Harden RAG telemetry and production guardrails - #136
Merged
Conversation
- Add explicit degraded-mode signaling across answer/search APIs\n- Expand RAG telemetry with shared-cache miss diagnostics\n- Add prompt-injection eval cases and route-level latency thresholds\n- Fail production readiness when raw query text persistence is enabled\n\nCo-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 958ed42def
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 091eef45d9
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
The clozapine injection control deterministically classifies as medication_dose_risk, so align expectedQueryClass to avoid false query-class mismatches in eval quality gates.\n\nCo-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 86f722f0ee
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Co-authored-by: Copilot App <223556219+Copilot@users.noreply.github.com>
BigSimmo
enabled auto-merge
July 2, 2026 10:20
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 9324c66a88
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Comment on lines
+7245
to
+7246
| shared_cache_status: search.telemetry.shared_cache_status, | ||
| shared_cache_miss_reason: search.telemetry.shared_cache_miss_reason, |
There was a problem hiding this comment.
When this generated answer is stored by setCachedAnswer below, a first request that missed the shared search cache persists shared_cache_status: "miss" and the miss reason inside answer.latencyTimings. Later exact-query answer-cache hits go through getCachedAnswer, which only refreshes total_latency_ms, so those requests are still reported as shared-cache misses even though no shared-cache lookup happened; clear these fields before answer-cache storage or when serving answer-cache hits.
Useful? React with 👍 / 👎.
Comment on lines
5207
to
+5209
| const answerQualityTier: RagAnswer["answerQualityTier"] = | ||
| answer.answerQualityTier ?? | ||
| (answer.modelUsed ? "model_synthesis" : answer.routingMode === "extractive" ? "source_only" : undefined); | ||
| (answer.modelUsed ? "model_synthesis" : inferredSourceOnlyFallback ? "source_only" : undefined); |
There was a problem hiding this comment.
When a fast generation is discarded and recovered with buildExtractiveAnswer, the fallback keeps routingMode: "extractive" but line 7310 later restores modelUsed to the attempted model. In that scenario this ternary labels the response as model_synthesis, so the new degradedMode is inactive even though the client received a source-only extractive fallback; check the extractive/source-only route before modelUsed, or set answerQualityTier on the fallback explicitly.
Useful? React with 👍 / 👎.
BigSimmo
pushed a commit
that referenced
this pull request
Jul 30, 2026
…ocation CI caught what I did not: `static-pr` failed on `check:outstanding-issues` with #131-#134 duplicated and two `issues:next-id` markers. Cause: main's PR #1424 allocated #131-#134 for its own findings at the same time this branch held #131-#135, and `merge=union` did what union does — kept both sides under the same ids. That is #112's documented limit: union preserves concurrent appends but cannot allocate unique ids, so the structural gate is the only thing that catches it. My error was pushing without re-running that gate. The previous push resolved a `docs/branch-review-ledger.md` conflict, and I validated only that file before pushing to win the race against main — but the same merge also touched `docs/outstanding-issues.md`. `verify:cheap` would have caught it locally. Main's rows keep #131-#134 (already merged and referenced elsewhere); this branch's five renumber to #136-#140, one marker at 141, and the cold-cache cross-reference in process-hardening follows its row. Two of main's new rows also make a planned addition here redundant: #134 is the absent ledger merge driver and #133 is the outstanding-issues merge churn — both hit during this branch's work, both already captured upstream, so nothing new is filed for them.
BigSimmo
added a commit
that referenced
this pull request
Jul 30, 2026
…t did (#1427) * test(phone-scroll): prove the drag delivered before asserting the chrome hid CI run 30518866604 failed `ui-phone-scroll.spec.ts:423` on expect(getByTestId('universal-header-collapse')) .toHaveAttribute('data-scroll-hidden', 'true') // received "" after the full 10s auto-retry, and the classifier recorded it as "needs investigation". The assertion was right; the scroll never happened. `dragScrollBy` moved the scroller with `scrollTop +=`, which clamps silently at the end of the range, and returned nothing. When a page lays out shorter than the test assumed — content still settling under full-suite CI load — a 720px request delivers a fraction of that, the chrome correctly stays visible because document-detail chrome only hides past `scrollTop > 120`, and the failure surfaces ten seconds later looking like a product regression. The helper also resolved the scroll owner once up front, so a mid-drag layout change left it pushing an element that had stopped scrolling. - `dragScrollBy` now re-resolves the owner each step and returns the distance actually travelled. - `dragScrollUntilHidden` waits for the remaining downward runway (a condition wait, not a settle sleep), drags, and fails naming the shortfall if the drag could not cross the threshold. Used at the four sites that assert a hide immediately after a fixed-distance drag. - `addPhoneScrollRunway` waits for its 1600px filler to reach layout instead of sleeping 50ms. All 14 call sites already depend on that runway existing. Every assertion is byte-identical: a genuinely stuck header still fails exactly as before, once the drag is proven to have happened. No `.first()` was added (#93's stop rule) and no tolerance was relaxed. * ci: shard Production UI across three runners Measured on 2026-07-30 from the Actions API, two full UI-scope PR runs (30520443076, 30519912667): `Production UI` took 15m26-16m31 of a 16.8-18.6 minute run — 83-89% of wall clock — while every other job finished by minute 4 and then waited. Playwright itself reported `339 passed (13.5m)`; the balance is the isolated production build. That single job is also where the churn cost lands: 42% of PR runs in the sampled window were cancelled (25 of 60 completed), almost all superseded mid-Production-UI. Sharding is across runners, not workers. `workers: 1`, `fullyParallel: false` and `retries: 0` are unchanged inside each shard, so determinism is identical and per-runner load falls — which matters because #93's duplicate page root is load-dependent. `run-playwright.mjs` already forwards argv to `playwright test`, so `--shard` needed no runner change. The shard count is measured, not chosen. `fullyParallel: false` makes a spec file indivisible, so shard sizes are lumpy and more shards is not monotonically faster. Over the 340 required chromium tests: N=3 -> 121/106/113 largest 121 N=4 -> 121/106/96/17 largest 121 (same critical path, one more runner) N=6 -> 65/56/106/5/91/17 largest 106 N=5 -> 121/106/0/96/17 and N=8 -> two empty shards N=4 buys nothing over N=3, and any N with an empty shard would go red because `test:e2e:pr` deliberately omits `--pass-with-no-tests`. Expected critical path ~15.5 -> ~7 min, assuming per-test cost is roughly uniform. `fail-fast: false` so a failing shard cannot cancel its siblings and re-create the cancelled-vs-failed ambiguity #95 removed. Artifact names are shard-scoped because upload-artifact runs with `overwrite: false`. Branch protection requires only the `pr-required` aggregate, and `needs` on a matrix job yields the roll-up of all shards, so the aggregate is unchanged. Also adds `restore-keys` to both Playwright browser caches: without a prefix fallback a lockfile bump forced a cold browser download in every UI job at once, now three times over. * ci: bound the codex auto-resolve jobs and serialise the visual config Two inconsistencies found while mapping the pipeline, neither load-bearing but both silent: - `codex-autofix-review-comments.yml` was the only workflow in the repo with no `timeout-minutes` on either job, so both inherited GitHub's 360-minute default for work that reads PR metadata and posts one comment. - `playwright.visual.config.ts` set neither `workers` nor `fullyParallel`, so it inherited Playwright's default `workers = 50% of CPUs`. The production config pins both to serial deliberately; the visual lane was quietly opting out of the anti-flake posture the rest of the suite is configured for. * chore(gates): pin the documented gate count to the real chain Both numbers were wrong. `CLAUDE.md` said 24 static/consistency gates against an actual 25 — `check:assets` landed before that line was written, so it was wrong at authoring — and the `gates` skill said "check 2 of 26" against an actual 28. A stale count is not cosmetic here. The skill's whole point at that line is that `verify:cheap` stops at the first failure and everything after it never ran; an agent that believes the chain is 26 long cannot say how much a mid-chain failure skipped. `check:gate-manifest` already derives the real count from `verify:cheap:internal`, so it now asserts the documented numbers against it. The assertions fail closed: if the anchor phrasing disappears, the guard reports a lost anchor rather than passing on a document it no longer checks. Mutation-proven: reverting the skill to "26" fails with ".claude/skills/gates/SKILL.md says 26 where the chain has 28". * docs(issues): capture the CI review's deferred findings Five items from the CI/testing review that should not be changed blind: - #125 `ui_changed` matches all of `src/app`, so an API-only diff pays the 15-minute UI gate. Narrowing it can hide a real regression, so it needs a decision plus a compensating check rather than a quieter filter. - #126 the Playwright build writes to a per-run distDir, so Next's build cache is cold every run (~2 min, now ~29% of the sharded critical path). Fixing it means suppressing the runner's documented always-cleanup, which must not ship without executing the runner. - #127 the advisory UI lane spends ~3 min per UI PR on 5 mockup tests; there are currently zero `@quarantine` tests for it to cover. - #128 CI Triage is complete and self-tested but inert pending a repo variable. - #129 four `changes` outputs are computed and consumed by nothing, and `coverage_changed` fires on any non-doc file. * docs(ledger): record the ci-testing-review pass at this HEAD * ci: re-measure the shard split on the merged tree and refresh stale gate counts The merge changed both numbers this branch had recorded. Shard balance, re-measured against 342 required chromium tests (was 340): N=3 -> 121/111/110 largest 121 N=4 -> 121/106/98/17 largest 121 N=3 remains correct — one 121-test spec group bounds both, so N=4 spends an extra runner for the same critical path. The re-measure command is now in the workflow comment so the next person does not have to rediscover it. Gate counts: merging main added `check:gitleaks-pinned` and `check:pr-mergeability` to `verify:cheap:internal`, so the documented counts went stale the moment the merge landed — 25 -> 27 static, 28 -> 30 total. The guard added earlier in this branch caught it immediately rather than letting the docs drift again, which is the whole reason it exists. Also records the `ui-critical-fast` interaction: the UI critical path is now that 15-test fail-fast job plus the slowest shard, not the full 13.5-minute suite, so neither of this branch's pre-merge timings can be read on its own. * docs(issues): rebuild the ledger after a union-merge duplication The `merge=union` driver on `docs/outstanding-issues.md` preserves concurrent appends, but when both sides restructure the same region it concatenates them wholesale. Merging the latest main did exactly that: every open row appeared twice and both `issues:next-id` markers survived — 66 duplicate-id errors from `check:outstanding-issues`, which is precisely the failure that gate exists to catch (#112). Resolved by rebuilding on main's canonical file rather than by hand-editing the duplicated table: reset to `origin/main`, then re-apply this branch's five captured rows at #131-#135 (main had advanced its allocation to #130 while this branch was open, so the earlier #128-#132 numbering collided again) and re-apply the #127 narrowing note. Marker bumped to 136. Union merge cannot allocate unique ids; only the structural gate can catch when it has produced an invalid file. It did. * ci: record the measured shard result, correcting the predicted one First real run of the sharded shape (CI 30530618838, all green, whole run 13m39 against a 16.8-18.6 min unsharded baseline): ui-critical-fast 15 tests 3m14 Production UI (1) 121 tests 9m36 Production UI (2) 111 tests 6m54 Production UI (3) 110 tests 6m20 The prediction was wrong by ~40%. ~6.8 min was expected for the largest shard from 121/342 tests x 13.5 min; 9m36 happened. Per-test cost is not uniform — 111 tests took 6m54 while 121 took 9m36 — so a count-balanced split understates the slowest shard whenever the slow specs land in one group. `--shard` can only balance by count; balancing by duration would mean splitting the slow spec files themselves. The win is real but smaller than claimed, and the workflow comment and process-hardening now carry the measured numbers plus the reason the arithmetic misleads, so the next person re-measures instead of re-deriving. Also merges origin/main. The ledger conflict was GitHub-visible only: that file carries merge=union locally, which GitHub does not honour (#129). Resolved by keeping the one genuinely new record and dropping three that main already had elsewhere in the file — append-only forbids dropping a record that exists once, not keeping a second copy. Superseding record appended for this HEAD, since the prior one asserted a root cause that #127's trace evidence refutes. * docs(issues): renumber this branch's rows above main's concurrent allocation CI caught what I did not: `static-pr` failed on `check:outstanding-issues` with #131-#134 duplicated and two `issues:next-id` markers. Cause: main's PR #1424 allocated #131-#134 for its own findings at the same time this branch held #131-#135, and `merge=union` did what union does — kept both sides under the same ids. That is #112's documented limit: union preserves concurrent appends but cannot allocate unique ids, so the structural gate is the only thing that catches it. My error was pushing without re-running that gate. The previous push resolved a `docs/branch-review-ledger.md` conflict, and I validated only that file before pushing to win the race against main — but the same merge also touched `docs/outstanding-issues.md`. `verify:cheap` would have caught it locally. Main's rows keep #131-#134 (already merged and referenced elsewhere); this branch's five renumber to #136-#140, one marker at 141, and the cold-cache cross-reference in process-hardening follows its row. Two of main's new rows also make a planned addition here redundant: #134 is the absent ledger merge driver and #133 is the outstanding-issues merge churn — both hit during this branch's work, both already captured upstream, so nothing new is filed for them. * docs(issues): rebuild against main's current id allocation The union merge duplicated the whole open and archive tables again (two header rows, every id twice) because main restructured the file while this branch held rows in it. Same resolution as before and for the same reason: rebuild on main's canonical file rather than hand-editing a doubled table, then re-apply this branch's five rows. Main is now at next-id=135, so they land as #135-#139 with the marker at 140. None of the five is duplicated upstream — checked by summary before re-applying. This is the third renumber of the same five rows in one PR. That is not a mistake being repeated, it is #133 ("outstanding-issues conflicts on nearly every main advance") happening: any branch that holds rows in this file re-collides every time main lands one. Worth weighing whether captures should land in their own PR ahead of the work rather than riding along with it. * docs: record PR 1427 review --------- Co-authored-by: Claude <noreply@anthropic.com>
BigSimmo
added a commit
that referenced
this pull request
Jul 30, 2026
13 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Validation