| 1 | package extension |
| 2 | |
| 3 | import ( |
| 4 | "context" |
| 5 | "errors" |
| 6 | "fmt" |
| 7 | "maps" |
| 8 | "strings" |
| 9 | |
| 10 | "reasonix/internal/provider" |
| 11 | "reasonix/internal/tool" |
| 12 | ) |
| 13 | |
| 14 | // Activator binds live resources (MCP sidecars, watchers) to a freshly |
| 15 | // frozen snapshot. Stage 2 wires none; the seam exists so later stages plug |
| 16 | // in without touching the build pipeline. A nil *RuntimeSet result is |
| 17 | // treated as an empty set bound to the snapshot's generation. |
| 18 | type Activator func(ctx context.Context, snap *RuntimeSnapshot) (*RuntimeSet, error) |
| 19 | |
| 20 | // Builder assembles a RuntimeSnapshot from registered contributors through a |
| 21 | // fixed pipeline: Discover → Parse → Validate → Resolve → Assemble → Freeze → |
| 22 | // Activate. The pipeline is linear by design — every contributor sees the |
| 23 | // same rules, and every consumer reads the same frozen result. |
| 24 | type Builder struct { |
| 25 | contributors []Contributor |
| 26 | generation uint64 |
| 27 | systemPrompt string |
| 28 | activator Activator |
| 29 | conflictPolicy ConflictPolicy |
| 30 | } |
| 31 | |
| 32 | // NewBuilder returns an empty builder with generation 0. |
| 33 | func NewBuilder() *Builder { return &Builder{} } |
| 34 | |
| 35 | // AddContributor registers contributors. Registration order never influences |
| 36 | // the snapshot — the catalog, winner rules, and interceptor chains all sort |
| 37 | // on contribution data, not arrival order — so callers may register in any |
| 38 | // order. |
| 39 | func (b *Builder) AddContributor(contributors ...Contributor) *Builder { |
| 40 | b.contributors = append(b.contributors, contributors...) |
| 41 | return b |
| 42 | } |
| 43 | |
| 44 | // WithGeneration sets the snapshot generation. Generations pair with |
| 45 | // RuntimeSet.CloseIfGeneration to keep stale cleanup from closing a newer |
| 46 | // runtime's resources. |
| 47 | func (b *Builder) WithGeneration(gen uint64) *Builder { |
| 48 | b.generation = gen |
| 49 | return b |
| 50 | } |
| 51 | |
| 52 | // WithSystemPrompt sets the assembled system prompt text. It participates in |
| 53 | // CacheHash because it is part of the provider-visible request prefix. |
| 54 | func (b *Builder) WithSystemPrompt(prompt string) *Builder { |
| 55 | b.systemPrompt = prompt |
| 56 | return b |
| 57 | } |
| 58 | |
| 59 | // WithActivator installs the activation seam. The default activator returns |
| 60 | // an empty RuntimeSet bound to the snapshot generation. |
| 61 | func (b *Builder) WithActivator(a Activator) *Builder { |
| 62 | b.activator = a |
| 63 | return b |
| 64 | } |
| 65 | |
| 66 | // ConflictPolicy selects how resolution treats same-tier duplicates of one |
| 67 | // canonical ID claimed by distinct sources for a shadowed kind. |
| 68 | type ConflictPolicy int |
| 69 | |
| 70 | const ( |
| 71 | // ConflictFail is the default: a disputed ID aborts the build with a |
| 72 | // ConflictError. v2 extensions use it — a package must never silently |
| 73 | // override another package's capability. |
| 74 | ConflictFail ConflictPolicy = iota |
| 75 | // ConflictCollect keeps the deterministic winner (highest tier, then |
| 76 | // first registration) and records the dispute on the snapshot's |
| 77 | // Diagnostics instead of failing the build. Boot's legacy assembly uses |
| 78 | // it: those resources already resolved their clashes inside their own |
| 79 | // discovery passes, and surfacing a residual dispute must never change |
| 80 | // whether a session boots. Malformed contributions still fail validation, |
| 81 | // and replacement-slot disputes still fail resolution — only shadowing |
| 82 | // conflicts are collected. |
| 83 | ConflictCollect |
| 84 | ) |
| 85 | |
| 86 | // WithConflictPolicy sets how same-tier multi-source duplicates are treated. |
| 87 | // The zero value is ConflictFail; see ConflictPolicy. |
| 88 | func (b *Builder) WithConflictPolicy(p ConflictPolicy) *Builder { |
| 89 | b.conflictPolicy = p |
| 90 | return b |
| 91 | } |
| 92 | |
| 93 | // ValidationError reports one malformed contribution. Build collects all of |
| 94 | // them so a broken manifest surfaces every problem in one pass. |
| 95 | type ValidationError struct { |
| 96 | Kind ContributionKind |
| 97 | ID string |
| 98 | Reason string |
| 99 | } |
| 100 | |
| 101 | func (e *ValidationError) Error() string { |
| 102 | if e.ID == "" { |
| 103 | return fmt.Sprintf("extension: invalid %s contribution: %s", e.Kind, e.Reason) |
| 104 | } |
| 105 | return fmt.Sprintf("extension: invalid %s %q: %s", e.Kind, e.ID, e.Reason) |
| 106 | } |
| 107 | |
| 108 | // Build runs the full pipeline and returns the frozen snapshot plus its bound |
| 109 | // runtime resources. Any validation, conflict, or activation error aborts the |
| 110 | // build: publishing half-resolved state would let a losing contribution leak |
| 111 | // into the runtime. Under ConflictCollect a shadowing conflict no longer |
| 112 | // aborts: the deterministic winner is kept and the dispute is recorded on the |
| 113 | // snapshot's Diagnostics. |
| 114 | func (b *Builder) Build(ctx context.Context) (*RuntimeSnapshot, *RuntimeSet, error) { |
| 115 | raw, err := b.discover(ctx) |
| 116 | if err != nil { |
| 117 | return nil, nil, err |
| 118 | } |
| 119 | parsed := parseContributions(raw) |
| 120 | if err := validateContributions(parsed); err != nil { |
| 121 | return nil, nil, err |
| 122 | } |
| 123 | resolved, replacements, conflicts, err := resolveContributions(parsed, b.conflictPolicy) |
| 124 | if err != nil { |
| 125 | return nil, nil, err |
| 126 | } |
| 127 | snap := b.assemble(resolved, replacements, conflicts) |
| 128 | runtimeSet, err := b.activate(ctx, snap) |
| 129 | if err != nil { |
| 130 | return nil, nil, err |
| 131 | } |
| 132 | return snap, runtimeSet, nil |
| 133 | } |
| 134 | |
| 135 | // discover asks every contributor for its offerings and stamps the |
| 136 | // per-contributor registration sequence. An empty Origin defaults to the |
| 137 | // contributor name so conflict reports always have something meaningful to |
| 138 | // say. |
| 139 | func (b *Builder) discover(ctx context.Context) ([]Contribution, error) { |
| 140 | var out []Contribution |
| 141 | for _, c := range b.contributors { |
| 142 | contribs, err := c.Contribute(ctx) |
| 143 | if err != nil { |
| 144 | return nil, fmt.Errorf("extension: contributor %q: %w", c.Name(), err) |
| 145 | } |
| 146 | for i, ct := range contribs { |
| 147 | ct.Order = i |
| 148 | if ct.Source.Origin == "" { |
| 149 | ct.Source.Origin = c.Name() |
| 150 | } |
| 151 | out = append(out, ct) |
| 152 | } |
| 153 | } |
| 154 | return out, nil |
| 155 | } |
| 156 | |
| 157 | // parseContributions normalizes raw contributions. Stage 2 has no manifest |
| 158 | // decoding to do — adapters hand over typed payloads — so parsing is limited |
| 159 | // to ID hygiene; the stage exists so later manifest formats slot into the |
| 160 | // pipeline without reordering it. |
| 161 | func parseContributions(in []Contribution) []Contribution { |
| 162 | out := make([]Contribution, len(in)) |
| 163 | for i, ct := range in { |
| 164 | ct.ID = strings.TrimSpace(ct.ID) |
| 165 | out[i] = ct |
| 166 | } |
| 167 | return out |
| 168 | } |
| 169 | |
| 170 | // validateContributions rejects malformed contributions before any winner |
| 171 | // rules run, so resolution never has to guess what an invalid entry meant. |
| 172 | func validateContributions(cs []Contribution) error { |
| 173 | var errs []error |
| 174 | for _, ct := range cs { |
| 175 | if !knownKind(ct.Kind) { |
| 176 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "unknown kind"}) |
| 177 | continue |
| 178 | } |
| 179 | if ct.ID == "" { |
| 180 | errs = append(errs, &ValidationError{Kind: ct.Kind, Reason: "empty ID"}) |
| 181 | continue |
| 182 | } |
| 183 | if strings.ContainsAny(ct.ID, " \t\n") { |
| 184 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "ID contains whitespace"}) |
| 185 | } |
| 186 | if !knownScope(ct.Source.Scope) { |
| 187 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: fmt.Sprintf("unknown scope %q", ct.Source.Scope)}) |
| 188 | } |
| 189 | switch ct.Kind { |
| 190 | case KindTool: |
| 191 | errs = append(errs, validateToolContribution(ct)...) |
| 192 | case KindProvider: |
| 193 | if _, _, ok := splitProviderRef(ct.ID); !ok { |
| 194 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "provider ID must be a <name>/<model> ref"}) |
| 195 | } |
| 196 | case KindInterceptor: |
| 197 | if !knownInterceptorPoint(InterceptorPoint(ct.ID)) { |
| 198 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "unknown interceptor point"}) |
| 199 | } |
| 200 | if err := ValidatePriority(ct.Priority); err != nil { |
| 201 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: err.Error()}) |
| 202 | } |
| 203 | } |
| 204 | } |
| 205 | return errors.Join(errs...) |
| 206 | } |
| 207 | |
| 208 | // validateToolContribution enforces the tool ID contract: lowercase names, |
| 209 | // the mcp__<server>__<tool> namespace for MCP-backed tools, and a payload the |
| 210 | // assembler can render into a provider schema. |
| 211 | func validateToolContribution(ct Contribution) []error { |
| 212 | var errs []error |
| 213 | if ct.ID != strings.ToLower(ct.ID) { |
| 214 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "tool IDs must be lowercase"}) |
| 215 | } |
| 216 | if strings.HasPrefix(ct.ID, tool.MCPNamePrefix) { |
| 217 | if _, _, ok := tool.SplitMCPName(ct.ID); !ok { |
| 218 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "malformed MCP tool name, want mcp__<server>__<tool>"}) |
| 219 | } |
| 220 | } else if _, isMCP := ct.Payload.(tool.MCPMetadata); isMCP { |
| 221 | // An MCP-backed tool outside the mcp__ namespace would collide with |
| 222 | // built-in names and bypass MCP-specific policy checks. |
| 223 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "MCP tool IDs must start with mcp__"}) |
| 224 | } |
| 225 | if _, ok := toolSchemaOf(ct); !ok { |
| 226 | errs = append(errs, &ValidationError{Kind: ct.Kind, ID: ct.ID, Reason: "payload must be tool.Tool, tool.ContractEntry, or provider.ToolSchema"}) |
| 227 | } |
| 228 | return errs |
| 229 | } |
| 230 | |
| 231 | // ValidToolID reports whether id satisfies the kernel's tool-ID contract: |
| 232 | // lowercase, and a well-formed mcp__<server>__<tool> name when the MCP |
| 233 | // namespace prefix is present. It encodes the ID-shape half of |
| 234 | // validateToolContribution (keep the two in sync); assemblers wrapping a |
| 235 | // pre-kernel legacy registry use it to skip names that predate the contract |
| 236 | // instead of failing the whole build. |
| 237 | func ValidToolID(id string) bool { |
| 238 | if id != strings.ToLower(id) { |
| 239 | return false |
| 240 | } |
| 241 | if strings.HasPrefix(id, tool.MCPNamePrefix) { |
| 242 | _, _, ok := tool.SplitMCPName(id) |
| 243 | return ok |
| 244 | } |
| 245 | return true |
| 246 | } |
| 247 | |
| 248 | // toolSchemaOf renders a tool contribution's payload into a provider schema. |
| 249 | // Parameters are canonicalized here — once — so every consumer, including |
| 250 | // CacheHash, sees identical bytes regardless of how the contributor marshaled |
| 251 | // them. |
| 252 | func toolSchemaOf(ct Contribution) (provider.ToolSchema, bool) { |
| 253 | switch p := ct.Payload.(type) { |
| 254 | case tool.Tool: |
| 255 | return provider.ToolSchema{ |
| 256 | Name: p.Name(), |
| 257 | Description: p.Description(), |
| 258 | Parameters: provider.CanonicalizeSchema(p.Schema()), |
| 259 | }, true |
| 260 | case tool.ContractEntry: |
| 261 | return provider.ToolSchema{ |
| 262 | Name: p.Name, |
| 263 | Description: p.Description, |
| 264 | Parameters: provider.CanonicalizeSchema(p.Schema), |
| 265 | }, true |
| 266 | case provider.ToolSchema: |
| 267 | p.Parameters = provider.CanonicalizeSchema(p.Parameters) |
| 268 | return p, true |
| 269 | default: |
| 270 | return provider.ToolSchema{}, false |
| 271 | } |
| 272 | } |
| 273 | |
| 274 | // resolveContributions applies the winner rules and returns the effective |
| 275 | // contribution set, the replacement-slot owners, and — under ConflictCollect — |
| 276 | // the shadowing disputes it resolved without failing. |
| 277 | // |
| 278 | // Shadowed kinds (tools, skills, commands, MCP servers, providers, prompts, |
| 279 | // themes, UI actions, strategies): the highest-tier contribution wins the |
| 280 | // canonical ID; distinct sources tied at that tier are a hard ConflictError |
| 281 | // under ConflictFail — the kernel refuses to pick a winner the user didn't |
| 282 | // ask for — or a recorded ConflictError under ConflictCollect, with the same |
| 283 | // deterministic winner kept. Duplicates from a single source collapse to the |
| 284 | // first registration, mirroring the first-root-wins behavior inside today's |
| 285 | // discovery passes. |
| 286 | // |
| 287 | // Hooks and interceptors are additive: every contribution survives and |
| 288 | // nothing ever conflicts. |
| 289 | // |
| 290 | // Replacement claims come from payloads implementing SlotClaimer; a second |
| 291 | // claimant for a slot is a hard SlotConflictError under both policies — a |
| 292 | // slot replaces runtime behavior outright, so there is no shadowing winner |
| 293 | // to keep. |
| 294 | func resolveContributions(cs []Contribution, policy ConflictPolicy) (resolved []Contribution, replacements map[Slot]ContributionSource, conflicts []ConflictError, err error) { |
| 295 | type key struct { |
| 296 | kind ContributionKind |
| 297 | id string |
| 298 | } |
| 299 | groups := map[key][]Contribution{} |
| 300 | var order []key |
| 301 | for _, ct := range cs { |
| 302 | k := key{ct.Kind, ct.ID} |
| 303 | if _, seen := groups[k]; !seen { |
| 304 | order = append(order, k) |
| 305 | } |
| 306 | groups[k] = append(groups[k], ct) |
| 307 | } |
| 308 | |
| 309 | var errs []error |
| 310 | resolved = make([]Contribution, 0, len(cs)) |
| 311 | claims := NewReplaceClaims() |
| 312 | for _, k := range order { |
| 313 | group := groups[k] |
| 314 | if additiveKind(k.kind) { |
| 315 | resolved = append(resolved, group...) |
| 316 | } else { |
| 317 | winner, sources, conflicted := resolveGroup(group) |
| 318 | if conflicted { |
| 319 | conflict := ConflictError{Kind: k.kind, ID: k.id, Sources: sources} |
| 320 | if policy == ConflictCollect { |
| 321 | // The dispute is surfaced on the snapshot; the winner |
| 322 | // rules above still decide what the runtime sees. |
| 323 | conflicts = append(conflicts, conflict) |
| 324 | resolved = append(resolved, winner) |
| 325 | continue |
| 326 | } |
| 327 | errs = append(errs, &conflict) |
| 328 | continue |
| 329 | } |
| 330 | resolved = append(resolved, winner) |
| 331 | } |
| 332 | } |
| 333 | // Claims are collected across all contributions, not just winners: a |
| 334 | // losing contribution must not silently keep a slot it declared, because |
| 335 | // slots replace runtime behavior regardless of catalog shadowing. |
| 336 | for _, ct := range cs { |
| 337 | claimer, ok := ct.Payload.(SlotClaimer) |
| 338 | if !ok { |
| 339 | continue |
| 340 | } |
| 341 | for _, slot := range claimer.ReplacementSlots() { |
| 342 | if err := claims.Claim(slot, ct.Source); err != nil { |
| 343 | errs = append(errs, err) |
| 344 | } |
| 345 | } |
| 346 | } |
| 347 | if err := errors.Join(errs...); err != nil { |
| 348 | return nil, nil, nil, err |
| 349 | } |
| 350 | return resolved, claims.Claims(), conflicts, nil |
| 351 | } |
| 352 | |
| 353 | // resolveGroup picks the winning contribution for one (kind, id): the first |
| 354 | // registration among the highest-tier entries, plus whether the top tier is |
| 355 | // disputed between distinct sources. The winner is returned even when the |
| 356 | // group is disputed so a ConflictCollect build keeps resolving to the same |
| 357 | // deterministic entry; ConflictFail callers discard it. |
| 358 | func resolveGroup(group []Contribution) (winner Contribution, sources []ContributionSource, conflicted bool) { |
| 359 | best := -1 |
| 360 | for _, ct := range group { |
| 361 | if r := tierRank(ct.Source.Scope); r > best { |
| 362 | best = r |
| 363 | } |
| 364 | } |
| 365 | var top []Contribution |
| 366 | for _, ct := range group { |
| 367 | if tierRank(ct.Source.Scope) == best { |
| 368 | top = append(top, ct) |
| 369 | } |
| 370 | } |
| 371 | winner = top[0] |
| 372 | for _, ct := range top[1:] { |
| 373 | if ct.Order < winner.Order { |
| 374 | winner = ct |
| 375 | } |
| 376 | } |
| 377 | if sources, ok := conflictingSources(group); ok { |
| 378 | return winner, sources, true |
| 379 | } |
| 380 | return winner, nil, false |
| 381 | } |
| 382 | |
| 383 | // assemble freezes the resolved set into an immutable snapshot. Everything |
| 384 | // derivable is derived here — schemas rendered and sorted, chains grouped and |
| 385 | // ordered, hashes computed — so snapshot accessors stay trivial copies. |
| 386 | // conflicts are the shadowing disputes a ConflictCollect build resolved with |
| 387 | // its ordinary winner rules; they are frozen onto the snapshot as Diagnostics |
| 388 | // in pipeline (first-appearance) order. |
| 389 | func (b *Builder) assemble(resolved []Contribution, replacements map[Slot]ContributionSource, conflicts []ConflictError) *RuntimeSnapshot { |
| 390 | catalog := NewCatalog() |
| 391 | catalog.Add(resolved...) |
| 392 | |
| 393 | schemas := make([]provider.ToolSchema, 0) |
| 394 | for _, ct := range catalog.ByKind(KindTool) { |
| 395 | schema, ok := toolSchemaOf(ct) |
| 396 | if ok { |
| 397 | schemas = append(schemas, schema) |
| 398 | } |
| 399 | } |
| 400 | schemas = normalizeToolSchemas(schemas) |
| 401 | |
| 402 | chains := map[InterceptorPoint][]Contribution{} |
| 403 | for _, ct := range catalog.ByKind(KindInterceptor) { |
| 404 | point := InterceptorPoint(ct.ID) |
| 405 | chains[point] = append(chains[point], ct) |
| 406 | } |
| 407 | for point, chain := range chains { |
| 408 | chains[point] = SortInterceptors(chain) |
| 409 | } |
| 410 | |
| 411 | repl := make(map[Slot]ContributionSource, len(replacements)) |
| 412 | maps.Copy(repl, replacements) |
| 413 | |
| 414 | diagnostics := make([]string, 0, len(conflicts)) |
| 415 | for i := range conflicts { |
| 416 | diagnostics = append(diagnostics, conflicts[i].Error()) |
| 417 | } |
| 418 | |
| 419 | systemHash, toolsHash, cacheHash := computeCacheShape(b.systemPrompt, schemas) |
| 420 | |
| 421 | catalog.freeze() |
| 422 | return &RuntimeSnapshot{ |
| 423 | generation: b.generation, |
| 424 | catalog: catalog, |
| 425 | systemPrompt: b.systemPrompt, |
| 426 | toolSchemas: schemas, |
| 427 | interceptorChain: chains, |
| 428 | replacements: repl, |
| 429 | diagnostics: diagnostics, |
| 430 | cacheHash: cacheHash, |
| 431 | systemHash: systemHash, |
| 432 | toolsHash: toolsHash, |
| 433 | } |
| 434 | } |
| 435 | |
| 436 | // activate binds runtime resources through the configured Activator, or |
| 437 | // returns an empty set. The snapshot is already frozen at this point: an |
| 438 | // activator must observe, never mutate. |
| 439 | func (b *Builder) activate(ctx context.Context, snap *RuntimeSnapshot) (*RuntimeSet, error) { |
| 440 | if b.activator == nil { |
| 441 | return NewRuntimeSet(snap.Generation()), nil |
| 442 | } |
| 443 | runtimeSet, err := b.activator(ctx, snap) |
| 444 | if err != nil { |
| 445 | return nil, fmt.Errorf("extension: activate generation %d: %w", snap.Generation(), err) |
| 446 | } |
| 447 | if runtimeSet == nil { |
| 448 | runtimeSet = NewRuntimeSet(snap.Generation()) |
| 449 | } |
| 450 | return runtimeSet, nil |
| 451 | } |
| 452 |