Metadata
| Status | done |
|---|---|
| Assigned | agent-928 |
| Agent identity | 6841f97a7a8e6cbc213051ce0ffb3c85c92c80cb61fd699428ab3fed6cb6ba94 |
| Created | 2026-07-26T10:25:10.266058219+00:00 |
| Started | 2026-07-28T12:16:01.887291440+00:00 |
| Completed | 2026-07-28T13:23:41.384381534+00:00 |
| Tokens | 26936152 in / 47278 out |
Description
Objective
Integrate the authoritative lifecycle, admission deferral, lazy evaluation creation, dedicated Pi bounded-evaluation lane, and deep-readonly FLIP; then roll evaluation back on through controlled canaries. Preserve native Codex execution behavior and executor neutrality without changing direct Codex/Claude task scope.
Rollout (enforced order)
- Start with automatic bounded evaluation, FLIP, and all hard gates disabled. Persist an explicit rollout stage; config/reload must reject attempts to skip forward.
- Run the credential-free Fake-Pi full lifecycle suite and real daemon/TUI human-flow smoke.
- Run a live low-risk Pi/Luna bounded advisory-evaluation canary when the configured Pi route is available. Record success evidence before stage advancement.
- Run one explicitly requested deep-readonly system FLIP canary. Verify its observation-only boundary and genuine latent-intent/counterfactual findings; do not call a shallow bounded grader FLIP.
- Verify zero evaluation for sources that never ran, zero stuck pending-eval states, zero duplicate records/verdicts, zero worker/build slot use, correct admission-deferral behavior, and idempotent daemon restarts.
- Only after steps 2–5 pass, enable
auto_evaluate=truein advisory bounded mode witheval_gate_all=false. Leave eager/global FLIP disabled and use explicit/selective high-risk policy. - Observe multiple real source completions and record rollback thresholds. This task may never enable a global hard gate. A later operator may enable a hard gate only for an explicit high-risk policy after reviewing canary evidence.
Produce an operator runbook, machine-readable canary evidence, and before/after Viz evidence.
Validation
-
A RED integration scenario is authored first and fails before wiring:
pi_evaluation_rollout_requires_canary_success_before_enablecovers disabled start, stage-skip rejection, canary evidence, advisory enablement, and rollback. - Credential-free Fake-Pi fixtures run the entire publish → execute → lazy bounded evaluation → advisory verdict → terminal source flow; cancellation, skip, admission deferral, and non-spawn sources create no evaluation work.
- A live scripted human flow uses the real daemon plus terminal/TUI/Viz controls, not library-only calls, and shows no eager clutter plus legible on-demand evidence.
- Daemon restart during each stage is idempotent; there are no duplicate records/verdicts, stuck pending-eval states, or worker/build/worktree use by evaluation.
- The bounded Pi evaluation canary passes before the explicit deep-readonly FLIP canary; deep FLIP's read-only boundary and cross-system understanding are demonstrated separately.
- Native Codex task routing/config remains selectable and is not silently rewritten or used as Pi fallback; the credential-free regression need not invoke an external Codex login.
-
Advisory
auto_evaluate=truecan be enabled only after recorded canary success;eval_gate_allremains false, global FLIP remains disabled, and attempts to globally hard-gate in this rollout are rejected. - Rollback command/path is exercised through the actual operator terminal flow and documented.
-
Permanent smoke scenario(s) are added under
tests/smoke/scenarios/and registered with this task intests/smoke/manifest.tomlowners. -
cargo fmt --check,cargo clippy, relevant suites,cargo build, andcargo testpass with no regressions.
Depends on
Required by
Log
- 2026-07-26T10:25:10.251947711+00:00 Task paused
- 2026-07-26T10:25:30.969038445+00:00 Task published
- 2026-07-28T12:14:35.930837504+00:00 Spawn preparation deferred before launch permit; rollback is complete and no circuit-breaker charge was recorded. Repair the reported checkout/configuration condition and retry: spawn transaction for agent-928 rolled back (task remains dispatchable; rollback diagnostics: complete): failed to establish isolated-worktree observer baseline: observer source tuple mismatch
- 2026-07-28T12:16:02.283531307+00:00 Spawned by coordinator --executor pi --model openai-codex:gpt-5.6-sol --isolation required-worktree
- 2026-07-28T12:16:18.890033259+00:00 Starting integration: inspect dependency commits/WIP, then author RED rollout scenario before wiring staged canaries.
- 2026-07-28T12:20:49.099432280+00:00 RED first confirmed: permanent pi_evaluation_rollout_requires_canary_success_before_enable scenario fails on missing ; now wiring persisted ordered controller and guards.
- 2026-07-28T13:16:46.406155532+00:00 Live canary passed on pi:openai-codex:gpt-5.6-luna: bounded verdict consumed (score 0.70), explicit deep FLIP first failed closed on locator grammar then succeeded on bounded same-route retry with latent intent + two counterfactual probes + cross-component findings; two subsequent real advisory completions each produced one record/verdict. Machine evidence and Viz digests recorded.
- 2026-07-28T13:17:40.849215854+00:00 Validation: fmt/check, clippy, cargo build --locked, focused lazy/dedicated/deep suites, the RED-first staged rollout smoke, and all four owned daemon/TUI/Fake-Pi/admission smokes pass. Full cargo test completed with unrelated branch-preexisting failures (profile env poison, fail/watchdog legacy tests, spawn-isolation drift, outdated executor tests); no evaluation rollout/lazy/dedicated/deep failure.
- 2026-07-28T13:17:40.924806581+00:00 Committed: 93df3f16 — pushed to remote
- 2026-07-28T13:23:41.384381823+00:00 Task marked as done