返回 DeepSeek-Reasonix
AGENT_CORE_SIMPLIFICATION.md
根目录 / docs / AGENT_CORE_SIMPLIFICATION.md
1 # Agent Core Simplification
2
3 This document tracks the agent-core simplification effort: the behavioral
4 contract of the simplified loop, the metrics used to compare before/after, and
5 where each contract item is tested. The Chinese version lives in
6 `AGENT_CORE_SIMPLIFICATION.zh-CN.md`.
7
8 ## Target loop
9
10 ```text
11 build request
12 -> provider stream
13 -> clean final: done
14 -> tool call: execute, next step
15 -> request error: unified retry
16 -> unhandled error: explicit failure
17 ```
18
19 ## Product decisions
20
21 - Normal requests are executor-only; the planner is opt-in (`planner_model`).
22 - Normal agents ship with synthetic continuation disabled; Goal, review,
23 guardian, and typed-report flows keep their own constraints.
24 - Compaction defaults to a single summary; chunked/tree-reduce recovery is
25 explicit only (manual `/compact` and marked recovery workflows).
26 - Final readiness, tool safety, cancellation, budgets, and explicit whole-file
27 read pauses stay as hard boundaries. Ordinary partial reads do not freeze
28 independent work; see [Read evidence lifecycle](READ_EVIDENCE_LIFECYCLE.md).
29 - Old configs and old session state stay readable for one release; the new
30 runtime never executes the old fallbacks.
31
32 ## Metrics baseline
33
34 Reuse the existing usage and e2ebench instrumentation; no new fallback
35 telemetry is added. Capture before/after per phase with:
36
37 - `reasonix run --metrics <path>` — per-run `RunMetrics`: token/cost totals,
38 `usage_by_source` (executor/planner/subagent/compaction/... request calls),
39 `retries`, `compactions`, `steps`.
40 - `go run ./cmd/e2ebench -task <task> -json` — per-task request counts,
41 `usage_by_source`, trajectory digest (stream retries, reasoning replays,
42 empty-final retries, TTFT, requests by source), wall time, cache hit/miss.
43
44 | Metric | Source |
45 |---|---|
46 | model requests per normal turn | `usage_by_source["executor"].Calls` / trajectory `ExecutorRequests` |
47 | planner requests | `usage_by_source["planner"].Calls` / trajectory `PlannerRequests` |
48 | reviewer/evaluator/guardian requests | `usage_by_source["recovery_reviewer"|"goal_evaluator"]`, guardian assessment usage |
49 | synthetic continuations | executor requests per clean turn above 1; trajectory `EmptyFinalRetries` |
50 | stream retries | trajectory `StreamRetries` / `RunMetrics.Retries` |
51 | compaction requests and summary spans | `RunMetrics.Compactions`, compaction telemetry notice (`spans=`, `reqs=`) |
52 | first text latency | trajectory `TTFTMs` |
53 | turn latency | `RunMetrics.DurationMs` / bench `WallMs` |
54 | tool execution success | bench task solved rate / `SolvedThenBroken` |
55 | protocol-error terminations | trajectory retry-exhausted outcomes |
56
57 Compare at least: normal Q&A, normal code edit, long-context/tool-heavy
58 (`context-pressure` tasks). Goal per phase: a normal request is one executor
59 request chain, no implicit planner/reviewer/evaluator requests, no extra
60 continuation after a clean final, default compaction never enters multi-span
61 summaries, and hard-safety failure rates do not increase.
62
63 ## Contract tests
64
65 New consolidated suite: `internal/agent/agent_contract_test.go`.
66
67 | Contract item | Test |
68 |---|---|
69 | clean final makes exactly one model request | `TestContractCleanFinalMakesOneModelRequest` |
70 | tool call executes, loop advances | `TestContractToolCallAdvancesToNextStep` |
71 | thinking survives the unified retry (frozen request) | `TestContractThinkingSurvivesUnifiedRetry` |
72 | no long-lived fallback state after retry exhaustion | `TestContractNoLongLivedFallbackStateAfterRetryExhaustion` |
73 | clean final adds no synthetic continuation | `TestContractCleanFinalAddsNoSyntheticContinuation` |
74 | reasoning-only clean stop completes | `TestRunAcceptsReasoningOnlyFinalAnswer` |
75 | zero content retries via unified `EMPTY_RESPONSE` path | `TestRunRetriesZeroContentWithTheSameFrozenRequest` |
76 | exhausted retries return an explicit protocol error | `TestRunStopsAfterExhaustedZeroContentRetriesWithoutCommittingEmptyMessages` |
77 | strict-provider missing reasoning: one frozen-request retry | `TestRunSilentlyRecoversMissingToolCallReasoning` and the replay suites in `loop_e2e_test.go`/`retry_e2e_test.go` (#9776 repair) |
78 | incomplete-read gate | `incomplete_read_test.go` |
79 | final readiness | `final_readiness_test.go` |
80 | cancellation | `cancel_test.go` |
81 | task/token/cost budgets | `run_budget_test.go` |
82 | tool permission/malformed args | `argument_validation_test.go`, gate tests |
83 | ordinary request skips planner | `TestCoordinatorOrdinaryRequestDoesNotCallPlanner`, `TestDecidePlannerRouteExplicitOnly` |
84 | planner prose without `submit_plan` fails | `TestCoordinatorPlanAndExecuteRequiresSubmittedPlan` |
85 | planner failure does not run executor | `TestCoordinatorFailsClosedWhenPlannerFails` |
86 | unusable `planner_model` is a config error | `TestBuildFailsWhenPlannerModelIsUnresolvable` |
87 | ordinary Agent adds no todo continuation | `TestStandardTodoContinuationDisabledByDefault` |
88 | pressure compaction stays single-summary | `TestPressureCompactionDoesNotCallChunkedFold` |
89 | missing guardian/recovery models fail closed | `TestBuildFailsWhenGuardianModelIsUnresolvable`, `TestBuildFailsWhenRecoveryModelIsUnresolvable` |
90
91 ## Gates
92
93 Every phase runs:
94
95 ```bash
96 go test -count=1 ./internal/agent/... ./internal/control/... ./internal/config/... ./internal/boot/...
97 go test -race ./internal/agent/... ./internal/provider/...
98 go vet ./...
99 go run ./tools/repolint
100 ```
101
101 lines MARKDOWN