Conversation
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
a8a10ed to
3956fff
Compare
|
FAILED — acceptance incomplete at The brief's earlier conflict/no-CI/draft caveats were checked against GitHub: this head is OPEN, not draft, MERGEABLE, and CI exists. The description still records 8 unresolved earlier local test failures: 7 matching its comparison checkpoint and 1 intermittent hard-reset shutdown proof failure reproduced on both revisions. The checkpoint already contains implementation Phases 1 and 2; it is not pristine main. Those local results were not rerun or relabeled here. Final GitHub recheck (2026-09-14 10:01 UTC): head unchanged; OPEN, draft=false, MERGEABLE. Check summary: 1 NEUTRAL, 24 SUCCESS. Observed on the exact candidate, one isolated iOS session:
Privacy: CLEAN for the fresh NOT RUN: fresh MCP cases for Why live acceptance stopped: first build hit the known simulator install race; the single prescribed stop-Metro/build retry installed and launched the Test App. Hermes then exposed one matching target, but the validator's final process probe returned Workspace/Test App base: Cleanup proven: owned runner closed, Metro stopped, integration restored, session released, transport stopped, and final validator cleanup reports empty authority plus absent owned simulator/listener/integration residue. Foreign devices were untouched. The subsystem-removal proposal is outside this verdict. 
|
|
PASS — tested refusal recognition/reporting at GitHub currently reports OPEN, not draft, MERGEABLE, with 24 successful checks and one neutral check. The earlier conflict/no-CI caveats no longer match that state. The description’s eight earlier local failures remain documented: seven matching its comparison checkpoint and one flaky hard-reset shutdown proof. That checkpoint includes implementation Phases 1 and 2; these failures were not rerun or relabeled here. The repaired workspace base was
The requested pre-bind SESSION_AUTHORITY_REQUIRED variant was not provoked: source auto-binding and earlier install/device gates emitted other codes. The foreign-Metro/UDID variant was also NOT RUN; no second Metro or foreign lane was used. Fresh reporting checks passed: 17 recognized occurrences / six complete-key groups. The Android runner group spans two tools; the main bundle group spans four. Every key, count, tool set, member signature, date and recorded provenance matched the store. Diagnostics did not record recovery. Both shipped host CLIs agreed in JSON and text after normalizing generated-at timestamps only. JSON carries signatures/provenance; their omission from the systemic text section is intentional. Reporting left 24,193 bytes, mtime 1789384740834568538 ns, hash and directory contents unchanged. The earlier exact-head legacy result is carried forward as instructed: 12 legacy-derived groups / 942 occurrences, without rewriting history. Structured-code/prose precedence, first-content-item selection and the 16 KiB UTF-8 boundaries are unit-verified by existing coverage; no synthetic acceptance harness was added. Real out-of-set codes remained Privacy: no UUID, known device identifier/serial, hostname or local/home path was found in any field of the 19 fresh stored records. All use redaction version 2. The actual Metro refusal contained no UDID, so this is not live proof of the unobserved UDID-bearing variant. Reverse cleanup passed with Evidence: 
|
|
FAILED — independent exact-head QA of #1034 at A failed native Android journey cannot become a full PASS. This is evidence for the captain, not a merge gate. No approval or merge decision. The Android first native surface after a fresh owned build/pin/attachOnly was the Expo development-server picker, not the app. Pin exhausted 120s with 0 Hermes targets. The APK does contain iOS at this same pair passed: usable Welcome baseline, hide Reporting recheck on this run’s private store: four real codes ( Cleanup: owned runner/Metro/transport closed, integration restored, owned simulator and emulator removed. Foreign inventory unchanged (two booted iOS simulators; physical USB phone untouched). 
|
|
FAILED — QA input-correction experiment for #1034 at Binding the existing Expo Dev Client URL before the owned Android build did not produce a visible app result from the original saved action. This is evidence for the captain, not a merge gate. No approval or merge decision. No product source was patched. What changed vs prior Android runs. Previous exact-head QA bound the device without a URL, then built, then pinned (0 Hermes targets / 120s; first native surface was the Expo picker). This run called What still failed. Picker-first and iOS, CLI/reporting, and Desktop were NOT RUN (preserved / out of scope). Foreign iOS simulators and the attached USB phone were not touched. Owned emulator and Metro were removed; package scripts restored to The bind-first input correction did not avoid the Android saved-action wall. No product patch is justified from this run. |
![cursor[bot]](https://avatars.githubusercontent.com/in/1210556?s=60&v=4){kind=small}
| if (!isAuthorityRefusalCode(code)) return null; | ||
| return { | ||
| code, | ||
| axis: AUTHORITY_AXES.find((candidate) => candidate === axis) ?? null, |
There was a problem hiding this comment.
Product-wrong: axis in the systemic key splits one refusal into several rows by emission path.
This reads axis from meta.axis and authorityRefusalSystemicKey hashes it into the group key. But the product does not emit axis consistently for the same code:
SessionAuthorityErrorthroughauthorityGate.wrap→authorityFailure→failResult(msg, code, authorityErrorMeta(err))→meta.axisset fromregistry.ts errorAxes(BUNDLE_HANDSHAKE_UNAVAILABLE → 'B',RUNNER_OWNERSHIP_MISMATCH → 'R', …).- Plain
Error('BUNDLE_HANDSHAKE_UNAVAILABLE: …')(index.ts:910,:1395,dev-client-authority.ts:90,:286) →authorityFailurefallback →failResult(message, code)with no meta →axis: null. failResult(msg, 'RUNNER_OWNERSHIP_MISMATCH')atagent-device-wrapper.ts:2040/tools/device-session.ts:487→ no meta →axis: null.- Every pre-PR row: the stored symptom is prose
CODE: message, sodecodeLegacyAuthorityRefusalgivesaxis: null. Legacy evidence can never join a gate-emitted group. HANDOFF_NOT_AUTHORIZEDis not inerrorAxesat all, so even the gate emits it without an axis.
Reproduced with the shipped CLI at this head (claude-plugin/rn-dev-agent-core/dist/experience-trends.js --json) on three BUNDLE_HANDSHAKE_UNAVAILABLE/android rows (gate-emitted axis:'B' count 3; no-meta axis:null count 2; legacy prose count 4):
[
{ "code": "BUNDLE_HANDSHAKE_UNAVAILABLE", "axis": null, "platform": "android", "count": 6, "tools": ["cdp_status","device_snapshot"], "provenance": ["legacy-derived","recorded"] },
{ "code": "BUNDLE_HANDSHAKE_UNAVAILABLE", "axis": "B", "platform": "android", "count": 3, "tools": ["cdp_run_action"], "provenance": ["recorded"] }
]That is #981's "N unrelated rows instead of one systemic pattern" again, at a coarser grain. experience-systemic-trends.test.ts:462-473 pins the split as intended (unknown-axis gets its own group), so the tests encode the gap rather than catch it; the "legacy joins recorded" fixture there uses a JSON symptom the pre-PR recorder never wrote (legacy symptoms were extractSymptom → parsed.error, i.e. prose).
Suggested fix: for the six supported codes the axis is a fixed function of the code (authority-gate.ts axisErrors / registry.ts errorAxes). Derive it from a code→axis table here and treat a supplied meta.axis that disagrees as conflicting metadata (refuse/unknown, per this PR's own precedence rule) — or drop axis from the key and keep it as a reported attribute. Add a regression: same code + platform, one row with meta.axis, one without, one legacy prose row → exactly one systemic group. The changeset's "including unambiguous legacy evidence" only becomes true after this.
| const provenance = Object.hasOwn(record, 'authorityRefusal') ? 'recorded' : 'legacy-derived'; | ||
| const facts = | ||
| provenance === 'recorded' | ||
| ? recordedRefusalFacts(record.authorityRefusal) | ||
| : decodeLegacyAuthorityRefusal(record.symptom); |
There was a problem hiding this comment.
Correctness: the reporter re-admits via prose what the recorder declined via structure.
Every record without an authorityRefusal key is treated as legacy and decoded from record.symptom. But post-PR the recorder also writes key-less records when it evaluated the event and concluded "not a refusal":
- structured
codeoutside the six while theerrorprose starts with a refusal code → recorder: structured wins, no facts; reporter: prose wins, refusal; - envelope over
MAX_AUTHORITY_ENVELOPE_BYTES(16 KiB) →decodeAuthorityRefusalPayloadreturnsnull, but the symptom is stillCODE: …→ reporter admits it aslegacy-derivedwithaxis: null(feeding the axis split noted inauthority-refusal.ts).
So "Structured codes take precedence over prose, including unknown codes" holds for the recorder but not for the report. Suggested fix: make the recorder's verdict explicit — write authorityRefusal: null on every new failure record in buildFailureRecord — and here skip records where the key is present and null; legacy-decode only when the key is genuinely absent. recordedRefusalFacts' malformed-extension handling is fine as is.
| const REFUSAL_CAUSES = { | ||
| SESSION_AUTHORITY_REQUIRED: [], | ||
| METRO_ORIGIN_MISMATCH: [], | ||
| RUNNER_OWNERSHIP_MISMATCH: [], | ||
| HANDOFF_NOT_AUTHORIZED: [], | ||
| NON_GIT_MANIFEST_REQUIRED: [], | ||
| BUNDLE_HANDSHAKE_UNAVAILABLE: [], | ||
| } as const satisfies Record<AuthorityRefusalCode, readonly string[]>; |
There was a problem hiding this comment.
Low: cause is a dead dimension right now. Every vocabulary here is empty, and authorityErrorMeta (registry.ts:375) never emits cause even though SessionAuthorityError.details.cause exists (startup-cleanup.ts:205 sets managed-metro-stop-proof-missing). Result: cause is always null in the key and the text report always prints cause: unknown. Not blocking, but either wire details.cause into authorityErrorMeta with a first vocabulary entry, or drop the column from the text output until one exists so the report does not advertise a dimension nothing populates.
| // Bump whenever a redaction rule changes: stored records stamped with an older | ||
| // version are re-sanitized under the current rules before the next rewrite. | ||
| export const REDACTION_RULES_VERSION = 1; | ||
| export const REDACTION_RULES_VERSION = 2; |
There was a problem hiding this comment.
Low / non-blocking, but two silent side effects of the version bump plus the new [ID_REDACTED] rule (line 81) are worth a line in the changeset:
- Legacy rows carry
evidencePointers: ["event:<uuid-with-dashes>", …]. Re-sanitization turns each intoevent:[ID_REDACTED], and the nextboundedPointersSet-dedupe collapses up to three pointers into one. Acceptable (they were opaque), but nothing announces it. normalizeSymptomShaperuns on the already-sanitized symptom. Pre-PR a UDID normalized to<id>; now it is the literal[id_redacted](11 chars, no digit, so the<id>rule does not match). A UUID-bearing failure that was recurring before the upgrade opens a fresh signature once after it, and the old row ages out over 14 days. One-time discontinuity, no action needed beyond awareness.
Code review summary (head
|
![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: c66ee38743
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".
| assert.ok(calls.includes('stale')); | ||
| }); | ||
|
|
||
| test('GH#202 runFlowParked: parking trace records begin, released, and committed', async () => { |
There was a problem hiding this comment.
These three new test cases expand a grandfathered .js test file, even though repository policy requires all new test code to use .ts or .tsx. Move the added coverage into a TypeScript test rather than growing the JavaScript baseline.
AGENTS.md reference: AGENTS.md:L45-L47
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: e4a2b29e09
ℹ️ About Codex in GitHub
Codex has been enabled to automatically 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 👍.
When you sign up for Codex through ChatGPT, Codex can also answer questions or update the PR, like "@codex address that feedback".
| [/~\/[A-Za-z0-9_./-]+/g, '[PATH_REDACTED]'], | ||
| [/\/(Users|home|opt|var|tmp|etc|private|Volumes)\/[A-Za-z0-9_./-]+/g, '[PATH_REDACTED]'], | ||
| [/(com|org|io|dev|net)\.[a-zA-Z0-9_-]+\.[a-zA-Z0-9_.-]+/g, '[BUNDLE_REDACTED]'], | ||
| [/[0-9a-f]{8}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{4}-[0-9a-f]{12}/gi, '[ID_REDACTED]'], |
There was a problem hiding this comment.
When a simulator UDID appears in the CDP bridge log or legacy telemetry, this new rule protects only in-process evidence: scripts/collect-feedback.sh still passes those sources through a redact sed program with no UUID rule before including them in submitted feedback. This leaves identifiers such as 12345678-1234-1234-1234-123456789ABC intact despite the nearby contract requiring these sanitizer lists to stay aligned; add the equivalent rule to the root collector and regenerate both packaged host copies.
AGENTS.md reference: AGENTS.md:L314-L323
Useful? React with 👍 / 👎.
There was a problem hiding this comment.
Landed on this head. Root scripts/collect-feedback.sh and both packaged host copies now apply the same UUID → [ID_REDACTED] rule as the in-process sanitizer (b2ac1254 / 8ae939d).
Take current main host bundles through the rebase, then rebuild so authority-refusal reporting ships in both plugin copies. Co-authored-by: Anton Lykhoyda <lykhoyda@gmail.com>
A recognized authority refusal persists the producer's error text after sanitizeString, and METRO_ORIGIN_MISMATCH names the bound device by its simulator UDID. The shared redaction rules did not cover UUIDs, so the UDID landed in patterns.jsonl. Add one UUID rule to REDACTION_RULES and bump REDACTION_RULES_VERSION so retained rows re-sanitize on their next rewrite. Evidence pointers keep a random id but drop the dashes so the rule does not collapse them. Regenerate both host runtimes.
…traces Axis is a reported attribute, not identity, so one code+platform stays one group. New failures persist an explicit non-membership verdict, and code-less action timeouts keep their runner traces.
The three parking-trace cases added in c66ee38 grew the grandfathered gh-202-maestro-flow-parks-l2.test.js. New test code must be TypeScript, so they now live in runner-diagnostics.test.ts beside the other withRunnerDiagnosticsContext coverage, and the .js file is restored to its main bytes. Call-order assertions already pinned by the existing JS cases are not repeated. Co-authored-by: Anton Lykhoyda <lykhoyda@gmail.com>
The in-process evidence rules gained a UUID rule in d4d7de4, but scripts/collect-feedback.sh still sanitized CDP bridge log and legacy telemetry tails with no such rule, so simulator UDIDs survived into submitted feedback. Add the equivalent sed rule to the root collector, regenerate both packaged host copies, and pin the collector path in the existing UDID redaction test. Co-authored-by: Anton Lykhoyda <lykhoyda@gmail.com>
Co-authored-by: Anton Lykhoyda <lykhoyda@gmail.com>
2262b6f to
8ae939d
Compare
Lykhoyda
changed the title
Lykhoyda
added
the
needs-qa
There was a problem hiding this comment.
HANDOFF: needs-qa
QA requested (needs-qa).
Linux automation cannot run device QA. GetSafe Project will spawn a new local Mac session.
QA_PR: #1034
PLATFORM: all
IMAGE_WIDTH: 800
POST_COMMENT: true
Claim: Report recognized authority refusals (SESSION_AUTHORITY_REQUIRED, METRO_ORIGIN_MISMATCH, RUNNER_OWNERSHIP_MISMATCH, HANDOFF_NOT_AUTHORIZED, NON_GIT_MANIFEST_REQUIRED, BUNDLE_HANDSHAKE_UNAVAILABLE) as one code-and-platform systemic group across gate, Error, failResult, and legacy-prose emission shapes (axis is reported but omitted from the grouping key; conflicting or missing axis becomes null). Persist an explicit authorityRefusal: null negative verdict on new and matching historical-row upserts so prose cannot re-admit non-membership, without backfilling untouched rows. Recognize code-less cdp_run_action TIMEOUT and producer timedOut and add observation-only flow-park / flow-stage diagnostics around existing park/stage/relaunch/cleanup/origin boundaries; do not change errors, cleanup, authority, or timeouts. The qa-observe-screen-match.yaml fixture is intentionally untouched. Prior note on this pair: Android QA FAILED; iOS already passed.
Changed paths:
packages/rn-dev-agent-core/src/experience/authority-refusal.tspackages/rn-dev-agent-core/src/experience/evidence.tspackages/rn-dev-agent-core/src/experience/trends.tspackages/rn-dev-agent-core/src/experience/runner-diagnostics.tspackages/rn-dev-agent-core/src/experience-trends.tspackages/rn-dev-agent-core/src/tools/maestro-run.tspackages/rn-dev-agent-core/test/unit/experience-authority-refusal.test.tspackages/rn-dev-agent-core/test/unit/experience-systemic-trends.test.tspackages/rn-dev-agent-core/test/unit/runner-diagnostics.test.tsseed-experience/common-failures.yamlscripts/collect-feedback.sh(and host copies).changeset/systemic-authority-refusals.md- Host runtime bundles under
packages/{claude,codex}-plugin/rn-dev-agent-core/dist/
Local runner must:
- New session (do not reuse an old chat; plugin ≥ 1.0.8 with /qa-pr)
- Version preflight vs GitHub main + latest release before any device work
- Own disposable git worktree (never the primary checkout)
- Own dedicated simulator/emulator named with the PR identity, e.g.
QA PR 1017 short-title - Plugin-repo PRs bind workspace test-app/
- Report back on this PR with gh --attach and
- Overall PASS or FAIL (PR/fix). Never PARTIAL.
Sent by Cursor Automation: New bug for QA
|
FAILED — Android FAILED · iOS NOT RUN · overall FAILED at exact head What this head is for: the reporting corrections (A1/A2) and the B1 timeout traces are published here; the conditional Identities
Android — FAILEDReplay call: B1 timeout trace — retained
There is no Inference, bounded: the long wait sits inside the managed relaunch boundary after native stage 0 returned — not before runner parking and not in initial runner execution. This does not prove the internal cause, and it does not prove that changing Confound recorded honestly: host load rose above 300 during the Android journey (samples 356.6, 332.7) despite the boot gate being satisfied. That is a material performance confound; it does not turn a failed acceptance into a pass, and it does not erase the trace. iOS — NOT RUNThe owned simulator was allocated but never booted. After Android cleanup the one-minute host load stayed above 280 (samples 346.4, 365.1, 354.8, 282.6), repeatedly failing the mandatory below-10 boot gate, and the run was finalized rather than booting into a starved host. No iOS build, baseline, replay, or repeat exists on this head. Missing platform coverage is FAILED overall; nothing about iOS was tested, so it is recorded as not run rather than as a product failure. Cleanup and an external inventory changeOwned cleanup is proven: runner close →
DispositionReturn the B1 trace to the implementation owner: the managed relaunch interval is now directly bounded by diagnostic events. Investigate that boundary under controlled host conditions before choosing a behavioral fix. The fresh iOS journey remains owed once the host is idle. #1037 is out of scope. No merge. 
|
|
PR: #1034 Additive iOS-only coverage from a second independent local Mac session at the same head. It fills the iOS NOT RUN gap in #1034 (comment); it does not re-run or contradict that run's Android TIMEOUT. Overall stays FAIL because Android failed there. This is initial needs-qa evidence, not a merge gate. No approval, branch edit, or merge.
Plugin under test: isolated worktree at iOS journey
Reporting claim, observed on the private store (real product refusals only; nothing injected)
NOT RUN here: B1 code-less TIMEOUT Side note (not a verdict input): interrupting the managed Repro steps (iOS)
Expected: iOS journey passes end to end; refusals aggregate per code+platform. Cleanup: runner close → ios
ios-flow.mp4 |
Quiet-host Android observation at
|
| Event | Monotonic ms |
|---|---|
| flow-park committed | 1,678.843 |
| stage 0 execute-begin | 1,679.111 |
| payload-verify passed (1.1.24 pin-cache) | 2,308.552 |
| stage 0 execute-complete | 10,115.943 |
stage 0 relaunch-begin (stopApp:true) |
10,115.985 |
| stage 1 execute-begin | 130,178.946 |
| stage 1 cleanup-begin | 130,180.260 |
| tool-outcome FAIL | 130,942.023 |
No relaunch-complete. Gap after relaunch-begin: 120,062.961 ms. Native stage 0 returned; authored steps did not.
Load (1-minute)
| Phase | 1-min |
|---|---|
| Gate / boot | 4.52 |
| After pin | 18.32 |
| Action start | 12.79 |
| During relaunch interval | 12.17–19.48 |
| Timeout | 15.85 |
| Series max (around build) | 25.91 |
QA4's load >300 confound does not apply here. The same stall reproduced on a quiet host.
Cleanup
Owned emulator, private AVD, unused validator simulator, Metro, and package integration were removed. Foreign shared AVDs Pixel_9_Pro and Pixel_10a were observed present and unused. No USB phone. Action bytes unchanged.
iOS remains NOT RUN for this head. Product acceptance remains FAILED. No merge.
2 participants
Intent
Captain (2026-09-15): "#1034 should be fixed" and "don't worry about the tokens. Finish the 1034". Earlier: address the review feedback on that same PR.
Make #1034 landable on the same branch. Keep the reporting intent of #981 (systemic authority-refusal recognition and reporting). Do not rewrite the authority system. Android QA is still FAILED; iOS already passed at this pair.
A1. In authority-refusal.ts, drop axis from the systemic key so one code+platform is one group across emission shapes (gate-emitted meta.axis, plain Error, failResult with no meta, legacy prose). Axis stays a reported attribute; conflicting or missing axis becomes null on the group. Add the mixed-emission regression.
A2. Recorder writes authorityRefusal: null as an explicit negative verdict on new failure records; reporter already skips present-null and must not re-admit via prose. Persist that explicit null on matching historical-row upserts. Do not backfill every untouched historical row.
B1. Preserve timeout traces: recognize code-less cdp_run_action TIMEOUT (and the producer timedOut surface), add only flow-park and flow-stage diagnostic events around existing park/stage/relaunch/cleanup/origin boundaries. Do not change errors, cleanup, authority, or timeouts.
B2. Change workspace qa-observe-screen-match.yaml launchApp.stopApp true to false only if a new current-session trace proves the cold native launch is the causal failure. That fixture edit is separately routed; do not invent a product redesign if the trace disproves it.
What Changed
SESSION_AUTHORITY_REQUIRED,METRO_ORIGIN_MISMATCH,RUNNER_OWNERSHIP_MISMATCH,HANDOFF_NOT_AUTHORIZED,NON_GIT_MANIFEST_REQUIRED,BUNDLE_HANDSHAKE_UNAVAILABLE) from gate envelopes, plain errors, and legacy prose into dedicatedFF_*families, stores an explicitauthorityRefusalverdict (ornull) plus a code-and-platformsystemicKeyon each record, and no longer marks a refusal as recovered when a later diagnostic call succeeds.experience-trendsCLI gain asystemicRefusalssection that groups retained refusals per code and platform across tools, merges conflicting axis/cause tonull, labels recovery as not verified and current authority state as unknown, and documents that--sinceonly scopes new patterns.cdp_run_actionTIMEOUT andmaestro_runtimedOutfailures and emitflow-park/flow-stagelifecycle events around park, stage, relaunch, origin, and cleanup boundaries; stored symptoms and the feedback collector scripts additionally redact UUID-shaped identifiers (redaction rules bumped to v2), with the six refusal families added to the seed failure catalog and a patch changeset.Risk Assessment
✅ Low: Every required intent criterion (A1 axis-free key with null-on-conflict, A2 explicit null verdict on new and upserted rows without backfill, B1 timeout recognition plus observation-only diagnostics, B2 fixture untouched) is source-verified with behavioral tests, the store is explicitly non-load-bearing, the diagnostics inserts cannot throw or alter control flow, and no wrong-value path was found; the only findings are a removable constant-field trio and a changeset wording nit.
Testing
Built the core package, ran the seven targeted unit test files for the refusal decoder, recorder, systemic trends, trends CLI, relaunch, runner diagnostics, and instrumentation (all passed), then demonstrated the intent end to end by driving the real recorder and the compiled rn-experience-trends CLI: mixed emission shapes of one code+platform collapse into a single systemic group with null axis and the explicit-null verdict is not re-admitted via prose, and a code-less cdp_run_action TIMEOUT retains its flow-park/flow-stage trace bundle. The B2 fixture is untouched in the diff. Transient build outputs and node_modules were removed from the worktree.
Evidence: End-to-end transcript: stored rows, CLI text report, JSON systemic group, timeout bundle events
Evidence: rn-experience-trends --json report showing one METRO_ORIGIN_MISMATCH/android group across four emission shapes
Evidence: Retained runner-diagnostics bundle for a code-less cdp_run_action TIMEOUT with flow-park/flow-stage events
Evidence: Demo script used to generate the end-to-end evidence
Evidence: CLI text report excerpt (systemic section)
Evidence: Targeted unit test transcript (122 pass, 0 fail)
Pipeline
Updates from git push no-mistakes
✅ **intent** - passed
✅ No issues found.
✅ **Rebase** - passed
✅ No issues found.
packages/rn-dev-agent-core/src/experience/trends.ts:48- Simplification: SystemicRefusalTrend carries three literal-typed constant fields on every row (recoveryEvidence: 'not-verified', currentAuthorityState: 'unknown', scope: 'retained-local-history'). They can only ever hold one value, so they convey no per-row information; the intent's requirement that groups never claim recovery or a currently blocked session is already satisfied by the absence of any recovery/current-state field plus the README and CLI disclaimer lines. No intent requirement (A1/A2/B1/B2 or the bug: repeated authority refusals never surface as one systemic issue #981 reporting intent) needs these fields. Recommended remedy: remove the three fields from the interface, the group constructor, and any test fixtures that assert them. This changes the JSON report shape, so it is the author's call..changeset/systemic-authority-refusals.md:6- The changeset body is two sentences; the repo owner's stated convention for changesets monorepos is a single-sentence summary because the text ships to the published CHANGELOG. Collapse to one sentence, e.g. join the UUID-redaction clause with 'and' or fold it into the first sentence.✅ **Test** - passed
✅ No issues found.
yarn install --immutableandyarn workspace rn-dev-agent-core build(tsc) so dist-importing tests could runnode --test --test-concurrency=2 test/unit/experience-authority-refusal.test.ts test/unit/experience-evidence.test.ts test/unit/experience-systemic-trends.test.ts test/unit/experience-trends-cli.test.ts test/unit/gh-708-mid-flow-relaunch.test.ts test/unit/runner-diagnostics.test.ts test/unit/instrumentation.test.ts(122 pass, 0 fail)End-to-end demonode e2e-systemic-refusal-demo.mjs(evidence dir): seeded a temp store via ExperienceRecorder with gate-emitted/plain-Error/no-meta/legacy-prose refusals plus an unrecognised-code prose row, randist/experience-trends.js --since 2026-06-01in text and--json, asserted one group, count 4, axis null, maestro_run excluded, provenance legacy-derived+recorded, no recovery or blocked-session claimEnd-to-end demo B1: ranrunFlowParkedunderwithRunnerDiagnosticsContext('cdp_run_action'), recorded a code-less TIMEOUT via ExperienceRecorder, asserted the retained bundle has failureCode TIMEOUT and flow-park begin/released/committed plus flow-stage execute-begin eventsVerified the diff does not modify qa-observe-screen-match.yaml (B2 conditional not triggered)Removed generated dist, .yarn/install-state.gz, and node_modules from the worktree after testing✅ **Document** - passed
✅ No issues found.
✅ **Lint** - passed
✅ No issues found.
✅ **Push** - passed
✅ No issues found.