Build step 6 of 6 — Phase 7 (umbrella #162)
Build doc: documentation/plans/phase-7-build-plan.md §3 Step 6
Type: operator (live multi-hour soak — not a /build-step code change)
What to build
Run the full /improve-bot-advised loop unattended to confirm the staleness signal → soak scheduling → daemon recovery path works end-to-end against the live system. This was split out of the old Step 3 (which buried a 1h live soak inside a code step's Done-when, per plan-review §22).
Done when (operator runs + reports)
- With reward rules freshly edited, run
/improve-bot-advised --self-improve-code --hours 8.
- Staleness detected on iteration 1 (reward edit newer than newest checkpoint).
- A
soak improvement is emitted and runs to completion (2–4h, clamped ≤ 50% of budget).
- Post-soak deterministic eval shows a measurable Elo/WR delta (sign doesn't matter — proves the path works end-to-end).
- Loop resumes and finishes remaining budget without starving other iterations.
Operator commands
# edit a reward rule first so the policy is provably stale, then:
uv run python -m orchestrator.staleness <current-version> # sanity: is_stale should be true
uv run improve-bot-advised --self-improve-code --hours 8
What to look for
| Check |
Expected |
| Iteration 1 staleness |
is_stale: true, reason cites reward-edit-after-checkpoint |
| Soak emitted + clamped |
One type: soak, duration ≤ min(remaining/2, 4h) |
| Daemon recovery |
API-only backend restarts cleanly after the soak; data/ intact |
| Budget |
Soak-type time ≤ 50% of total --hours |
| End-to-end |
Post-soak eval delta non-zero; loop completes remaining budget |
Depends on
Steps 1–5 (#180–#184) all shipped + green.
Parallel-safe with
none — strictly sequential (this is the phase-end validation gate; it observes the behavior Steps 1–5 build).
Produces
A run-log / observation report (operator-authored). No code artifact.
Operator workflow notes
- This is a
Type: operator step: /build-phase halts here and hands off to the operator; it does not spawn /build-step.
- It is the Phase 7 validation gate (plan §5/§6).
Synced from documentation/plans/phase-7-build-plan.md by /repo-sync at acd147f
Build step 6 of 6 — Phase 7 (umbrella #162)
Build doc: documentation/plans/phase-7-build-plan.md §3 Step 6
Type: operator (live multi-hour soak — not a /build-step code change)
What to build
Run the full
/improve-bot-advisedloop unattended to confirm the staleness signal → soak scheduling → daemon recovery path works end-to-end against the live system. This was split out of the old Step 3 (which buried a 1h live soak inside a code step's Done-when, per plan-review §22).Done when (operator runs + reports)
/improve-bot-advised --self-improve-code --hours 8.soakimprovement is emitted and runs to completion (2–4h, clamped ≤ 50% of budget).Operator commands
What to look for
Depends on
Steps 1–5 (#180–#184) all shipped + green.
Parallel-safe with
none — strictly sequential (this is the phase-end validation gate; it observes the behavior Steps 1–5 build).
Produces
A run-log / observation report (operator-authored). No code artifact.
Operator workflow notes
Type: operatorstep:/build-phasehalts here and hands off to the operator; it does not spawn/build-step.Synced from documentation/plans/phase-7-build-plan.md by /repo-sync at acd147f