modelrig 0.4.0 → 0.6.0

package/CHANGELOG.md

@@ -10,6 +10,244 @@ published package diverge.
 This project adheres to [Semantic Versioning](https://semver.org/). Dates are
 UTC.
 
+## [0.6.0] — 2026-08-23
+
+Released with the docs/registry mirror in one R-A ceremony. Headline for router
+codebases: the candidate ladder can now fall through on **quality** (`qualityGate`
+→ `quality_rejected`), bake-offs measure **regret against a reference arm** under a
+declared objective (legacy routes stay cost-first, byte-identical), the console
+gains the **project tree** (`/projects`) and the **analytics API**
+(`/v1/analytics/meta` + `/query`, metadata only — never content), and the
+OpenAI-compatible proxy accepts the `developer` role. ZERO DDL across the line.
+
+Regret-bounded bake-offs (RB-1, near-optimal) — the bake-off objective flips from
+"cheapest that passes" to **near-optimal at a reasonable price**. New additive
+route YAML: `objective` (`mode: quality-first | balanced | cost-first`,
+`max_regret`, `price_ceiling`, `lambda`), `bakeoff` (`reference: auto | provider/model`,
+`reference_envelope_usd`, `judge: { model, max_samples }`), and `value_checks`
+(JSON-pointer assertions). Every bake-off can run a **reference (ceiling) arm**
+and report each arm's **regret vs the reference with a seeded-bootstrap CI and n**.
+Quality per sample resolves eval-case golden → `value_checks` → an opt-in capped,
+metered judge (judge model ∈ declared candidates) → 1/0 conformance; it composes
+with Q1's `qualityGate` (deterministic checks first, then the gate/judge). The
+proposal gate is objective-aware: `cost-first` keeps Q1's gate byte-identical
+(golden); a declared `quality-first`/`balanced` objective with no bound reports
+regret but proposes nothing (no silent actuation). `modelrig bakeoff --objective
+<mode>` overrides per run. The report/PR body state regret, cost and the ceiling
+on one line. ZERO DDL. See `docs/bakeoffs.md` §5.
+
+Analytics surface (WI-1b) — the read side of the attempt envelope. `GET
+/v1/analytics/meta` describes every dimension and metric with its classification
+and `since` version; `POST /v1/analytics/query` runs the SELECT-only grammar over
+an org's inferences (metadata only — enums, numbers, hashes, identifiers; never
+content; bake-off replay rows excluded). `modelrig query "<grammar>"` /
+`--saved <id>` run the same grammar locally. Console: saved questions on
+`/costs`, `/coverage`, `/routes/[route]`. OTLP-out gains `modelrig.meta.*` span
+attributes for the metadata-safe envelope fields. Behaviour change: a free-text
+`actor` is dropped with one warning (metadata-safe rule). New root exports:
+`parseQuery`, `compileQuery`, `evaluateAnalyticsQuery`, `describeAnalyticsSchema`,
+`planAnalyticsRead`, `runTelemetryQuery`, `SAVED_QUERIES`, `savedQueryById`, the
+column registries and their types.
+ZERO DDL.
+
+Quality gate (#223, Q1) — the candidate ladder can now fall through on *quality*,
+not just shape. Pass `rig.run(route, { qualityGate })` and a schema-valid output
+your predicate rejects is recorded as a new `quality_rejected` failure (its own
+retry budget, never the repair rung) and the ladder advances. The same predicate
+scores bake-offs (`rig.bakeoff({ qualityGate })`) — a quality pass rate and an
+`effectiveUsdPer1kPassing` metric beside the conformant one. SDK lane only: a
+function cannot cross the hosted boundary; the route-declared hosted judge is Q2.
+
+### Added — quality gate (#223, Q1)
+
+- **`RunOptions.qualityGate?`** — a builder predicate
+  `(output, meta) => { ok, score?, reason? }` (sync or async) evaluated on a
+  schema-VALID output before it is accepted. `ok:false` records a
+  `quality_rejected` attempt (message carries the reason + score), captures the
+  rejected text for replay, and falls the ladder through on the class's OWN retry
+  budget (`policy.retries.quality_rejected`, default 0) — the repair rung is NEVER
+  entered from this path. A throwing predicate rejects fail-closed (`reason:
+  "qualityGate threw: …"`, one warning per process). Tool-call turns and
+  `schema: null` routes are handled per spec §4.1. `runStream` inherits it.
+- **`FailureClass "quality_rejected"`** — retryable (own non-fungible budget),
+  never terminal, with an errors-that-teach entry (`docsUrl`
+  → routing-reliability.html#quality_rejected) and a suggested `policy.retries`
+  edit. `policy.retries.quality_rejected` is accepted by the loader.
+- **`RunMeta.quality?`** `{ score, verdict, judgedBy }` — present when a gate ran
+  (`verdict: "pass"` on a returned run; `judgedBy: null` for an SDK predicate).
+  SDK-owned `tags.quality` / `tags.quality_score` are stamped on the attempt row.
+- **`rig.bakeoff` / `rig.bakeoffFromEvalCases` gain `qualityGate?`** — scored on
+  each replayed sample into `ReplaySampleResult.valueAccuracy` + `qualityPassed`;
+  `VariantMetrics` gains `qualityPassCi95` and `effectiveUsdPer1kPassing` (beside,
+  never redefining, `effectiveUsdPer1kConformant`), and `computeProposal` gains a
+  third clause (challenger quality-pass CI-lower ≥ incumbent's). The `modelrig
+  bakeoff` CLI cannot carry a function — use programmatic `rig.bakeoff` until Q2.
+- **Exports:** `QualityGate`, `QualityVerdict`, `QualityGateMeta` from the root.
+- **Boundary B1 rewritten** to "Route-declared judge (hosted lane)" (`not_built`,
+  tripwire `/\bjudge\??\s*:/`) now the SDK predicate ships; the routing-reliability
+  page, llms.txt (C11), README, `/definitions`, and the v4 setup prompt are swept
+  to match.
+
+Project tree (G2b) — the console gains `/projects` (picker) and `/projects/[value]`
+(project → run → step → call → attempts, cost + quality + lane on every node) over
+one new read, `GET /v1/console/tree`. Server + console only; the SDK change below
+is documentation. ZERO DDL.
+
+- **`@step` / `step()` sets the ambient step key.** Documented in `step.ts` that
+  entering a step scope mints `label#n` and makes it the run context's ambient
+  step: an `artifact.save` with no explicit `stepKey` attaches to it — for saves
+  inside the scope AND for saves after it, until the next step mints a new key
+  (the ambient key is a running "most recent step", not a lexical bracket).
+  `run-context.ts` clarifies that within one run's parallel step waves only an
+  explicit `stepKey` prevents cross-attribution (`run.scope()` separates distinct
+  runs, not overlapping steps in one run).
+
+Screening (B0 + B1) — two bug-class fixes and a shared detector contract.
+- **Bake-off review samples are scrubbed before they persist.** A `--keep-outputs`
+  bake-off now runs each sample's input and output through the rig's content-custody
+  scrub mode (`detect` counts onto `arm_meta.scrub`, `redact` masks, `block` keeps a
+  metadata-only sample). Default `off` ⇒ rows byte-identical to before.
+- **Proxy role enum.** `messages[].role` must be `system|developer|user|assistant|tool`
+  (`developer` normalises to the system slot); anything else is rejected with a
+  teaching message before dispatch (HTTP 422). Previously an arbitrary role was cast
+  through.
+- **Raw-lane multi-turn encoding (behaviour change).** A genuine multi-turn transcript
+  is now flattened into the adapter's user prompt as one JSON object per line instead
+  of `role: content` text, so a crafted turn can never read as a synthetic system
+  line. Single-user-turn (and system + single user) requests are byte-identical.
+- **Shared `Detector` interface.** `Detector`, `DetectionSpan`, `detectSpans`, and
+  `scrubDetectors` are exported from the package root; `scrubContent` is unchanged.
+  `buildBakeoffSamples` gains a scrub-resolver argument.
+
+## [0.5.0] — 2026-08-22
+
+Released with the docs/registry mirror in one R-A ceremony. Headline for router
+codebases: `rig.runRaw` inside a run context now records a step (the `/runs`
+tree fills in), raw-lane timeouts are tier-aware and teach, the `trial:`
+candidate flag + console picker ship (deploy the server before any YAML uses
+`trial:`), and the console rollups show cost and quality together by route or
+tag. `modelrig-oracle` 0.2.0 ships alongside (`propose_candidates`).
+
+Raw-lane defaults (#245) — the raw lane teaches its own timeouts. The `timeoutMs`
+default becomes tier-aware, a `timeout` failure carries a teaching `fixHint`, an
+empty no-tools gemini response carries a hint naming grounding, and the raw-lane
+defaults are published as a generated table that cannot drift from the runtime.
+
+Declare-first experiment ergonomics (routing-console R3, #239) — a `trial:`
+candidate you declare but never serve live, a console challenger picker, and a
+one-command bake-off that measures it on real captured traffic before you
+promote it.
+
+### Changed
+
+- **Raw-lane `timeoutMs` default is now tier-aware** (was a flat 120 s):
+  `standard`/`priority` default to 300 s and `flex` to 900 s (the 15-minute budget
+  ModelRig advertises) — a passthrough should not be shorter-fused than the tier it
+  advertises, and grounded/large prompts routinely exceed 120 s even at standard.
+  An explicit `timeoutMs` **always wins** (no pre-dispatch rejection — a short
+  deadline is a choice), and a `timeout` failure on the raw lane now carries a
+  `fixHint` naming the elapsed time, the tier, that tier's budget, and the two
+  fixes (raise `timeoutMs`, or drop to `serviceTier: "standard"`). The routed
+  lane's required `policy.timeout_ms` is unchanged. Callers who relied on a fast
+  120 s failure should set `timeoutMs` explicitly.
+
+### Added — declare-first experiment (routing-console R3)
+
+- **`candidates[].trial: true`** (route bundle) — a DECLARED candidate that is
+  NEVER served live. The loader parses and validates it, marks an all-trial
+  variant `trialOnly`, and warns on a trial candidate no variant references;
+  `resolveCandidates` drops trial candidates on every live path (`rig.run`,
+  `runStream`, `runBundle`, the hosted gateway slug lane) and keeps them only in
+  bake-off replay. Selecting a trial variant on a live call is a
+  `RouteConfigError` that names the bake-off. This lets a customer measure a new
+  model on real captured traffic before promoting it. Additive: a route with no
+  trial candidate is byte-identical to before.
+  - CAVEAT — **deploy the server before any customer YAML carries `trial:`.** An
+    SDK/server older than 0.5.0 IGNORES `trial:` (the loader has always been
+    tolerant of unknown candidate keys) and its resolver has no trial-mode filter,
+    so it would serve the candidate LIVE — including an older hosted-gateway build
+    reading the flag through the mirror. Roll the server first, then let routes
+    start declaring `trial:`. The console's declare-YAML says so inline.
+- **`requirementPredicates(route)` / `evaluateChallengerEligibility(...)`** (new
+  exports) — decompose a route's `require` + json rung into capability-flag
+  requirements and answer "can this model be a trial challenger?", pinned in
+  agreement with the live resolver's `satisfiesRequirements`.
+- **`buildDeclareTrialChange(bundle, challenger)` / `trialVariantName(...)`** and
+  a trial-winner branch of **`buildSwapChange`** (new exports) — the paste-ready
+  declare-YAML (challenger appended `trial: true`, a `trial-<provider>-<model>`
+  variant, `version: n+1`) and the promote-YAML that drops the winner's trial
+  flag, moves it to position 0, and removes the trial variant. No filesystem, no
+  git — the human applies the PR (F6).
+- **`ManualChallengerTrigger` + `manualChallengerTriggerHash`**, and
+  **`proposalsClient.createManualChallenger`** — a console/oracle-initiated
+  declare-first proposal (`kind:"bakeoff"`, idempotent per `(route, challenger)`).
+- **`RunMeta`/`Rig` additions:** `rig.capabilities` and `rig.searchAvailable` are
+  exposed so the challenger picker answers eligibility with the exact live flags;
+  `modelrig bakeoff --from-proposal <id>` reads the trial variant from a
+  manual_challenger proposal (default vs that one variant).
+- **`resolveRouteBundle` / `hashResolvedBundle`** (new exports) — the canonical
+  `bundle_sha` helpers, used by the hosted-lane bake-off to prove the caller org's
+  mirrored bundle is the SAME bundle the deployment holds locally (a route-name
+  collision on a multi-tenant mirror otherwise baked off a foreign org's route).
+
+### Added
+
+- **A grounding-naming `fixHint` on an empty no-tools gemini response.** When a
+  call that declared no tools comes back with no text parts (a grounding turn
+  that produced only `functionCall` parts, for example), the `content_invalid`
+  "empty response" failure now names grounding as the likely cause. Note: the
+  text extraction itself is unchanged in effect — the SDK already concatenated
+  mixed `functionCall` + text parts — so the customer-zero empty-response report
+  (#245 secondary) remains an OPEN investigation pending a captured response.
+
+### Docs
+
+- **Raw-lane defaults table** — a generated table (`RAW_DEFAULTS`, rendered at the
+  `<!-- RAW_DEFAULTS -->` marker in `route-bundles.md` and in `llms.txt`) lists
+  every raw-lane knob and its default, with the tier-aware `timeoutMs` read from
+  the same constant the runtime resolves from — so the published defaults cannot
+  drift from what `rig.runRaw` sends. README / quickstart / migration-playbook
+  gain the "flex ⇒ set `timeoutMs` in minutes" one-liner.
+
+Raw-lane artifact graph + tags parity (G1). A `rig.runRaw` / `rig.runRawStream`
+call inside a `rig.artifacts.run.start` context now records a ground-truth step,
+so a router or report pipeline that wraps raw (BYOK) calls in a run context gets
+a real per-run tree at `/runs/[id]` instead of an empty one. ZERO DDL.
+
+### Added — raw-lane graph
+
+- **Raw calls record a ground-truth step.** Inside a run context,
+  `runRaw`/`runRawStream` emit a step (on success AND failure, buffered AND
+  streaming) whose `inference_id` joins the metered `lane=raw` row — parity with
+  the routed lane's auto-step. Fail-open: a buffer error there never affects the
+  returned result. Outside a run context the dispatch is byte-identical to 0.4.0
+  (no step row). Raw steps are **metadata only** — no artifact content is written
+  on the raw lane.
+- **`stepKey` on BOTH lanes.** `RunOptions.stepKey` (routed) and
+  `RawRunInput.stepKey` (raw) name the step. Resolution is `stepKey ?? tags.step
+  ?? auto` (`route#n` / `provider/model#n`), validated `[A-Za-z0-9_.:#-]{1,128}`;
+  an invalid value is warned once and dropped to the auto key (never thrown). A
+  raw row's step **task** is `tags.step ?? provider/model`.
+- **Ambient artifact step.** `artifact.save` without a `stepKey` now attaches to
+  the ambient step (the most recent step in the run context), so the two grouping
+  mechanisms compose — a raw/routed call and the artifacts saved around it land in
+  one group.
+- **Required-dimension hygiene on the raw lane.** `runRaw`/`runRawStream` run the
+  same once-per-process "required dimension missing from tags" warning the routed
+  lane does.
+
+### Added — docs
+
+- **instrumentation.md** gains "Tags, dimensions, and the graph": tags = flat
+  filter/rollup (`/costs`), the graph = the per-run tree (`/runs/[id]`), how
+  `stepKey` + ambient artifacts compose, and the raw-lane behavior.
+- **The per-service key rule** is now stated in the quickstart, the migration
+  playbook (F-6), and llms.txt: `MODELRIG_API_KEY` gates the exporter *per
+  process* — set it on EVERY service, worker, and cron, or a keyless process
+  buffers to its own local SQLite and ships nothing.
+- **route-bundles.md** points at where `rig.yaml` `dimensions` are declared; the
+  **README** documents `stepKey` on the raw lane.
+
 ## [0.4.0] — 2026-08-22
 
 Released with the docs/registry mirror in one R-A ceremony. NOT in this tag:

package/README.md

@@ -144,6 +144,19 @@ never stores it. It records one metered `lane=raw` telemetry row at the
 provider's list price (zero ModelRig margin) and throws a typed `RigFailureError`
 on failure, exactly like `rig.run`.
 
+Called inside a `rig.artifacts.run.start` context, `runRaw`/`runRawStream` also
+records a ground-truth **step** in the run's graph (`/runs/[id]`) — since
+modelrig 0.5.0, on success and on failure. Name it with `stepKey` (the same field
+`RunOptions` exposes on the routed lane; precedence `stepKey` > `tags.step` > an
+auto `provider/model#n` key), so a raw call and the artifacts you `save` around it
+group under one step.
+
+The raw-lane `timeoutMs` default is **tier-aware** since 0.5.0 — `standard`/
+`priority` 300 s, `flex` 900 s; an explicit `timeoutMs` always wins, and a
+`timeout` failure teaches the tier + budget in its `fixHint`. **flex ⇒ set
+`timeoutMs` in minutes** (default 900 s since 0.5.0). See the generated
+[raw-lane defaults table](https://modelrig.dev/route-bundles.html#raw-lane-defaults-rigrunraw).
+
 **runRaw-only construction (`routesDir: null`) — 0.4.0+.** A
 Lane-B deployment that ONLY calls `runRaw`/`runRawStream` has no routes. Build it
 with `routesDir: null` and `createRig` skips route loading, the `rig.yaml`
@@ -166,9 +179,55 @@ Every attempt is recorded in a local SQLite file (`.modelrig/telemetry.db`)
 tagged with your `tags`. Failures are typed —
 `RigFailureError.failure.class` is one of `content_invalid`, `capacity_shed`,
 `network`, `refusal`, `cache_invalid`, `timeout`, `config_auth`, `budget_exhausted`,
-`invariant_violation` — with a separate retry budget per class, so a flaky
+`invariant_violation`, `quality_rejected` — with a separate retry budget per class, so a flaky
 network never spends your content-validation retries. What each class means, whether it retries, and how it backs off is on [Routing & reliability](https://modelrig.dev/routing-reliability.html#5-failure-classes).
 
+## Metadata capture ("capture now, analyze later")
+
+Every inference records a versioned, **metadata-safe** attempt envelope
+(`AttemptMeta@v1`) on `inferences.meta` — in both the routed and the raw lane,
+from the same producer. It is the metadata the adapters already see and used to
+discard: finish reason, refusal, reasoning tokens, the provider-reported model
+snapshot and request id, coarse HTTP status / error code, sampling echo, tool
+names (declared kept, undeclared hashed) with counts, per-component cost, request
+and output **hashes** (never the content), request/response byte sizes, response
+char count, and a coarse context bucket.
+
+**It never contains free text.** Every field is an enum, a number, a boolean, a
+hash, or an identifier. The producer drops any string that is over 128 chars or
+contains whitespace, and the hosted ingest endpoint re-applies the same rule as a
+backstop — so analysis later is possible without prompts or outputs ever being
+stored here. Building the envelope never blocks or changes a run: on any failure
+it is simply absent (`meta` is null).
+
+Runs carry a companion `RunMeta@v1` on `runs.meta`, set at `run.start`:
+
+```ts
+const h = rig.run.start({
+  pipeline: "nightly-report",
+  meta: { v: 1, trigger: "cron", actor: "scheduler", config_hash: "…" },
+  // pipeline_sha auto-fills from git when you don't set it (never fails the run)
+});
+```
+
+### Reserved dimensions
+
+Business dimensions live in `tags`. Five keys are **well-known** and become
+first-class group-bys in cost and coverage views — use them verbatim so your data
+lines up with the console's built-ins:
+
+| key | meaning |
+|---|---|
+| `subject` | the customer/tenant the work is for — an **opaque** id, hashed if it looks like PII |
+| `feature` | the product feature the call serves |
+| `user` | the end user (an opaque id) |
+| `session` | the conversation/session the call belongs to |
+| `cost_center` | the team/budget to attribute spend to |
+
+`modelrig validate` warns on near-misses — a declared `rig.yaml` dimension named
+`customer`, `tenant`, `user_id`, `team`, … is nudged toward the reserved key it
+resembles. They stay ordinary tags; nothing is required.
+
 ## CLI
 
 The package ships a `modelrig` binary:
@@ -204,8 +263,11 @@ on the quickstart:
 - **`modelrig serve` needs the `modelrig-server` package** — it is not bundled
   here.
 - **What the ladder does and does not do.** Fall-through is structural and
-  infrastructural, never a quality judgment; the full boundary (quality gate,
-  raw-lane structure, latency ordering…) is generated and tripwire-tested on
+  infrastructural, plus an OPTIONAL SDK-lane `qualityGate` predicate you supply
+  (a schema-valid output it rejects becomes
+  `quality_rejected`); the route-declared
+  hosted judge is not built yet, and the full boundary (raw-lane structure, latency
+  ordering…) is generated and tripwire-tested on
   [Routing & reliability](https://modelrig.dev/routing-reliability.html#10-not-yet-implemented-the-boundary).
 
 ## License

package/dist/adapters/anthropic.d.ts.map

@@ -1 +1 @@
-{"version":3,"file":"anthropic.d.ts","sourceRoot":"","sources":["../../src/adapters/anthropic.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAEhB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAEhE,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,gBAAgB,CAAC;AAG9D,MAAM,WAAW,sBAAsB;IACrC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AASD;yDACyD;AACzD,wBAAgB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAEhE;AAED,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,cAAc,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAclF;AAED,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAwB3E;AAED,UAAU,aAAa;IACrB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,sDAAsD;IACtD,QAAQ,CAAC,MAAM,CAAC,EAAE,IAAI,CAAC;CACxB;AAED;;mEAEmE;AACnE,MAAM,WAAW,oBAAoB;IACnC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,aAAa,CAAC,EAAE;QAAE,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAA;KAAE,CAAC;CACzD;AAED,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B;wFACoF;IACpF,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;+DAC2D;IAC3D,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,GAAG,SAAS,oBAAoB,EAAE,CAAC;IAC3D,QAAQ,CAAC,QAAQ,EAAE,aAAa,CAAC;QAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IACtF,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,aAAa,EAAE,CAAC;IAC1C,QAAQ,CAAC,WAAW,CAAC,EAAE;QAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;KAAE,GAAG,uBAAuB,CAAC;CACnG;AAED,wBAAgB,oBAAoB,CAClC,KAAK,EAAE,MAAM,EACb,GAAG,EAAE,YAAY,EACjB,MAAM,UAAQ,GACb,qBAAqB,CAuDvB;AAED;iEACiE;AACjE,MAAM,WAAW,sBAAsB;IACrC,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,uBAAuB,CAAC,EAAE,MAAM,CAAC;IAC1C,QAAQ,CAAC,2BAA2B,CAAC,EAAE,MAAM,CAAC;IAC9C,QAAQ,CAAC,cAAc,CAAC,EAAE;QACxB,QAAQ,CAAC,yBAAyB,CAAC,EAAE,MAAM,CAAC;QAC5C,QAAQ,CAAC,yBAAyB,CAAC,EAAE,MAAM,CAAC;KAC7C,CAAC;CACH;AAED,0EAA0E;AAC1E,MAAM,WAAW,6BAA6B;IAC5C,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,gBAAgB,EAAE,MAAM,CAAC;IAClC,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;CACrC;AAED;;;;;;GAMG;AACH,wBAAgB,4BAA4B,CAC1C,CAAC,EAAE,sBAAsB,GACxB,6BAA6B,CAa/B;AAED,wBAAgB,sBAAsB,CAAC,MAAM,EAAE,sBAAsB,GAAG,eAAe,CAuMtF"}
+{"version":3,"file":"anthropic.d.ts","sourceRoot":"","sources":["../../src/adapters/anthropic.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;GAUG;AAGH,OAAO,KAAK,EAIV,cAAc,EACd,eAAe,EAEhB,MAAM,mBAAmB,CAAC;AAE3B,OAAO,KAAK,EAAE,iBAAiB,EAAE,YAAY,EAAE,MAAM,UAAU,CAAC;AAEhE,OAAO,KAAK,EAAE,uBAAuB,EAAE,MAAM,gBAAgB,CAAC;AAK9D,MAAM,WAAW,sBAAsB;IACrC,QAAQ,CAAC,MAAM,EAAE,MAAM,CAAC;CACzB;AASD;yDACyD;AACzD,wBAAgB,uBAAuB,CAAC,OAAO,EAAE,MAAM,GAAG,OAAO,CAEhE;AAED,wBAAgB,iBAAiB,CAAC,OAAO,EAAE,cAAc,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO,CAclF;AAED,wBAAgB,wBAAwB,CAAC,OAAO,EAAE,MAAM,GAAG,iBAAiB,CAwB3E;AAED,UAAU,aAAa;IACrB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,sDAAsD;IACtD,QAAQ,CAAC,MAAM,CAAC,EAAE,IAAI,CAAC;CACxB;AAED;;mEAEmE;AACnE,MAAM,WAAW,oBAAoB;IACnC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;IACtB,QAAQ,CAAC,aAAa,CAAC,EAAE;QAAE,QAAQ,CAAC,IAAI,EAAE,WAAW,CAAA;KAAE,CAAC;CACzD;AAED,MAAM,WAAW,qBAAqB;IACpC,QAAQ,CAAC,KAAK,EAAE,MAAM,CAAC;IACvB,QAAQ,CAAC,UAAU,EAAE,MAAM,CAAC;IAC5B;wFACoF;IACpF,QAAQ,CAAC,WAAW,CAAC,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IACxB;+DAC2D;IAC3D,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,GAAG,SAAS,oBAAoB,EAAE,CAAC;IAC3D,QAAQ,CAAC,QAAQ,EAAE,aAAa,CAAC;QAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,OAAO,EAAE,MAAM,CAAA;KAAE,CAAC,CAAC;IACtF,QAAQ,CAAC,MAAM,EAAE,IAAI,CAAC;IACtB,QAAQ,CAAC,KAAK,CAAC,EAAE,SAAS,aAAa,EAAE,CAAC;IAC1C,QAAQ,CAAC,WAAW,CAAC,EAAE;QAAE,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAC;QAAC,QAAQ,CAAC,IAAI,EAAE,MAAM,CAAA;KAAE,GAAG,uBAAuB,CAAC;CACnG;AAED,wBAAgB,oBAAoB,CAClC,KAAK,EAAE,MAAM,EACb,GAAG,EAAE,YAAY,EACjB,MAAM,UAAQ,GACb,qBAAqB,CAuDvB;AAED;iEACiE;AACjE,MAAM,WAAW,sBAAsB;IACrC,QAAQ,CAAC,YAAY,CAAC,EAAE,MAAM,CAAC;IAC/B,QAAQ,CAAC,uBAAuB,CAAC,EAAE,MAAM,CAAC;IAC1C,QAAQ,CAAC,2BAA2B,CAAC,EAAE,MAAM,CAAC;IAC9C,QAAQ,CAAC,cAAc,CAAC,EAAE;QACxB,QAAQ,CAAC,yBAAyB,CAAC,EAAE,MAAM,CAAC;QAC5C,QAAQ,CAAC,yBAAyB,CAAC,EAAE,MAAM,CAAC;KAC7C,CAAC;CACH;AAED,0EAA0E;AAC1E,MAAM,WAAW,6BAA6B;IAC5C,QAAQ,CAAC,QAAQ,EAAE,MAAM,CAAC;IAC1B,QAAQ,CAAC,YAAY,EAAE,MAAM,CAAC;IAC9B,QAAQ,CAAC,gBAAgB,EAAE,MAAM,CAAC;IAClC,QAAQ,CAAC,kBAAkB,EAAE,MAAM,CAAC;CACrC;AAED;;;;;;GAMG;AACH,wBAAgB,4BAA4B,CAC1C,CAAC,EAAE,sBAAsB,GACxB,6BAA6B,CAa/B;AAED,wBAAgB,sBAAsB,CAAC,MAAM,EAAE,sBAAsB,GAAG,eAAe,CAsOtF"}

package/dist/adapters/anthropic.js

@@ -22,6 +22,7 @@ exports.normalizeAnthropicInputUsage = normalizeAnthropicInputUsage;
 exports.createAnthropicAdapter = createAnthropicAdapter;
 const sdk_1 = __importDefault(require("@anthropic-ai/sdk"));
 const tool_mapping_1 = require("./tool-mapping");
+const attempt_meta_map_1 = require("./attempt-meta-map");
 const sampling_1 = require("./sampling");
 const DEFAULT_MAX_TOKENS = 32000;
 const OUTPUT_TOOL_NAME = "route_output";
@@ -159,6 +160,8 @@ function createAnthropicAdapter(config) {
         supportsTools: () => true,
         async call(req) {
             const model = req.candidate.model;
+            // WI-1: native attempt envelope for this call (see openai.ts for the pattern).
+            let attemptMeta;
             // F1: post-stream failures still billed tokens — attach usage.
             const fail = (cls, message, fixHint, usage) => ({
                 ok: false,
@@ -170,6 +173,7 @@ function createAnthropicAdapter(config) {
                     ...(fixHint !== undefined ? { fixHint } : {}),
                 },
                 ...(usage !== undefined ? { usage } : {}),
+                ...(attemptMeta !== undefined ? { meta: attemptMeta } : {}),
             });
             const wantJson = req.outputSchema !== null;
             try {
@@ -210,9 +214,20 @@ function createAnthropicAdapter(config) {
                 };
                 let completionTokens = 0;
                 let stopReason = null;
+                // WI-1 native fields: message.id + message.model arrive at message_start.
+                let responseModel = "";
+                let responseId = "";
                 for await (const event of events) {
-                    if (event.type === "message_start" && event.message?.usage !== undefined) {
-                        inputUsage = normalizeAnthropicInputUsage(event.message.usage);
+                    if (event.type === "message_start") {
+                        if (event.message?.usage !== undefined) {
+                            inputUsage = normalizeAnthropicInputUsage(event.message.usage);
+                        }
+                        if (responseModel === "" && typeof event.message?.model === "string") {
+                            responseModel = event.message.model;
+                        }
+                        if (responseId === "" && typeof event.message?.id === "string") {
+                            responseId = event.message.id;
+                        }
                     }
                     else if (event.type === "content_block_start" &&
                         wantTools &&
@@ -257,13 +272,25 @@ function createAnthropicAdapter(config) {
                 };
                 // Tool-call turn (spec §4.4): `arguments` is the raw JSON string the
                 // model streamed (OpenAI parity) — text may be empty, skip the guards.
-                if (wantTools && toolBlocks.size > 0) {
-                    const toolCalls = [...toolBlocks.values()].map((b, i) => ({
+                const toolCalls = wantTools && toolBlocks.size > 0
+                    ? [...toolBlocks.values()].map((b, i) => ({
                         id: b.id !== "" ? b.id : `call_${i}`,
                         type: "function",
                         function: { name: b.name, arguments: b.json === "" ? "{}" : b.json },
-                    }));
-                    return { ok: true, text, usage, toolCalls };
+                    }))
+                    : undefined;
+                // WI-1: native attempt envelope. refusal derives from stop_reason=refusal.
+                attemptMeta = (0, attempt_meta_map_1.buildAdapterMeta)("anthropic", {
+                    req,
+                    finishReasonRaw: stopReason,
+                    refusal: stopReason === "refusal",
+                    modelSnapshot: responseModel,
+                    providerRequestId: responseId,
+                    toolCalls,
+                    streaming: false,
+                });
+                if (toolCalls !== undefined) {
+                    return { ok: true, text, usage, toolCalls, meta: attemptMeta };
                 }
                 if (stopReason === "refusal") {
                     return fail("refusal", "Anthropic refused this request (stop_reason=refusal)", undefined, usage);
@@ -274,11 +301,12 @@ function createAnthropicAdapter(config) {
                 if (stopReason === "max_tokens" && wantJson) {
                     return fail("content_invalid", "Anthropic structured output truncated (stop_reason=max_tokens)", "raise maxOutputTokens for this route", usage);
                 }
-                return { ok: true, text, usage };
+                return { ok: true, text, usage, meta: attemptMeta };
             }
             catch (error) {
                 const message = error instanceof Error ? error.message : String(error);
                 const { cls, fixHint } = classifyAnthropicFailure(message);
+                attemptMeta = (0, attempt_meta_map_1.transportErrorMeta)(error);
                 return {
                     ok: false,
                     failure: {
@@ -289,6 +317,7 @@ function createAnthropicAdapter(config) {
                         cause: error,
                         ...(fixHint !== undefined ? { fixHint } : {}),
                     },
+                    meta: attemptMeta,
                 };
             }
         },

package/dist/adapters/attempt-meta-map.d.ts

@@ -0,0 +1,63 @@
+/**
+ * Per-adapter NATIVE metadata mapping (WI-1, spec §2). ONE place that turns a
+ * provider's native finish/refusal/usage/id signals into the additive
+ * `AdapterResult.meta` (`Partial<AttemptMetaV1>`). Adapters map ONLY what the
+ * provider reports; the producer (`telemetry/attempt-meta.ts`) adds
+ * hashes/sizes/cost and enforces the classification rule. Absent = unknown;
+ * nothing here is inferred.
+ *
+ * `refusal` derivation, per adapter:
+ *  - openai / grok / deepseek / deepinfra / fireworks (OpenAI dialect): the
+ *    stream carried a non-empty `choices[].delta.refusal` string, OR
+ *    `finish_reason === "content_filter"`.
+ *  - anthropic: `stop_reason === "refusal"`.
+ *  - gemini: `finishReason` is `SAFETY` / `PROHIBITED_CONTENT` / `RECITATION`.
+ * The refusal TEXT itself is never carried (content) — only the boolean.
+ */
+import type { AttemptMetaV1, FinishReason } from "../telemetry/attempt-meta";
+import type { AdapterRequest, ProviderId, ToolCall } from "../types/provider";
+/** Map a provider's native finish/stop reason to the `FinishReason` enum
+ * (spec §2). Unknown/absent ⇒ undefined (never inferred); an unrecognised
+ * non-empty value ⇒ `"other"`. */
+export declare function mapFinishReason(provider: ProviderId, raw: string | null | undefined): FinishReason | undefined;
+/** The declared sampling echo (spec §2 `sampling`) — the values ModelRig sent,
+ * read back from the request. topP → top_p, maxOutputTokens → max_output_tokens
+ * (the raw lane's top-level or the routed lane's `sampling` block). Numbers
+ * only; undefined when the caller declared none. */
+export declare function samplingEcho(req: AdapterRequest): {
+    temperature?: number;
+    top_p?: number;
+    max_output_tokens?: number;
+} | undefined;
+/** Aggregate normalized tool calls into `{ name, count }` (RAW names — the
+ * producer hashes undeclared ones). Undefined when the turn made none. */
+export declare function aggregateToolCalls(toolCalls: readonly ToolCall[] | undefined): {
+    name: string;
+    count: number;
+}[] | undefined;
+/** The signals an adapter collects on ONE completed call. All optional. */
+export interface AdapterMetaSignals {
+    readonly req: AdapterRequest;
+    readonly finishReasonRaw?: string | null;
+    readonly refusal?: boolean;
+    readonly reasoningTokens?: number;
+    readonly modelSnapshot?: string;
+    readonly providerRequestId?: string;
+    readonly region?: string;
+    readonly toolCalls?: readonly ToolCall[];
+    readonly grounding?: {
+        citations?: number;
+        searchProvider?: string;
+    };
+    /** true = the run loop streamed deltas (stream()); false/absent = buffered. */
+    readonly streaming?: boolean;
+}
+/** Build one adapter's NATIVE `meta` (spec §2). Only defined signals appear —
+ * everything else is unknown and omitted. The producer takes it from here. */
+export declare function buildAdapterMeta(provider: ProviderId, s: AdapterMetaSignals): Partial<AttemptMetaV1>;
+/** A transport/SDK error's coarse envelope (spec §2 `http_status`/`error_code`).
+ * Reads `error.status` (number) and a short `error.code` / `error.error.code`
+ * slug — NEVER the message (that is free text). The producer's classification
+ * rule drops a code that turns out to be free text. */
+export declare function transportErrorMeta(error: unknown): Partial<AttemptMetaV1>;
+//# sourceMappingURL=attempt-meta-map.d.ts.map

package/dist/adapters/attempt-meta-map.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"attempt-meta-map.d.ts","sourceRoot":"","sources":["../../src/adapters/attempt-meta-map.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;GAeG;AAEH,OAAO,KAAK,EAAE,aAAa,EAAE,YAAY,EAAE,MAAM,2BAA2B,CAAC;AAC7E,OAAO,KAAK,EAAE,cAAc,EAAE,UAAU,EAAE,QAAQ,EAAE,MAAM,mBAAmB,CAAC;AAE9E;;kCAEkC;AAClC,wBAAgB,eAAe,CAC7B,QAAQ,EAAE,UAAU,EACpB,GAAG,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,GAC7B,YAAY,GAAG,SAAS,CA6C1B;AAED;;;oDAGoD;AACpD,wBAAgB,YAAY,CAC1B,GAAG,EAAE,cAAc,GAClB;IAAE,WAAW,CAAC,EAAE,MAAM,CAAC;IAAC,KAAK,CAAC,EAAE,MAAM,CAAC;IAAC,iBAAiB,CAAC,EAAE,MAAM,CAAA;CAAE,GAAG,SAAS,CAOlF;AAED;0EAC0E;AAC1E,wBAAgB,kBAAkB,CAChC,SAAS,EAAE,SAAS,QAAQ,EAAE,GAAG,SAAS,GACzC;IAAE,IAAI,EAAE,MAAM,CAAC;IAAC,KAAK,EAAE,MAAM,CAAA;CAAE,EAAE,GAAG,SAAS,CAQ/C;AAED,2EAA2E;AAC3E,MAAM,WAAW,kBAAkB;IACjC,QAAQ,CAAC,GAAG,EAAE,cAAc,CAAC;IAC7B,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IACzC,QAAQ,CAAC,OAAO,CAAC,EAAE,OAAO,CAAC;IAC3B,QAAQ,CAAC,eAAe,CAAC,EAAE,MAAM,CAAC;IAClC,QAAQ,CAAC,aAAa,CAAC,EAAE,MAAM,CAAC;IAChC,QAAQ,CAAC,iBAAiB,CAAC,EAAE,MAAM,CAAC;IACpC,QAAQ,CAAC,MAAM,CAAC,EAAE,MAAM,CAAC;IACzB,QAAQ,CAAC,SAAS,CAAC,EAAE,SAAS,QAAQ,EAAE,CAAC;IACzC,QAAQ,CAAC,SAAS,CAAC,EAAE;QAAE,SAAS,CAAC,EAAE,MAAM,CAAC;QAAC,cAAc,CAAC,EAAE,MAAM,CAAA;KAAE,CAAC;IACrE,+EAA+E;IAC/E,QAAQ,CAAC,SAAS,CAAC,EAAE,OAAO,CAAC;CAC9B;AAED;8EAC8E;AAC9E,wBAAgB,gBAAgB,CAC9B,QAAQ,EAAE,UAAU,EACpB,CAAC,EAAE,kBAAkB,GACpB,OAAO,CAAC,aAAa,CAAC,CAyBxB;AAED;;;uDAGuD;AACvD,wBAAgB,kBAAkB,CAAC,KAAK,EAAE,OAAO,GAAG,OAAO,CAAC,aAAa,CAAC,CASzE"}