* test(qa): gate parity prose scenarios on real tool calls
Closes criterion 2 of the GPT-5.4 parity completion gate in #64227 ('no
fake progress / fake tool completion') for the two first/second-wave
parity scenarios that can currently pass with a prose-only reply.
Background: the scenario framework already exposes tool-call assertions
via /debug/requests on the mock server (see approval-turn-tool-followthrough
for the pattern). Most parity scenarios use this seam to require a specific
plannedToolName, but source-docs-discovery-report and subagent-handoff
only checked the assistant's prose text, which means a model could fabricate:
- a Worked / Failed / Blocked / Follow-up report without ever calling
the read tool on the docs / source files the prompt named
- three labeled 'Delegated task', 'Result', 'Evidence' sections without
ever calling sessions_spawn to delegate
Both gaps are fake-progress loopholes for the parity gate.
Changes:
- source-docs-discovery-report: require at least one read tool call tied
to the 'worked, failed, blocked' prompt in /debug/requests. Failure
message dumps the observed plannedToolName list for debugging.
- subagent-handoff: require at least one sessions_spawn tool call tied
to the 'delegate' / 'subagent handoff' prompt in /debug/requests. Same
debug-friendly failure message.
Both assertions are gated behind !env.mock so they no-op in live-frontier
mode where the real provider exposes plannedToolName through a different
channel (or not at all).
Not touched: memory-recall is also in the parity pack but its pass path
is legitimately 'read the fact from prior-turn context'. That is a valid
recall strategy, not fake progress, so it is out of scope for this PR.
memory-recall's fake-progress story (no real memory_search call) would
require bigger mock-server changes and belongs in a follow-up that
extends the mock memory pipeline.
Validation:
- pnpm test extensions/qa-lab/src/scenario-catalog.test.ts
Refs #64227
* test(qa): fix case-sensitive tool-call assertions and dedupe debug fetch
Addresses loop-6 review feedback on PR #64681:
1. Copilot / Greptile / codex-connector all flagged that the discovery
scenario's .includes('worked, failed, blocked') assertion is
case-sensitive but the real prompt says 'Worked, Failed, Blocked...',
so the mock-mode assertion never matches. Fix: lowercase-normalize
allInputText before the contains check.
2. Greptile P2: the expr and message.expr each called fetchJson
separately, incurring two round-trips to /debug/requests. Fix: hoist
the fetch to a set step (discoveryDebugRequests / subagentDebugRequests)
and reuse the snapshot.
3. Copilot: the subagent-handoff assertion scanned the entire request
log and matched the first request with 'delegate' in its input text,
which could false-pass on a stale prior scenario. Fix: reverse the
array and take the most recent matching request instead.
Validation: pnpm test extensions/qa-lab/src/scenario-catalog.test.ts
(4/4 pass).
Refs #64227
* test(qa): narrow subagent-handoff tool-call assertion to pre-tool requests
Pass-2 codex-connector P1 finding on #64681: the reverse-find pattern I
used on pass 1 usually lands on the FOLLOW-UP request after the mock
runs sessions_spawn, not the pre-tool planning request that actually
has plannedToolName === 'sessions_spawn'. The mock only plans that tool
on requests with !toolOutput (mock-openai-server.ts:662), so the
post-tool request has plannedToolName unset and the assertion fails
even when the handoff succeeded.
Fix: switch the assertion back to a forward .some() match but add a
!request.toolOutput filter so the match is pinned to the pre-tool
planning phase. The case-insensitive regex, the fetchJson dedupe, and
the failure-message diagnostic from pass 1 are unchanged.
Validation: pnpm test extensions/qa-lab/src/scenario-catalog.test.ts
(4/4 pass).
Refs #64227
* test(qa): pin subagent-handoff tool-call assertion to scenario prompt
Addresses the pass-3 codex-connector P1 on #64681: the pass-2 fix
filtered to pre-tool requests but still used a broad
`/delegate|subagent handoff/i` regex. The `subagent-fanout-synthesis`
scenario runs BEFORE `subagent-handoff` in catalog order (scenarios
are sorted by path), and the fanout prompt reads
'Subagent fanout synthesis check: delegate exactly two bounded
subagents sequentially' — which contains 'delegate' and also plans
sessions_spawn pre-tool. That produces a cross-scenario false pass
where the fanout's earlier sessions_spawn request satisfies the
handoff assertion even when the handoff run never delegates.
Fix: tighten the input-text match from `/delegate|subagent handoff/i`
to `/delegate one bounded qa task/i`, which is the exact scenario-
unique substring from the `subagent-handoff` config.prompt. That
pins the assertion to this scenario's request window and closes the
cross-scenario false positive.
Validation: pnpm test extensions/qa-lab/src/scenario-catalog.test.ts
(4/4 pass).
Refs #64227
* test(qa): align parity assertion comments with actual filter logic
Addresses two loop-7 Copilot findings on PR #64681:
1. source-docs-discovery-report.md: the explanatory comment said the
debug request log was 'lowercased for case-insensitive matching',
but the code actually lowercases each request's allInputText inline
inside the .some() predicate, not the discoveryDebugRequests
snapshot. Rewrite the comment to describe the inline-lowercase
pattern so a future reader matches the code they see.
2. subagent-handoff.md: the comment said the assertion 'must be
pinned to THIS scenario's request window' but the implementation
actually relies on matching a scenario-unique prompt substring
(/delegate one bounded qa task/i), not a request-window. Rewrite
the comment to describe the substring pinning and keep the
pre-tool filter rationale intact.
No runtime change; comment-only fix to keep reviewer expectations
aligned with the actual assertion shape.
Validation: pnpm test extensions/qa-lab/src/scenario-catalog.test.ts
(4/4 pass).
Refs #64227
* test(qa): extend tool-call assertions to image-understanding, subagent-fanout, and capability-flip scenarios
* Guard mock-only image parity assertions
* Expand agentic parity second wave
* test(qa): pad parity suspicious-pass isolation to second wave
* qa-lab: parametrize parity report title and drop stale first-wave comment
Addresses two loop-7 Copilot findings on PR #64662:
1. Hard-coded 'GPT-5.4 / Opus 4.6' markdown H1: the renderer now uses a
template string that interpolates candidateLabel and baselineLabel, so
any parity run (not only gpt-5.4 vs opus 4.6) renders an accurate
title in saved reports. Default CLI flags still produce
openai/gpt-5.4 vs anthropic/claude-opus-4-6 as the baseline pair.
2. Stale 'declared first-wave parity scenarios' comment in
scopeSummaryToParityPack: the parity pack is now the ten-scenario
first-wave+second-wave set (PR D + PR E). Comment updated to drop
the first-wave qualifier and name the full QA_AGENTIC_PARITY_SCENARIOS
constant the scope is filtering against.
New regression: 'parametrizes the markdown header from the comparison
labels' — asserts that non-default labels (openai/gpt-5.4-alt vs
openai/gpt-5.4) render in the H1.
Validation: pnpm test extensions/qa-lab/src/agentic-parity-report.test.ts
(13/13 pass).
Refs #64227
* qa-lab: fail parity gate on required scenario failures regardless of baseline parity
* test(qa): update readable-report test to cover all 10 parity scenarios
* qa-lab: strengthen parity-report fake-success detector and verify run.primaryProvider labels
* Tighten parity label and scenario checks
* fix: tighten parity label provenance checks
* fix: scope parity tool-call metrics to tool lanes
* Fix parity report label and fake-success checks
* fix(qa): tighten parity report edge cases
* qa-lab: add Anthropic /v1/messages mock route for parity baseline
Closes the last local-runnability gap on criterion 5 of the GPT-5.4 parity
completion gate in #64227 ('the parity gate shows GPT-5.4 matches or beats
Opus 4.6 on the agreed metrics').
Background: the parity gate needs two comparable scenario runs - one
against openai/gpt-5.4 and one against anthropic/claude-opus-4-6 - so the
aggregate metrics and verdict in PR D (#64441) can be computed. Today the
qa-lab mock server only implements /v1/responses, so the baseline run
against Claude Opus 4.6 requires a real Anthropic API key. That makes the
gate impossible to prove end-to-end from a local worktree and means the
CI story is always 'two real providers + quota + keys'.
This PR adds a /v1/messages Anthropic-compatible route to the existing
mock OpenAI server. The route is a thin adapter that:
- Parses Anthropic Messages API request shapes (system as string or
[{type:text,text}], messages with string or block content, text and
tool_result and tool_use and image blocks)
- Translates them into the ResponsesInputItem[] shape the existing shared
scenario dispatcher (buildResponsesPayload) already understands
- Calls the shared dispatcher so both the OpenAI and Anthropic lanes run
through the exact same scenario prompt-matching logic (same subagent
fanout state machine, same extractRememberedFact helper, same
'/debug/requests' telemetry)
- Converts the resulting OpenAI-format events back into an Anthropic
message response with text and tool_use content blocks and a correct
stop_reason (tool_use vs end_turn)
Non-streaming only: the QA suite runner falls back to non-streaming mock
mode so real Anthropic SSE isn't necessary for the parity baseline.
Also adds claude-opus-4-6 and claude-sonnet-4-6 to /v1/models so baseline
model-list probes from the suite runner resolve without extra config.
Tests added:
- advertises Anthropic claude-opus-4-6 baseline model on /v1/models
- dispatches an Anthropic /v1/messages read tool call for source discovery
prompts (tool_use stop_reason, correct input path, /debug/requests
records plannedToolName=read)
- dispatches Anthropic /v1/messages tool_result follow-ups through the
shared scenario logic (subagent-handoff two-stage flow: tool_use -
tool_result - 'Delegated task / Evidence' prose summary)
Local validation:
- pnpm test extensions/qa-lab/src/mock-openai-server.test.ts (18/18 pass)
- pnpm test extensions/qa-lab/src/mock-openai-server.test.ts extensions/qa-lab/src/cli.runtime.test.ts extensions/qa-lab/src/scenario-catalog.test.ts (47/47 pass)
Refs #64227
Unblocks #64441 (parity harness) and the forthcoming qa parity run wrapper
by giving the baseline lane a local-only mock path.
* qa-lab: fix Anthropic tool_result ordering in messages adapter
Addresses the loop-6 Copilot / Greptile finding on PR #64685: in
`convertAnthropicMessagesToResponsesInput`, `tool_result` blocks were
pushed to `items` inside the per-block loop while the surrounding
user/assistant message was only pushed after the loop finished. That
reordered the function_call_output BEFORE its parent user message
whenever a user turn mixed `tool_result` with fresh text/image blocks,
which broke `extractToolOutput` (it scans AFTER the last user-role
index; function_call_output placed BEFORE that index is invisible to it)
and made the downstream scenario dispatcher behave as if no tool output
had been returned on mixed-content turns.
Fix: buffer `tool_result` and `tool_use` blocks in local arrays during
the per-block loop, push the parent role message first (when it has any
text/image pieces), then push the accumulated function_call /
function_call_output items in original order. tool_result-only user
turns still omit the parent message as before, so the non-mixed
subagent-fanout-synthesis two-stage flow that already worked keeps
working.
Regression added:
- `places tool_result after the parent user message even in mixed-content
turns` — sends a user turn that mixes a `tool_result` block with a
trailing fresh text block, then inspects `/debug/last-request` to
assert that `toolOutput === 'SUBAGENT-OK'` (extractToolOutput found
the function_call_output AFTER the last user index) and
`prompt === 'Keep going with the fanout.'` (extractLastUserText picked
up the trailing fresh text).
Local validation: pnpm test extensions/qa-lab/src/mock-openai-server.test.ts
(19/19 pass).
Refs #64227
* qa-lab: reject Anthropic streaming and empty model in messages mock
* qa-lab: tag mock request snapshots with a provider variant so parity runs can diff per provider
* Handle invalid Anthropic mock JSON
* fix: wire mock parity providers by model ref
* fix(qa): support Anthropic message streaming in mock parity lane
* qa-lab: record provider/model/mode in qa-suite-summary.json
Closes the 'summary cannot be label-verified' half of criterion 5 on the
GPT-5.4 parity completion gate in #64227.
Background: the parity gate in #64441 compares two qa-suite-summary.json
files and trusts whatever candidateLabel / baselineLabel the caller
passes. Today the summary JSON only contains { scenarios, counts }, so
nothing in the summary records which provider/model the run actually
used. If a maintainer swaps candidate and baseline summary paths in a
parity-report call, the verdict is silently mislabeled and nobody can
retroactively verify which run produced which summary.
Changes:
- Add a 'run' block to qa-suite-summary.json with startedAt, finishedAt,
providerMode, primaryModel (+ provider and model splits),
alternateModel (+ provider and model splits), fastMode, concurrency,
scenarioIds (when explicitly filtered).
- Extract a pure 'buildQaSuiteSummaryJson(params)' helper so the summary
JSON shape is unit-testable and the parity gate (and any future parity
wrapper) can import the exact same type rather than reverse-engineering
the JSON shape at runtime.
- Thread 'scenarioIds' from 'runQaSuite' into writeQaSuiteArtifacts so
--scenario-ids flags are recorded in the summary.
Unit tests added (src/suite.summary-json.test.ts, 5 cases):
- records provider/model/mode so parity gates can verify labels
- includes scenarioIds in run metadata when provided
- records an Anthropic baseline lane cleanly for parity runs
- leaves split fields null when a model ref is malformed
- keeps scenarios and counts alongside the run metadata
This is additive: existing callers of qa-suite-summary.json continue to
see the same { scenarios, counts } shape, just with an extra run field.
No existing consumers of the JSON need to change.
The follow-up 'qa parity run' CLI wrapper (run the parity pack twice
against candidate + baseline, emit two labeled summaries in one command)
stacks cleanly on top of this change and will land as a separate PR
once #64441 and #64662 merge so the wrapper can call runQaParityReportCommand
directly.
Local validation:
- pnpm test extensions/qa-lab/src/suite.summary-json.test.ts (5/5 pass)
- pnpm test extensions/qa-lab/src/suite.summary-json.test.ts extensions/qa-lab/src/cli.runtime.test.ts extensions/qa-lab/src/scenario-catalog.test.ts (34/34 pass)
Refs #64227
Unblocks the final parity run for #64441 / #64662 by making summaries
self-describing.
* qa-lab: strengthen qa-suite-summary builder types and empty-array semantics
Addresses 4 loop-6 Copilot / codex-connector findings on PR #64689
(re-opened as #64789):
1. P2 codex + Copilot: empty `scenarioIds` array was serialized as
`[]` because of a truthiness check. The CLI passes an empty array
when --scenario is omitted, so full-suite runs would incorrectly
record an explicit empty selection. Fix: switch to a
`length > 0` check so '[] or undefined' both encode as `null`
in the summary run metadata.
2. Copilot: `buildQaSuiteSummaryJson` was exported for parity-gate
consumers but its return type was `Record<string, unknown>`, which
defeated the point of exporting it. Fix: introduce a concrete
`QaSuiteSummaryJson` type that matches the JSON shape 1-for-1 and
make the builder return it. Downstream code (parity gate, parity
run wrapper) can now import the type and keep consumers
type-checked.
3. Copilot: `QaSuiteSummaryJsonParams.providerMode` re-declared the
`'mock-openai' | 'live-frontier'` string union even though
`QaProviderMode` is already imported from model-selection.ts. Fix:
reuse `QaProviderMode` so provider-mode additions flow through
both types at once.
4. Copilot: test fixtures omitted `steps` from the fake scenario
results, creating shape drift with the real suite scenario-result
shape. Fix: pad the test fixtures with `steps: []` and tighten the
scenarioIds assertion to read `json.run.scenarioIds` directly (the
new concrete return type makes the type-cast unnecessary).
New regression: `treats an empty scenarioIds array as unspecified
(no filter)` — passes `scenarioIds: []` and asserts the summary
records `scenarioIds: null`.
Validation: pnpm test extensions/qa-lab/src/suite.summary-json.test.ts
(6/6 pass).
Refs #64227
* qa-lab: record executed scenarioIds in summary run metadata
Addresses the pass-3 codex-connector P2 on #64789 (repl of #64689):
`run.scenarioIds` was copied from the raw `params.scenarioIds`
caller input, but `runQaSuite` normalizes that input through
`selectQaSuiteScenarios` which dedupes via `Set` and reorders the
selection to catalog order. When callers repeat --scenario ids or
pass them in non-catalog order, the summary metadata drifted from
the scenarios actually executed, which can make parity/report
tooling treat equivalent runs as different or trust inaccurate
provenance.
Fix: both writeQaSuiteArtifacts call sites in runQaSuite now pass
`selectedCatalogScenarios.map(scenario => scenario.id)` instead of
`params?.scenarioIds`, so the summary records the post-selection
executed list. This also covers the full-suite case automatically
(the executed list is the full lane-filtered catalog), giving parity
consumers a stable record of exactly which scenarios landed in the
run regardless of how the caller phrased the request.
buildQaSuiteSummaryJson's `length > 0 ? [...] : null` pass-2
semantics are preserved so the public helper still treats an empty
array as 'unspecified' for any future caller that legitimately passes
one.
Validation: pnpm test extensions/qa-lab/src/suite.summary-json.test.ts
(6/6 pass).
Refs #64227
* qa-lab: preserve null scenarioIds for unfiltered suite runs
Addresses the pass-4 codex-connector P2 on #64789: the pass-3 fix
always passed `selectedCatalogScenarios.map(...)` to
writeQaSuiteArtifacts, which made unfiltered full-suite runs
indistinguishable from an explicit all-scenarios selection in the
summary metadata. The 'unfiltered → null' semantic (documented in
the buildQaSuiteSummaryJson JSDoc and exercised by the
"treats an empty scenarioIds array as unspecified" regression) was
lost.
Fix: both writeQaSuiteArtifacts call sites now condition on the
caller's original `params.scenarioIds`. When the caller passed an
explicit non-empty filter, record the post-selection executed list
(pass-3 behavior, preserving Set-dedupe + catalog-order
normalization). When the caller passed undefined or an empty array,
pass undefined to writeQaSuiteArtifacts so buildQaSuiteSummaryJson's
length-check serializes null (pass-2 behavior, preserving unfiltered
semantics).
This keeps both codex-connector findings satisfied simultaneously:
- explicit --scenario filter reorders/dedupes through the executed
list, not the raw caller input
- unfiltered full-suite run records null, not a full catalog dump
that would shadow "explicit all-scenarios" selections
Validation: pnpm test extensions/qa-lab/src/suite.summary-json.test.ts
(6/6 pass).
Refs #64227
* qa-lab: reuse QaProviderMode in writeQaSuiteArtifacts param type
* qa-lab: stage mock auth profiles so the parity gate runs without real credentials
* fix(qa): clean up mock auth staging follow-ups
* ci: add parity-gate workflow that runs the GPT-5.4 vs Opus 4.6 gate end-to-end against the qa-lab mock
* ci: use supported parity gate runner label
* ci: watch gateway changes in parity gate
* docs: pin parity runbook alternate models
* fix(ci): watch qa-channel parity inputs
* qa: roll up parity proof closeout
* qa: harden mock parity review fixes
* qa-lab: fix review findings — comment wording, placeholder key, exported type, ordering assertion, remove false-positive positive-tone detection
* qa: fix memory-recall scenario count, update criterion 2 comment, cache fetchJson in model-switch
* qa-lab: clean up positive-tone comment + fix stale test expectations
* qa: pin workflow Node version to 22.14.0 + fix stale label-match wording
* qa-lab: refresh mock provider routing expectation
* docs: drop stale parity rollup rewrite from proof slice
* qa: run parity gate against mock lane
* deps: sync qa-lab lockfile
* build: refresh a2ui bundle hash
* ci: widen parity gate triggers
---------
Co-authored-by: Eva <eva@100yen.org>
