An untrusted Codex project can silently lose its project-local MCP configuration, even in Full Access mode. Sending a catalog-only prompt in a project with a local MCP server returns no tool and gives no actionable explanation.
Read Codex's configuration layers for the current turn directory and display its disabled-project reason through the existing persisted system notices. The optional diagnostic falls back after a read failure or three seconds; configuration bodies are omitted from native protocol logs.
Fixes #15140
Triaged bug via the automated issue flow. The application displays Codex's explanation without changing project trust.
CodexAdapterV2.test.ts: 79 passed after merging main. This includes the disabled-project notice, enabled/non-project controls, timeout continuation, and configuration-log redaction.tsc-rs --noEmit: exit 0. Formatting hooks and git diff --check passed.Evidence index · Focused verification · Failing regression output
Independent proof audit reproduced the narrow dark capture with zero changed pixels and verified all six image-pair difference maps and counts. Strict matrix coverage remains incomplete for N3 (multiple disabled project reasons); that state is unverified.
Local web at http://localhost:8103. Native T3 browser, Chromium154 Linux x86_64 desktop UA, DPR2, root font16px, visualViewport.scale1. CSS viewports1440x900 and390x900. Explicit app Light/Dark themes. Whole viewport captured, chat scrolltop0, composer collapsed for narrow stills. Wide sidebar256CSS light,208CSS dark, matched per pair.
Native screenshot export cap1280px is implemented in apps/server/src/preview/ServerBrowserPage.ts (MAX_SCREENSHOT_WIDTH and CDP clip scale). Consequently1440x900CSS/DPR2 stills export1280x800PNG;390x900CSS/DPR2 exports780x1800PNG. Videos retain2880x1800 wide /780x1800 narrow. No images rescaled after capture. This is desktop browser width emulation, not native mobile.
Server, Vite and browser Date fixed2026-10-22T08:05:00.000Z; timers real. Both code versions restored identical pre-turn T3/provider snapshots, same project and thread IDs, exact prompt and local draft. Real provider latency differs; recordings are behavioral comparisons, not frame-aligned pixel diffs. Existing sidebar data copied into isolated home. No live home writes.
Route: /8a17860f-3c27-4a5f-8044-93154022e65d/3e6b8790-416b-4fe8-894c-cb7509147d32 . Title normalized through UI to Project MCP trust. Collapsed warning/answer positions are the intended visible change.
Diff0 is amplified binary zero-tolerance proof: white if max absolute RGB channel difference >0; black otherwise. DiffT uses >16/255, also binary. Counts refer to exported PNG pixels. No blanket claim of zero differences outside the change.
| Row | Before BASE 3b95e7831b | After HEAD 91197cb734 | Diff0 (BASE→HEAD) | DiffT16 (BASE→HEAD) |
|---|---|---|---|---|
| N1 1440 light | before1440light.png | after1440light.png | 1440light-diff0.png 4264 changed | 1440light-diffT16.png 3861 changed |
| N1 1440 dark | before1440dark.png | after1440dark.png | 1440dark-diff0.png 4272 changed | 1440dark-diffT16.png 3875 changed |
| N1 390 light | before390light.png | after390light.png | 390light-diff0.png 6209 changed | 390light-diffT16.png 5763 changed |
| N1 390 dark | before390dark.png | after390dark.png | 390dark-diff0.png 6170 changed | 390dark-diffT16.png 5776 changed |
| N0 trusted control | beforeControl.png | afterControl.png | Control-diff0.png 3215 changed | Control-diffT16.png 2758 changed |
| N2 390 dark expanded | before390dark.png | afterExpanded.png | Expanded-diff0.png 24630 changed | Expanded-diffT16.png 22348 changed |
What else changed: N1 wide light/dark and narrow dark show the actual warning and displaced No answer/action timestamp. Narrow light additionally has a tiny sidebar-toggle icon raster difference. N2 adds expanded detail and larger answer displacement. Trusted control has Yes in both and no warning; its auto-generated title timing differs (BASE raw prompt title, HEAD Check MCP Tool Availability), plus small submit-button readiness/focus styling. The control is not pixel-identical. Sidebar ages are frozen; no additional age change was identified in the core pairs. Diff masks show all actual differences, without masking or normalizing pixels.
No audio track was added. The following written narration accompanies each recording. Times are elapsed media seconds; ranges below are bounded using sampled frames, rather than claiming precise provider event latency.
| Row | Before BASE 3b95e7831b | After HEAD 91197cb734 | Narration / observed endpoint |
|---|---|---|---|
| T1 arrival | beforeT1.mp4 25.4032s; start08:18:48.061Z | afterT1.mp4 23.283367s; start08:26:02.439Z | 0s prepared exact prompt; by5s sent. Decoded held-frame PTS: BASE14.989133s still Thinking,16.398167s No (completed by20s), held to end. HEAD4.908533s Thinking without warning;9.517100s warning visible (by10s);14.442333s No, one warning held to end. Provider timing differs. These bounds supersede the earlier incorrect BASE-by15s and HEAD-warning-by5s narration. |
| T2/T5 detail/reload | beforeT25.mp4 5.969367s; start08:22:16.978Z | afterT25-final.mp4 64.956467s; start08:39:26.003Z | BASE has no warning to expand; reload preserves absence/No. HEAD exact verified timeline below. |
| T3 trust next turns | beforeT3.mp4 40.478933s; start08:22:58.287Z | afterT3.mp4 35.3048s; start08:29:04.111Z | Both begin from untrusted first-turn history. Trusted helper transition occurs outside public browser view at start; next same-thread catalog answer remains No (native session catalog caching), with no NEW warning. BASE trusted answer complete by10s; reverse untrusted transition at |
T2/T5 final HEAD media narration uses decoded frame PTS (sparse variable-frame-rate video, ~1fps average). Seeking with ffmpeg -ss can return the NEXT frame, so requested seek time is not a measured frame timestamp. Auditor decoded every frame to establish these bounds:
| Media time / bound | Narrated observation |
|---|---|
| 0s | Starts expanded from preceding inspection. |
| By11s, still collapsed at19s | Visible row header clicked to establish collapsed state; tool independently verified aria-expanded=false. |
| By22s | Expanded full reason, readable project and provider-home trust instruction. |
| By32.974700s | Collapsed again; verified aria-expanded=false. |
| 32.974700–40.275133s | Reload transition is bounded by the last collapsed frame and first blank-loading frame. |
| 40.275133s /40.359s | Blank loading, then splash. |
| 45.684267s /48.725667s | Splash, then partial page. |
| 49.947667s | Restored page, persisted notice collapsed. |
| ~50.62s; settled51.02s | Explicit postreload expansion; aria-expanded=true, height150.48CSS verified. |
| 51.02–64.92s | Complete recovery reason held visible through final decoded frame. Video duration64.956467s. |
Separate wall-clock action log (UTC, not inferred media PTS): recording start08:39:26.003; header collapse08:39:35.975; expand08:39:46.478; collapse08:39:58.515; reload08:40:05.858; postreload expand08:40:16.613; stop08:40:30.912. Tool call times and sparse media frames have different precision. Neither is relabelled as the other.
The earlier afterT25.mp4 and afterT25-corrected.mp4 are superseded; their collapse/final expansion sequence was insufficient. Clicking the inspector center can hit detail instead of the header; final recording uses the actual visible header and explicit state checks.
Safe trust transitions changed only disposable config.toml. No credentials or editor/terminal config content appears in recordings. Trust was restored UNTRUSTED; root then restored first-turn checkpoint for independent audit. No provider turns were issued after that restore.
Captured required: N1 all4 width/theme pairs; N0 trusted pair; N2 expanded pair; T1 arrival pair; T2/T5 detail/reload pair; T3 trust-transition pair with explicit endpoint limitation. Missing required conditional row: N3 stacked notices UNCOVERED. Real fixture exposes one disabled project layer; no real nested layer or labelled protocol harness capture was produced. Source review is not substituted for this row. Additional unverified: N4 whitespace harness, deterministic T4 failures/races, T6 goal continuation, remote/relay/tunnel runtime, native Windows desktop, native iOS/Android expansion/copy. Narrow web is not native proof. T3 final untrusted turn completion is not captured. It is not claimed by the stills or descriptions. No duplicate provider tests/checks run by capture worker. Parent separately reported79 adapter tests, targeted lint and server typecheck passing.
Unverified: Windows desktop (cannot-verify desktop) and native mobile runtime, plus the fixture limits above.
Model: GPT-6.1-Sol. Harness: Codex in T3 Code.
Fixes pingdotgg/t3code#15140
Added 7 commits
Verdict: Approved at 4b2c6cf
Macroscope's review found this PR approvable — Adds notices explaining why Codex skipped project configuration, without changing trust or configuration behavior. The diagnostic is time-bounded and tolerates unsupported servers, while config bodies are redacted from protocol logs.
You can add or adjust custom eligibility rules. Learn more.
✅ Thread transfer remains within every enforced ceiling.
| Provider | Metric | Main baseline | This PR | Impact | PR ceiling | |
|---|---|---|---|---|---|---|
| Codex | Total thread wire | 5.0 KiB | 5.0 KiB | 0 B (0.0%) | 6.8 KiB | ✅ |
| Codex | Thread snapshot wire | 3.8 KiB | 3.8 KiB | 0 B (0.0%) | 4.9 KiB | ✅ |
| Codex | Live turn WebSocket wire | 1.2 KiB | 1.2 KiB | 0 B (0.0%) | 2.0 KiB | ✅ |
| Codex | Live turn WebSocket decoded | 21.5 KiB | 21.5 KiB | 0 B (0.0%) | 24.4 KiB | ✅ |
| Codex | Live turn messages | 1 | 1 | 0 (0.0%) | 8 | ✅ |
| Claude | Total thread wire | 5.0 KiB | 5.0 KiB | 0 B (0.0%) | 6.8 KiB | ✅ |
| Claude | Thread snapshot wire | 3.8 KiB | 3.8 KiB | 0 B (0.0%) | 4.9 KiB | ✅ |
| Claude | Live turn WebSocket wire | 1.2 KiB | 1.2 KiB | 0 B (0.0%) | 2.0 KiB | ✅ |
| Claude | Live turn WebSocket decoded | 21.9 KiB | 21.9 KiB | 0 B (0.0%) | 24.4 KiB | ✅ |
| Claude | Live turn messages | 1 | 1 | 0 (0.0%) | 8 | ✅ |
Baseline: 3a8c250 · PR result: 4b2c6cf · Source CI: success
10 historical turns, 5 command tools per turn, 878.9 KiB retained MCP result per historical turn, and a 1.05 MiB retained result in the measured turn.
Updated in place by a trusted workflow. PR artifacts are strictly validated and never executed.
No actionable comments were generated in the recent review.
The adapter now reads Codex configuration layers before starting a turn and emits system notices for disabled project layers. Protocol logs redact values under the config key. Tests cover configuration-read errors and timeouts, notice emission, and log redaction.
Changes: Project Configuration Notices
| Layer / File(s) | Summary |
|---|---|
Read project configuration apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.test.ts, apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.testkit.ts | Before a native turn starts, the adapter reads configuration layers for the working directory with a three-second timeout and retains disabled project-layer reasons. Replay support handles configured responses and method-not-found errors. Tests cover a read that does not complete. |
Emit project-layer notices apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.test.ts | Root-turn registration emits completed system notice items for retained project-layer reasons. A test checks the notice and assistant message. |
Protocol Log Redaction
| Layer / File(s) | Summary |
|---|---|
Redact configuration values in protocol logs apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.ts, apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.test.ts | Protocol-value redaction replaces values under the config key. A test checks that top-level and project-layer database URLs do not appear in logged payloads. |
Estimated code review effort: 3 (Moderate) | ~20 minutes
Change: Bug fix · Severity of issue fixed: Medium
| 1 | sequenceDiagram |
| 2 | participant CodexAppServer |
| 3 | participant CodexAdapterV2 |
| 4 | participant RootTurnRegistration |
| 5 | CodexAppServer->>CodexAdapterV2: Return config/read layers |
| 6 | CodexAdapterV2->>CodexAdapterV2: Retain disabled project-layer reasons |
| 7 | CodexAdapterV2->>RootTurnRegistration: Provide retained reasons |
| 8 | RootTurnRegistration->>RootTurnRegistration: Emit completed system_notice items |
Suggested reviewers: t3dotgg
4b2c6No identified issue blocks merging the project-configuration notices and log redaction after normal checks.
| Check name | Status | Explanation |
|---|---|---|
| Title check | Passed | The title is concise, uses conventional commit format, and clearly identifies the main change: explaining skipped Codex project configuration. |
| Description check | Passed | The description includes all required sections. It explains the problem, implementation, scope and issue link, and detailed verification results. It also identifies unverified scenarios and limitation… |
| Linked Issues check | Passed | For [#15140], startTurn requests config/read with the current turn cwd and includeLayers: true. It collects disabledReason from disabled project layers and emits each reason as a persisted `… |
| Out of Scope Changes check | Passed | The whole-PR changes remain connected to [#15140]. Adapter changes add the configuration diagnostic, timeout handling, persisted notice emission, and native protocol redaction. Test and replay-harness… |
Explanation
The description includes all required sections. It explains the problem, implementation, scope and issue link, and detailed verification results. It also identifies unverified scenarios and limitations.
Explanation
For [#15140], startTurn requests config/read with the current turn cwd and includeLayers: true. It collects disabledReason from disabled project layers and emits each reason as a persisted system_notice. The diagnostic uses a three-second timeout and an empty fallback for request failure, so the turn can continue. Protocol logging redacts complete config values. Tests cover the project notice, control cases, timeout continuation, and configuration redaction. The proposed trust action and scoped refresh are not required because [#15140] identifies the missing diagnostic as the reported failure.
Explanation
The whole-PR changes remain connected to [#15140]. Adapter changes add the configuration diagnostic, timeout handling, persisted notice emission, and native protocol redaction. Test and replay-harness changes validate those behaviors and support older app-server responses. The changes do not modify project trust or add unrelated product behavior.
Note
This review was completed with usage-based billing: files reviewed beyond your plan's included limits are billed at $0.25/file. View usage-based billing.
50fe1243-4a73-4a44-9164-044aba7a9861Reviewing files that changed from the base of the PR and between 91197cb734f1b7cfc1ec440591eed38ff6e56dff and 4b2c6cf9ef2e202c2f850267ba057cb5e5d19696.
apps/server/src/orchestration-v2/Adapters/CodexAdapterV2.test.tsapps/server/src/orchestration-v2/Adapters/CodexAdapterV2.testkit.tsapps/server/src/orchestration-v2/Adapters/CodexAdapterV2.tsLimit details: You’ve used all 10 included reviews currently available.
Added 2 commits
Verified with real Codex turns in the web client, Full Access, Chromium 1280×800, light theme, and fresh isolated databases on both commits.
Both agents reported the tool absent. Base provided no trust explanation; head displayed Codex’s instruction naming the affected project and provider configuration file.
| Before | After |
|---|---|
| 3a8c250cb0 | 4b2c6cf9ef |
| Before: tool absent with no project-trust notice | After: expanded persisted recovery instruction with both paths |
| Before recording | After recording |
Synthetic seed · Reproduction driver and browser steps
AI-assisted; verified with the tests and tools listed above.
Not shown: