Restore desktop notifications for safety-access stops

[ipo]

Outcome

Use the current notification abstraction to emit desktop alerts for cyber/biology safety-access stops still missed upstream, without duplicate notification on already-covered terminal paths.

Governing evidence

Closed issue #17, commit bad1711718, and upstream 0.153.4 terminal-stop notification behavior.

The issue and this evidence define required behavior. Port functional intent onto the current upstream architecture; patch identity is irrelevant.

Non-goals

Do not change safety decisions, require a live desktop in automated tests, or add notification configuration.

Acceptance criteria and verification

Event-level tests prove exactly one notification for each missing safety-stop branch and no duplicate on upstream-covered paths. Include snapshots only if visible TUI text changes. Run the smallest focused just test -p <crate> checks that cover the changed crates under the repository reduced-priority wrapper. Regenerate schemas or snapshots only when this ticket changes them. For a large Rust change run scoped just fix -p <crate>, then just fmt; do not rerun tests after fix/format. Remove disposable build artifacts afterward.

Likely areas

TUI safety-stop event handling and notification abstraction tests.

Dependencies

Fully blocked by #81.

Do not begin ticket-owned mutation until every dependency is integrated; their landed contracts are authoritative.

Blocks #104.

Delivery

The accepted result must integrate as exactly one issue-linked commit with subject feat(tui): notify on safety-access stops. Tests, snapshots, generated files, and fixes belong in that commit; adaptation or test-only follow-up commits are not permitted.

Agent routing

agent_level: 2

Agent signature

gpt-5.6-sol@high:01a07441-98af-75c1-9950-8af46deb76de

Compile concurrency

Codex compilation is permitted and required where acceptance needs it. Every build, Rust test, Clippy/fix, or other compiling command must acquire the ticket's coordinator-assigned compile-slot lock. No more than three such commands may run concurrently across the coordinated ticket set.

Audit order

Freeze the implementation before compilation and run the normal fresh independent auditor step first. Only a candidate with an ACCEPT verdict may request a compile slot. A compile failure returns the ticket to repair and invalidates the affected audit evidence; re-audit the repair before another compile. Do not create a separate pre-compile review status.

[ipo]

Implementation started in ticket/98-safety-notifications from f1480a4b62.

[ipo]

IMPLEMENTATION_READY

Implemented and pushed commit c42a20ce1f (feat(tui): notify on safety-access stops).

Checks:

• just fmt passed.
• just test -p codex-tui reached crate compilation but is blocked by unrelated existing bedrock_catalog_tests.rs compile error: models_manager_without_cache now requires a second ResolvedModelCatalogOverlay argument.

Caveats: disposable build artifacts were produced under codex-rs/target; cleanup command was restricted by the environment.

Signature: gpt-5.6-sol@high:01a07441-98af-75c1-9950-8af46deb76de

[ipo]

IMPLEMENTATION_READY (resumed)

Reconciled the ticket branch with origin/patch-stack/rust-v0.153.4 at 3337c8ae350703aca002c516f8bf36918bca5ad9 via merge (no published history rewrite). Tip: dc58338b0c.

Checks:

• just test -p codex-tui safety_stops: 8 passed, 4092 skipped.
• just test -p codex-tui safety_alert_contract: 1 passed, 4099 skipped.
• cargo insta pending-snapshots: no pending snapshots.
• just fmt passed last.

Coverage proves one live cyber/biology safety-stop alert, suppression on replay, dedupe across paired error/completion events, and no alert for upstream-covered/unrelated paths.

Caveat: target artifacts were built in disposable /tmp/codex-tui-98-target and cleaned.

Signature: gpt-5.6-sol@high:01a07441-98af-75c1-9950-8af46deb76de

[ipo]

REJECT

Independent audit of dc58338 found a duplicate terminal path. In TurnStatus::Failed, deduplication only skips when last_non_retry_error was set by a prior Error event (protocol.rs:339-343). If the paired TurnCompleted(Failed) arrives first, safety_stop::on_safety_stop finalizes/adds the safety cell and emits SafetyAlert (safety_stop.rs:41-54), but does not record last_non_retry_error; a subsequent live non-retry Error for the same turn then runs the same handler again. This produces duplicate safety history and a second notify call, violating exactly-once behavior across paired error/completion paths. Existing tests cover only Error then Failed, not the reverse order. Focused tests were launched under the required reduced-priority wrapper with disposable target; compile was still in progress during this audit.

[ipo]

IMPLEMENTATION_READY (repair)

Fixed the auditor-reported reverse terminal-event ordering bug. Live safety stops now seed the dedupe tuple, and the Error handler symmetrically suppresses a paired duplicate; added Failed→Error regression coverage. Repair commit: 10ea7d981f (pushed, no history rewrite).

Checks:

• just fmt passed last.
• Focused just test -p codex-tui safety_stops is queued behind the mandated compile slot 2 lock with disposable target; it has not started because the shared queue contains long-running validations.
• Existing notification snapshot remains unchanged; prior snapshot check reported no pending snapshots.

Caveat: focused repair test completion is pending shared slot-2 queue availability.

Signature: gpt-5.6-sol@high:01a07441-98af-75c1-9950-8af46deb76de

[ipo]

IMPLEMENTATION_READY

Validation completed on commit 10ea7d981f: just test -p codex-tui safety_stops passed all 9 tests (4092 skipped). Disposable target cleaned up; worktree is clean.

[ipo]

ACCEPT

Re-audit of candidate 10ea7d9 confirms the prior reverse-order duplicate is fixed. Live safety stops now seed the same turn-id/message dedupe tuple used by both Error and TurnCompleted(Failed), so Error→Failed and Failed→Error each handle once. Distinct events remain distinguishable by turn ID or message. The new paired_failed_turn_and_live_error_emit_once regression test asserts one history cell, one pending SafetyAlert, and a stopped task. Updated IMPLEMENTATION_READY records just test -p codex-tui safety_stops passing all 9 tests with 4092 skipped and the disposable target cleaned. The cumulative diff preserves safety classification/decisions and replay suppression, adds no notification configuration or live-desktop dependency, and the notification snapshot matches the emitted display string. Non-compiling integrity checks: git diff --check passed and the worktree is clean.

[ipo]

Integrated as b4863513397b4fd603c8c14bfdbe3b66d95af94d on patch-stack/rust-v0.153.4. Fresh independent re-audit: ACCEPT. Validation: just test -p codex-tui safety_stops passed 9/9; integration remained patch-identical to accepted candidate and git diff --check passed.