exception | None = None |
| 3601 | try: |
| 3602 | result = reddit.search_and_enrich( |
| 3603 | reddit_query, from_date, to_date, depth=depth, |
| 3604 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 3605 | subreddits=subreddits, |
| 3606 | ) |
| 3607 | items = reddit.parse_reddit_response(result) |
| 3608 | if items: |
| 3609 | return items, {} |
| 3610 | sys.stderr.write( |
| 3611 | "[Reddit] ScrapeCreators primary returned no items, " |
| 3612 | "using public fallback\n" |
| 3613 | ) |
| 3614 | except Exception as exc: |
| 3615 | primary_failure = exc |
| 3616 | sys.stderr.write( |
| 3617 | f"[Reddit] ScrapeCreators primary failed " |
| 3618 | f"({type(exc).__name__}: {exc}), using public fallback\n" |
| 3619 | ) |
| 3620 | public_failure: Exception | None = None |
| 3621 | try: |
| 3622 | public_results = reddit_public.search_reddit_public( |
| 3623 | reddit_query, from_date, to_date, depth=depth, |
| 3624 | subreddits=subreddits, |
| 3625 | ) |
| 3626 | if public_results: |
| 3627 | if primary_failure is not None: |
| 3628 | state = reddit.classify_run_failure(str(primary_failure)) |
| 3629 | return public_results, _outcome_artifact( |
| 3630 | state, |
| 3631 | f"Reddit primary failed; public fallback returned " |
| 3632 | f"{len(public_results)} items: {primary_failure}", |
| 3633 | ) |
| 3634 | return public_results, {} |
| 3635 | sys.stderr.write( |
| 3636 | "[Reddit] Public fallback returned no items after " |
| 3637 | "ScrapeCreators primary miss\n" |
| 3638 | ) |
| 3639 | except Exception as exc: |
| 3640 | public_failure = exc |
| 3641 | sys.stderr.write( |
| 3642 | f"[Reddit] Public fallback also failed " |
| 3643 | f"({type(exc).__name__}: {exc})\n" |
| 3644 | ) |
| 3645 | failure = public_failure or primary_failure |
| 3646 | if failure is not None: |
| 3647 | state = reddit.classify_run_failure(str(failure)) |
| 3648 | raise SourceRunError( |
| 3649 | f"Reddit primary and fallback produced no results after failure: {failure}", |
| 3650 | state, |
| 3651 | ) |
| 3652 | return [], {} |
| 3653 | |
| 3654 | # Default: public Reddit first (free). ScrapeCreators backfills when the |
| 3655 | # free path is empty OR returns fewer than the configured thinness floor |
| 3656 | # (env.REDDIT_SC_MIN_ITEMS_VAR, default 0 = empty-only — today's |
| 3657 | # behavior, no extra credit spend unless the user opts in). |
| 3658 | try: |
| 3659 | min_items = int(config.get(env.REDDIT_SC_MIN_ITEMS_VAR) or 0) |
| 3660 | except (TypeError, ValueError): |
| 3661 | min_items = 0 |
| 3662 | public_results: list[dict] = [] |
| 3663 | public_failure: Exception | None = None |
| 3664 | try: |
| 3665 | public_results = reddit_public.search_reddit_public( |
| 3666 | reddit_query, from_date, to_date, depth=depth, |
| 3667 | subreddits=subreddits, dedicated_subreddits=dedicated_subreddits, |
| 3668 | ) or [] |
| 3669 | except Exception as exc: |
| 3670 | public_failure = exc |
| 3671 | sys.stderr.write( |
| 3672 | f"[Reddit] Public search failed ({type(exc).__name__}: {exc})" |
| 3673 | ) |
| 3674 | if not has_sc_key: |
| 3675 | sys.stderr.write("\n") |
| 3676 | state = reddit.classify_run_failure(str(exc)) |
| 3677 | raise SourceRunError(f"Reddit public search failed: {exc}", state) from exc |
| 3678 | sys.stderr.write(", using ScrapeCreators backup\n") |
| 3679 | # Enough free results, or no key to backfill with -> done. max(min_items, |
| 3680 | # 1) keeps the default (min_items=0) as empty-only AND treats exactly |
| 3681 | # `min_items` results as acceptable (no backfill) for min_items > 0. |
| 3682 | if len(public_results) >= max(min_items, 1) or not has_sc_key: |
| 3683 | return public_results, {} |
| 3684 | if public_results: |
| 3685 | sys.stderr.write( |
| 3686 | f"[Reddit] Free path returned {len(public_results)} " |
| 3687 | f"(below the {min_items}-item floor); backfilling with ScrapeCreators\n" |
| 3688 | ) |
| 3689 | try: |
| 3690 | result = reddit.search_and_enrich( |
| 3691 | reddit_query, from_date, to_date, depth=depth, |
| 3692 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 3693 | subreddits=subreddits, |
| 3694 | ) |
| 3695 | sc_items = reddit.parse_reddit_response(result) |
| 3696 | except Exception as exc: |
| 3697 | sys.stderr.write( |
| 3698 | f"[Reddit] ScrapeCreators backup also failed " |
| 3699 | f"({type(exc).__name__}: {exc})\n" |
| 3700 | ) |
| 3701 | state = reddit.classify_run_failure(str(exc)) |
| 3702 | return public_results, _outcome_artifact( |
| 3703 | state, |
| 3704 | f"Reddit backup failed after {len(public_results)} public items: {exc}", |
| 3705 | ) |
| 3706 | merged = _merge_reddit_items(public_results, sc_items) |
| 3707 | if public_failure is not None: |
| 3708 | state = reddit.classify_run_failure(str(public_failure)) |
| 3709 | return merged, _outcome_artifact( |
| 3710 | state, |
| 3711 | f"Reddit public search failed; backup returned {len(sc_items)} items: " |
| 3712 | f"{public_failure}", |
| 3713 | ) |
| 3714 | return merged, {} |
| 3715 | if source == "x": |
| 3716 | # One X source, an ordered chain of interchangeable backends. Try the |
| 3717 | # primary; fall through to the next only if it returns nothing or errors. |
| 3718 | chain = env.x_backend_chain(config) |
| 3719 | # Trust an explicit runtime backend as the primary (already resolved as |
| 3720 | # available), keeping the rest of the chain as failover backups. |
| 3721 | pinned = runtime.x_search_backend |
| 3722 | if pinned: |
| 3723 | chain = [pinned] + [b for b in chain if b != pinned] |
| 3724 | if not chain: |
| 3725 | raise RuntimeError("No X backend is available.") |
| 3726 | last_error = "" |
| 3727 | for i, backend in enumerate(chain): |
| 3728 | items, err = _fetch_x_backend(backend, subquery, from_date, to_date, depth, config) |
| 3729 | if items: |
| 3730 | if i > 0: |
| 3731 | print(f"[X] primary backend(s) returned nothing; used fallback '{backend}'", file=sys.stderr) |
| 3732 | if last_error: |
| 3733 | state = ( |
| 3734 | bird_x.classify_run_failure(last_error) |
| 3735 | if last_error.startswith("bird:") |
| 3736 | else http.classify_failure(message=last_error) |
| 3737 | ) |
| 3738 | return items, _outcome_artifact( |
| 3739 | state, |
| 3740 | f"X fallback '{backend}' returned {len(items)} items after {last_error}", |
| 3741 | ) |
| 3742 | return items, {} |
| 3743 | if err: |
| 3744 | last_error = f"{backend}: {err}" |
| 3745 | print(f"[X] backend '{backend}' failed ({err}); trying next", file=sys.stderr) |
| 3746 | if last_error: |
| 3747 | state = ( |
| 3748 | bird_x.classify_run_failure(last_error) |
| 3749 | if last_error.startswith("bird:") |
| 3750 | else http.classify_failure(message=last_error) |
| 3751 | ) |
| 3752 | raise SourceRunError(f"All X backends failed — {last_error}", state) |
| 3753 | return [], {} |
| 3754 | if source == "youtube": |
| 3755 | # Use raw_topic so expand_youtube_queries() generates diverse variants |
| 3756 | # from the original user topic, not the planner's narrowed search_query. |
| 3757 | yt_query = raw_topic or subquery.search_query |
| 3758 | result = None |
| 3759 | youtube_failure: str | None = None |
| 3760 | # ScrapeCreators key (when present) is the default-on backup tier: it |
| 3761 | # powers the per-video transcript fallback, the SC search fallback, and |
| 3762 | # comment enrichment. None when no key, which keeps everything keyless. |
| 3763 | sc_token = ( |
| 3764 | config.get("SCRAPECREATORS_API_KEY", "") |
| 3765 | if env.is_youtube_sc_available(config) else None |
| 3766 | ) |
| 3767 | # Try yt-dlp first; the SC transcript fallback covers per-video failures. |
| 3768 | if which("yt-dlp"): |
| 3769 | try: |
| 3770 | result = youtube_yt.search_and_transcribe( |
| 3771 | yt_query, from_date, to_date, depth=depth, token=sc_token, |
| 3772 | ) |
| 3773 | if result.get("error"): |
| 3774 | youtube_failure = str(result["error"]) |
| 3775 | except Exception as exc: |
| 3776 | youtube_failure = str(exc) |
| 3777 | result = None |
| 3778 | # Fall back to SC YouTube search if yt-dlp failed or isn't installed. |
| 3779 | if (result is None or not result.get("items")) and sc_token: |
| 3780 | try: |
| 3781 | result = youtube_yt.search_youtube_sc( |
| 3782 | yt_query, from_date, to_date, depth=depth, token=sc_token, |
| 3783 | ) |
| 3784 | if result.get("error"): |
| 3785 | youtube_failure = str(result["error"]) |
| 3786 | except Exception as exc: |
| 3787 | youtube_failure = str(exc) |
| 3788 | result = None |
| 3789 | if result is None: |
| 3790 | result = {"items": []} |
| 3791 | # Enrich top videos with comments (default-on when a key is present). |
| 3792 | items = youtube_yt.parse_youtube_response(result) |
| 3793 | if items and env.is_youtube_comments_available(config): |
| 3794 | youtube_yt.enrich_with_comments( |
| 3795 | items, token=config.get("SCRAPECREATORS_API_KEY", ""), |
| 3796 | ) |
| 3797 | if youtube_failure: |
| 3798 | state = youtube_yt.classify_run_failure(youtube_failure) |
| 3799 | attempted = state != schema.SKIPPED_UNCONFIGURED |
| 3800 | return items, _outcome_artifact(state, youtube_failure, attempted=attempted) |
| 3801 | return items, {} |
| 3802 | if source == "tiktok": |
| 3803 | # Use raw_topic so expand_tiktok_queries() generates diverse variants |
| 3804 | # from the original user topic, not the planner's narrowed search_query. |
| 3805 | tiktok_query = raw_topic or subquery.search_query |
| 3806 | result = tiktok.search_and_enrich( |
| 3807 | tiktok_query, |
| 3808 | from_date, |
| 3809 | to_date, |
| 3810 | depth=depth, |
| 3811 | token=env.get_tiktok_token(config), |
| 3812 | hashtags=tiktok_hashtags, |
| 3813 | creators=tiktok_creators, |
| 3814 | ) |
| 3815 | items = tiktok.parse_tiktok_response(result) |
| 3816 | if items and env.is_tiktok_comments_available(config): |
| 3817 | sc_token = config.get("SCRAPECREATORS_API_KEY", "") |
| 3818 | tiktok.enrich_with_comments(items, token=sc_token) |
| 3819 | return items, _result_outcome_artifact(source, result) |
| 3820 | if source == "instagram": |
| 3821 | # Use raw_topic so expand_instagram_queries() generates diverse variants |
| 3822 | # from the original user topic, not the planner's narrowed search_query. |
| 3823 | ig_query = raw_topic or subquery.search_query |
| 3824 | result = instagram.search_and_enrich( |
| 3825 | ig_query, |
| 3826 | from_date, |
| 3827 | to_date, |
| 3828 | depth=depth, |
| 3829 | token=env.get_instagram_token(config), |
| 3830 | ig_creators=ig_creators, |
| 3831 | ) |
| 3832 | items = instagram.parse_instagram_response(result) |
| 3833 | if items and env.is_instagram_comments_available(config): |
| 3834 | instagram.enrich_with_comments( |
| 3835 | items, token=config.get("SCRAPECREATORS_API_KEY", ""), |
| 3836 | ) |
| 3837 | return items, _result_outcome_artifact(source, result) |
| 3838 | if source == "linkedin": |
| 3839 | token = config.get("SCRAPECREATORS_API_KEY", "") |
| 3840 | result = linkedin.search_linkedin( |
| 3841 | subquery.search_query, |
| 3842 | from_date, |
| 3843 | to_date, |
| 3844 | depth=depth, |
| 3845 | token=token, |
| 3846 | ) |
| 3847 | items = linkedin.parse_linkedin_response( |
| 3848 | result, from_date=from_date, to_date=to_date |
| 3849 | ) |
| 3850 | # Articles never appear in post search — surface them (high signal) |
| 3851 | # via a bounded profile-enrichment lane on person topics. |
| 3852 | items += linkedin.enrich_articles( |
| 3853 | items, raw_topic or topic, token, from_date=from_date, to_date=to_date |
| 3854 | ) |
| 3855 | return items, _result_outcome_artifact(source, result) |
| 3856 | if source == "hackernews": |
| 3857 | result = hackernews.search_hackernews(subquery.search_query, from_date, to_date, depth=depth) |
| 3858 | return ( |
| 3859 | hackernews.parse_hackernews_response(result, query=subquery.search_query), |
| 3860 | _result_outcome_artifact(source, result), |
| 3861 | ) |
| 3862 | if source == "stocktwits": |
| 3863 | # Pass raw_topic so symbol detection sees the full topic, not the |
| 3864 | # narrowed per-subquery search_query (same rationale as reddit). |
| 3865 | result = stocktwits.search_stocktwits( |
| 3866 | raw_topic or topic or subquery.search_query, from_date, to_date, depth=depth) |
| 3867 | return ( |
| 3868 | stocktwits.parse_stocktwits_response(result, query=subquery.search_query), |
| 3869 | _result_outcome_artifact(source, result), |
| 3870 | ) |
| 3871 | if source == "dripstack": |
| 3872 | result = dripstack.search_dripstack( |
| 3873 | subquery.search_query, from_date, to_date, depth=depth) |
| 3874 | relevance_topic = raw_topic or topic or subquery.search_query |
| 3875 | return ( |
| 3876 | dripstack.parse_dripstack_response(result, query=relevance_topic), |
| 3877 | _result_outcome_artifact(source, result), |
| 3878 | ) |
| 3879 | if source == "digg": |
| 3880 | result = digg.search_digg(subquery.search_query, from_date, to_date, depth=depth) |
| 3881 | items = digg.parse_digg_response(result, query=subquery.search_query) |
| 3882 | # Enrichment with attached X posts is deferred to |
| 3883 | # _finalize_items_by_source so it runs on the items that actually |
| 3884 | # survive dedupe rather than on top-K of the raw fanout. |
| 3885 | return items, _result_outcome_artifact(source, result) |
| 3886 | if source == "arxiv": |
| 3887 | result = arxiv.search_arxiv(subquery.search_query, from_date, to_date, depth=depth) |
| 3888 | # Relevance keys off the stable research topic, not the per-subquery |
| 3889 | # search_query, so off-topic narrowing does not let weak matches through. |
| 3890 | relevance_topic = raw_topic or topic or subquery.search_query |
| 3891 | return ( |
| 3892 | arxiv.parse_arxiv_response(result, query=relevance_topic), |
| 3893 | _result_outcome_artifact(source, result), |
| 3894 | ) |
| 3895 | if source == "techmeme": |
| 3896 | result = techmeme.search_techmeme(subquery.search_query, from_date, to_date, depth=depth) |
| 3897 | relevance_topic = raw_topic or topic or subquery.search_query |
| 3898 | return ( |
| 3899 | techmeme.parse_techmeme_response(result, query=relevance_topic), |
| 3900 | _result_outcome_artifact(source, result), |
| 3901 | ) |
| 3902 | if source == "trustpilot": |
| 3903 | # Brand-shape gate keys off the stable research topic, not the narrowed |
| 3904 | # per-subquery search_query, so the company is detected consistently. |
| 3905 | relevance_topic = raw_topic or topic or subquery.search_query |
| 3906 | result = trustpilot.search_trustpilot( |
| 3907 | relevance_topic, from_date, to_date, depth=depth, config=config, |
| 3908 | explicit_domain=trustpilot_domain, |
| 3909 | domain_is_hint=trustpilot_domain_is_hint, |
| 3910 | ) |
| 3911 | return ( |
| 3912 | trustpilot.parse_trustpilot_response(result, query=relevance_topic), |
| 3913 | _result_outcome_artifact(source, result), |
| 3914 | ) |
| 3915 | if source == "bluesky": |
| 3916 | result = bluesky.search_bluesky(subquery.search_query, from_date, to_date, depth=depth, config=config) |
| 3917 | return bluesky.parse_bluesky_response(result), _result_outcome_artifact(source, result) |
| 3918 | if source == "threads": |
| 3919 | result = threads.search_threads( |
| 3920 | subquery.search_query, from_date, to_date, |
| 3921 | depth=depth, |
| 3922 | token=config.get("SCRAPECREATORS_API_KEY"), |
| 3923 | ) |
| 3924 | return threads.parse_threads_response(result), _result_outcome_artifact(source, result) |
| 3925 | if source == "truthsocial": |
| 3926 | result = truthsocial.search_truthsocial(subquery.search_query, from_date, to_date, depth=depth, config=config) |
| 3927 | return truthsocial.parse_truthsocial_response(result), _result_outcome_artifact(source, result) |
| 3928 | if source == "polymarket": |
| 3929 | result = polymarket.search_polymarket(subquery.search_query, from_date, to_date, depth=depth) |
| 3930 | # Relevance filtering keys off the stable original research topic, not the |
| 3931 | # per-subquery search_query (which narrows differently on each fanout pass |
| 3932 | # and would let off-topic markets through on broad subqueries while dropping |
| 3933 | # everything on narrow ones). |
| 3934 | relevance_topic = raw_topic or topic or subquery.search_query |
| 3935 | return ( |
| 3936 | polymarket.parse_polymarket_response(result, topic=relevance_topic), |
| 3937 | _result_outcome_artifact(source, result), |
| 3938 | ) |
| 3939 | if source == "github": |
| 3940 | # Resolve once at the pipeline boundary so search and enrich |
| 3941 | # share the result; otherwise each call would re-run the env |
| 3942 | # lookup and gh-CLI subprocess fallback (up to 5s timeout each). |
| 3943 | token = github.resolve_token(config.get("GITHUB_TOKEN")) |
| 3944 | response = github.search_github(subquery.search_query, from_date, to_date, depth=depth, token=token) |
| 3945 | items = github.parse_github_response(response) |
| 3946 | # Note: an unauth rate-limit (response["error"]) is expected on the |
| 3947 | # tokenless anon tier and returns empty here rather than raising — github |
| 3948 | # is now always eligible, so raising would spam "github failed" on every |
| 3949 | # tokenless run. The condition is logged in github.search_github. |
| 3950 | items = github.enrich_with_comments(items, depth=depth, token=token) |
| 3951 | return items, _result_outcome_artifact(source, response) |
| 3952 | if source == "pinterest": |
| 3953 | result = pinterest.search_pinterest( |
| 3954 | subquery.search_query, from_date, to_date, |
| 3955 | depth=depth, |
| 3956 | token=env.get_pinterest_token(config), |
| 3957 | ) |
| 3958 | return pinterest.parse_pinterest_response(result), _result_outcome_artifact(source, result) |
| 3959 | if source == "xiaohongshu": |
| 3960 | return xiaohongshu_api.search_feeds( |
| 3961 | subquery.search_query, |
| 3962 | from_date, |
| 3963 | to_date, |
| 3964 | env.get_xiaohongshu_api_base(config), |
| 3965 | depth=depth, |
| 3966 | ), {} |
| 3967 | if source == "perplexity": |
| 3968 | return perplexity.search(subquery.search_query, date_range, config, deep=config.get("_deep_research", False)) |
| 3969 | raise RuntimeError(f"Unsupported source: {source}") |
| 3970 | |
| 3971 | |
| 3972 | def _google_key(config: dict[str, Any]) -> str | None: |
| 3973 | return config.get("GOOGLE_API_KEY") or config.get("GEMINI_API_KEY") or config.get("GOOGLE_GENAI_API_KEY") |
| 3974 | |
| 3975 | |
| 3976 | |
| 3977 | |
| 3978 | def _mock_stream_results(source: str, subquery: schema.SubQuery) -> tuple[list[dict], dict]: |
| 3979 | # Namespace URLs and the canned comment by topic: real runs never hand two |
| 3980 | # distinct stories byte-identical evidence, and discovery's same-story fold |
| 3981 | # (correctly) collapses topics that share it. Mock enrichment sub-runs feed |
| 3982 | # this fixture one topic per subquery, so the slug keeps them distinct. |
| 3983 | slug = re.sub(r"[^a-z0-9]+", "-", subquery.search_query.lower()).strip("-") or "topic" |
| 3984 | payloads = { |
| 3985 | "reddit": [ |
| 3986 | { |
| 3987 | "id": "R1", |
| 3988 | "title": f"{subquery.search_query} discussion thread", |
| 3989 | "url": f"https://reddit.com/r/example/comments/{slug}-1", |
| 3990 | "subreddit": "example", |
| 3991 | "date": dates.get_date_range(5)[0], |
| 3992 | "engagement": {"score": 120, "num_comments": 48, "upvote_ratio": 0.91}, |
| 3993 | "selftext": f"Community discussion about {subquery.search_query}.", |
| 3994 | "top_comments": [{"excerpt": f"Strong firsthand feedback from {subquery.search_query} users."}], |
| 3995 | "relevance": 0.82, |
| 3996 | "why_relevant": "Mock Reddit result", |
| 3997 | } |
| 3998 | ], |
| 3999 | "x": [ |
| 4000 | { |
| 4001 | "id": "X1", |
| 4002 | "text": f"People on X are discussing {subquery.search_query} right now.", |
| 4003 | "url": f"https://x.com/example/status/{slug}-1", |
| 4004 | "author_handle": "example", |
| 4005 | "date": dates.get_date_range(2)[0], |
| 4006 | "engagement": {"likes": 200, "reposts": 35, "replies": 18, "quotes": 4}, |
| 4007 | "relevance": 0.79, |
| 4008 | "why_relevant": "Mock X result", |
| 4009 | } |
| 4010 | ], |
| 4011 | "grounding": [ |
| 4012 | { |
| 4013 | "id": "WB1", |
| 4014 | "title": f"{subquery.search_query} article", |
| 4015 | "url": f"https://example.com/article/{slug}", |
| 4016 | "source_domain": "example.com", |
| 4017 | "snippet": f"Recent web reporting about {subquery.search_query}.", |
| 4018 | "date": dates.get_date_range(7)[0], |
| 4019 | "relevance": 0.88, |
| 4020 | "why_relevant": "Brave web search", |
| 4021 | } |
| 4022 | ], |
| 4023 | "digg": [ |
| 4024 | { |
| 4025 | "id": "mock1abc", |
| 4026 | "title": f"Digg cluster about {subquery.search_query}", |
| 4027 | "url": f"https://di.gg/ai/mock1abc-{slug}", |
| 4028 | "tldr": f"Curated cluster summarizing recent {subquery.search_query} discussion across the AI 1000.", |
| 4029 | "author": "", |
| 4030 | "date": dates.get_date_range(3)[0], |
| 4031 | "engagement": {"postCount": 8, "uniqueAuthors": 5, "rank": 2, "rank_score": 49.0}, |
| 4032 | "first_post_age": "3d", |
| 4033 | "posts": [ |
| 4034 | { |
| 4035 | "username": "exampledev", |
| 4036 | "display_name": "Example Dev", |
| 4037 | "category": "Engineer", |
| 4038 | "rank": 142, |
| 4039 | "body": f"Quote from the AI 1000 about {subquery.search_query}.", |
| 4040 | "post_type": "tweet", |
| 4041 | "x_url": "https://x.com/exampledev/status/1", |
| 4042 | "posted_at": dates.get_date_range(3)[0], |
| 4043 | }, |
| 4044 | ], |
| 4045 | "relevance": 0.84, |
| 4046 | "why_relevant": "Mock Digg cluster", |
| 4047 | }, |
| 4048 | { |
| 4049 | "id": "mock2def", |
| 4050 | "title": f"Second Digg cluster on {subquery.search_query}", |
| 4051 | "url": f"https://di.gg/ai/mock2def-{slug}", |
| 4052 | "tldr": f"Another angle on {subquery.search_query}.", |
| 4053 | "author": "", |
| 4054 | "date": dates.get_date_range(8)[0], |
| 4055 | "engagement": {"postCount": 3, "uniqueAuthors": 2, "rank": 18, "rank_score": 33.0}, |
| 4056 | "first_post_age": "8d", |
| 4057 | "posts": [], |
| 4058 | "relevance": 0.71, |
| 4059 | "why_relevant": "Mock Digg cluster", |
| 4060 | }, |
| 4061 | ], |
| 4062 | "arxiv": [ |
| 4063 | { |
| 4064 | "id": f"http://arxiv.org/abs/2606.00001v1-{slug}", |
| 4065 | "title": f"A Survey of {subquery.search_query}", |
| 4066 | "url": f"https://arxiv.org/abs/2606.00001v1-{slug}", |
| 4067 | "summary": f"We present a comprehensive study of {subquery.search_query} and its recent advances.", |
| 4068 | "author": "Ada Lovelace et al.", |
| 4069 | "authors": ["Ada Lovelace", "Alan Turing"], |
| 4070 | "date": dates.get_date_range(20)[0], |
| 4071 | "engagement": {}, |
| 4072 | "relevance": 0.86, |
| 4073 | "why_relevant": "Mock arXiv paper", |
| 4074 | }, |
| 4075 | ], |
| 4076 | "techmeme": [ |
| 4077 | { |
| 4078 | "id": f"https://www.techmeme.com/260627/p1-{slug}", |
| 4079 | "title": f"Major development in {subquery.search_query} reshapes the industry", |
| 4080 | "url": f"https://www.techmeme.com/260627/p1-{slug}", |
| 4081 | "source_name": "techcrunch.com", |
| 4082 | "date": dates.get_date_range(1)[0], |
| 4083 | "engagement": {}, |
| 4084 | "relevance": 0.83, |
| 4085 | "why_relevant": "Mock Techmeme headline", |
| 4086 | }, |
| 4087 | ], |
| 4088 | "dripstack": [ |
| 4089 | { |
| 4090 | "id": "DS1", |
| 4091 | "title": f"Deep dive: {subquery.search_query} from a paid newsletter", |
| 4092 | "url": f"https://newsletter.example.com/deep-dive-{slug}", |
| 4093 | "author": "newsletter.example.com", |
| 4094 | "date": dates.get_date_range(3)[0], |
| 4095 | "engagement": {}, |
| 4096 | "relevance": 0.85, |
| 4097 | "why_relevant": "Mock DripStack newsletter result", |
| 4098 | "snippet": f"Professional analyst coverage of {subquery.search_query}.", |
| 4099 | "metadata": { |
| 4100 | "publication_slug": "newsletter.example.com", |
| 4101 | "post_slug": "deep-dive", |
| 4102 | "relevance_score": 85, |
| 4103 | "match_confidence": "strong", |
| 4104 | }, |
| 4105 | }, |
| 4106 | ], |
| 4107 | "trustpilot": [ |
| 4108 | { |
| 4109 | "id": "example.com", |
| 4110 | "title": f"{subquery.search_query}: TrustScore 3.4", |
| 4111 | "url": f"https://www.trustpilot.com/review/{slug}.example.com", |
| 4112 | "summary": f"Across recent reviews, customers were split on {subquery.search_query}: some praised support, others cited delays.", |
| 4113 | "name": subquery.search_query, |
| 4114 | "trustScore": 3.4, |
| 4115 | "reviewCount": 128, |
| 4116 | "date": dates.get_date_range(1)[0], |
| 4117 | "engagement": {"reviews": 128, "trustScore": 3.4}, |
| 4118 | "relevance": 0.8, |
| 4119 | "why_relevant": "Mock Trustpilot sentiment", |
| 4120 | }, |
| 4121 | ], |
| 4122 | "jobs": [ |
| 4123 | { |
| 4124 | "id": "J1", |
| 4125 | "title": "Founding Enterprise Solutions Engineer", |
| 4126 | "url": f"https://boards.greenhouse.io/example/jobs/{slug}-1", |
| 4127 | "description": ( |
| 4128 | f"Work with enterprise customers on SSO, SOC 2, security, " |
| 4129 | f"and procurement workflows for {subquery.search_query}." |
| 4130 | ), |
| 4131 | "department": "Sales", |
| 4132 | "location": "San Francisco, CA", |
| 4133 | "date": dates.get_date_range(4)[0], |
| 4134 | "provider": "mock", |
| 4135 | "relevance": 0.8, |
| 4136 | "why_relevant": "Mock public job posting", |
| 4137 | }, |
| 4138 | { |
| 4139 | "id": "J2", |
| 4140 | "title": "Security Platform Engineer", |
| 4141 | "url": f"https://boards.greenhouse.io/example/jobs/{slug}-2", |
| 4142 | "description": "Build enterprise security, audit, and admin workflows.", |
| 4143 | "department": "Engineering", |
| 4144 | "location": "Remote", |
| 4145 | "date": dates.get_date_range(6)[0], |
| 4146 | "provider": "mock", |
| 4147 | "relevance": 0.78, |
| 4148 | "why_relevant": "Mock public job posting", |
| 4149 | }, |
| 4150 | ], |
| 4151 | } |
| 4152 | if source == "grounding": |
| 4153 | return payloads.get(source, []), { |
| 4154 | "label": subquery.label, |
| 4155 | "mock": True, |
| 4156 | "webSearchQueries": [subquery.search_query], |
| 4157 | "resultCount": 1, |
| 4158 | } |
| 4159 | return payloads.get(source, []), {} |
| 4160 | |
| 4161 |