diff options
| author | Adam Malczewski <[email protected]> | 2026-06-05 14:06:39 +0900 |
|---|---|---|
| committer | Adam Malczewski <[email protected]> | 2026-06-05 14:06:39 +0900 |
| commit | 9ae09aad5d8d6232c55932af0d496b888166065f (patch) | |
| tree | e62b59314ba065a531f9924151742c5adc7426ef /notes | |
| parent | 6733bbd47e6df8681fcf4b6815f82bd6b6922334 (diff) | |
| download | dispatch-9ae09aad5d8d6232c55932af0d496b888166065f.tar.gz dispatch-9ae09aad5d8d6232c55932af0d496b888166065f.zip | |
feat(observability): Phase A.2 — verbatim provider.request "after" capture + self-redaction (267 tests)
Threads the step span's correlated logger into provider.stream (new optional ProviderStreamOptions.logger) so provider-openai-compat opens a child provider.request span at the fetch edge, capturing the verbatim post-transform request + response status/cache-tokens/raw-error. Auth header self-redacted in the provider's OWN code (graduated mask tiers; no shared helper). Capture is fail-safe (never throws into the turn). Adds the first hermetic provider HTTP test (stream.test.ts: fetch mocked, 15 cases). Large payloads use attributes for now; the LogRecord.body channel is a deferred ABI design (notes §10).
Verified: tsc -b clean, 267 tests (250->+17), biome 0 warnings/0 infos. Live boot: provider.request shares turnId with prompt:before (before<->after diffable); auth-key leak count = 0 (self-redaction proven on a real request).
Diffstat (limited to 'notes')
| -rw-r--r-- | notes/observability-design.md | 32 |
1 files changed, 26 insertions, 6 deletions
diff --git a/notes/observability-design.md b/notes/observability-design.md index 6c09099..5e9c3a7 100644 --- a/notes/observability-design.md +++ b/notes/observability-design.md @@ -583,12 +583,32 @@ finish) → Unit 2 → host-bin wiring**. Optional overlap: start Unit 2 once Un **Verify:** `typecheck`/`test`/`check` clean; live boot → `host.logger` lines land in the journal file. -### Next step (after Phase A) — the "AFTER" capture -`provider.request` verbatim post-transform (D5) inside `provider-openai-compat`: the -exact serialized request bytes + raw response/error + cache tokens, auth header -self-redacted. This **completes full round-trip rebuild** and unlocks the -**before↔after diff** (kernel "before" vs wire "after" → pinpoints transform bugs). -Depends only on the Phase A Logger ABI; lives entirely in the provider extension. +### Phase A.2 — the "AFTER" capture (build plan) +`provider.request` verbatim post-transform (D5) inside `provider-openai-compat`: +exact serialized request + response status/cache-tokens/raw-error, auth self-redacted. +Completes full round-trip rebuild + the **before↔after diff**. +- **Contract (DONE, orchestrator):** `ProviderStreamOptions.logger?: Logger` + (`contracts/provider.ts`) — threads the step's correlated logger into `stream()` so + the `provider.request` span is a child of the step span (before↔after share + turnId/parentSpanId). Optional = non-breaking. +- **Unit K — kernel run-turn** (owner: kernel): pass the step span's logger into + `provider.stream(msgs, tools, { ...opts, logger })`. One file (`runtime/run-turn.ts`). +- **Unit P — provider-openai-compat** (owner): at the fetch edge, if `opts.logger`, + open a `provider.request` child span; capture the verbatim post-transform request + (URL, model, params, serialized body) + `cache_control` presence; on response, + status + cache-read/creation tokens (Usage) + (on error) raw error; **self-redact + the auth header in its own code** (graduated tiers, §6). First hermetic provider + HTTP test (`stream.test.ts`, mock `fetch` + real-capture fixtures). +- **Order:** contract frozen (done) → Unit K ∥ Unit P (disjoint: kernel vs provider). + +**DEFERRED — body-channel ABI (design decision, surface to user):** `LogRecord` has a +`body` field but the Logger/Span API exposes no way to set it — that's why +`prompt:before` (and now the request/response) use stringified `attributes`. Storing +large verbatim payloads in `body` (store-fat-serve-thin; both before & after) needs a +small ABI addition (e.g. `Span.setBody(body)`), worth doing before Phase B's query +layer. Until then captures use attributes — functional + reconstructable, just not +ideal for D9 `GROUP BY`. + *(Full per-extension prompt-segment provenance — D8 — comes later, with the context-filter chain.)* |
