Plan-review prompt: convergence stop rule, finding-origin telemetry, and fix-verification guards #10

Merged

vnprc merged 2 commits from agent/plan-review-convergence into master

2026-07-04 15:29:53 +01:00

allod-agent commented

2026-07-04 14:10:59 +01:00

Member

Implements the plan-review process changes from allod/memory#9. The 15-pass allod/nexus-host-key-rotation review that motivated the issue never converged, increasingly spent passes fixing contradictions earlier passes had introduced, and ran a three-pass self-verified regression chain on one section — all failure modes that trace back to the review-prompt template and the Agent Rotation guidance.

Changes

templates/plan-review-prompt.md

The findings summary must now classify each finding's origin (original-plan defect vs introduced by an earlier pass, naming the commit) and count only findings new to the pass; carried-over unresolved items move to Focus Areas so per-pass counts track a real severity trend instead of inflating it.
The zero-BLOCKER/zero-GAP/zero-QUESTION stop rule is replaced by a convergence heuristic: stop plan-text review when review-introduced findings outnumber original-plan findings for two consecutive passes, or after two consecutive passes with no BLOCKER and no original-plan GAP, then hand remaining focus areas to implementation review.
After any pass that commits a structural or design change, the next pass is a scoped diff review of that change by a different model than the author, not a full re-review.
Each pass must run an explicit SIMPLIFY sweep and report what it considered for deletion; repeated zero-SIMPLIFY passes on a growing plan are flagged as a smell rather than a clean bill.

dev-plans.md

Agent Rotation now tracks per-model fix stability in the review prompt's focus-area updates, prefers the most stable model for verification passes, and drops a model from the rotation after repeated same-feature regressions.

Risk

R1. Documentation and review-process guidance only — no runtime, provisioning, secrets, or generated-artifact behavior changes, and rollback is a straight revert. The residual question for human triage is whether the convergence thresholds ("two consecutive passes") terminate too early on a genuinely churny plan; they are single-line tunables if a real review shows them to be too eager.

Validation

No build or test harness exists in this docs repo. Verified by mapping each of the six issue items to a concrete edit, and by checking cross-file consistency: the template's origin classification feeds the convergence heuristic, and the template's scoped-diff verification pass ties to the fix-stability tracking added under Agent Rotation in dev-plans.md. Grep confirms no remaining references to the removed zero/zero/zero stop rule.

Closes allod/memory#9

Implements the plan-review process changes from allod/memory#9. The 15-pass `allod/nexus-host-key-rotation` review that motivated the issue never converged, increasingly spent passes fixing contradictions earlier passes had introduced, and ran a three-pass self-verified regression chain on one section — all failure modes that trace back to the review-prompt template and the Agent Rotation guidance. ## Changes `templates/plan-review-prompt.md` - The findings summary must now classify each finding's origin (original-plan defect vs introduced by an earlier pass, naming the commit) and count only findings new to the pass; carried-over unresolved items move to Focus Areas so per-pass counts track a real severity trend instead of inflating it. - The zero-BLOCKER/zero-GAP/zero-QUESTION stop rule is replaced by a convergence heuristic: stop plan-text review when review-introduced findings outnumber original-plan findings for two consecutive passes, or after two consecutive passes with no BLOCKER and no original-plan GAP, then hand remaining focus areas to implementation review. - After any pass that commits a structural or design change, the next pass is a scoped diff review of that change by a different model than the author, not a full re-review. - Each pass must run an explicit SIMPLIFY sweep and report what it considered for deletion; repeated zero-SIMPLIFY passes on a growing plan are flagged as a smell rather than a clean bill. `dev-plans.md` - Agent Rotation now tracks per-model fix stability in the review prompt's focus-area updates, prefers the most stable model for verification passes, and drops a model from the rotation after repeated same-feature regressions. ## Risk R1. Documentation and review-process guidance only — no runtime, provisioning, secrets, or generated-artifact behavior changes, and rollback is a straight revert. The residual question for human triage is whether the convergence thresholds ("two consecutive passes") terminate too early on a genuinely churny plan; they are single-line tunables if a real review shows them to be too eager. ## Validation No build or test harness exists in this docs repo. Verified by mapping each of the six issue items to a concrete edit, and by checking cross-file consistency: the template's origin classification feeds the convergence heuristic, and the template's scoped-diff verification pass ties to the fix-stability tracking added under Agent Rotation in `dev-plans.md`. Grep confirms no remaining references to the removed zero/zero/zero stop rule. Closes allod/memory#9

allod-agent added 1 commit

2026-07-04 14:11:00 +01:00

Plan-review prompt: convergence stop rule, finding-origin telemetry, fix-verification guards 4f35d4f842

Rework templates/plan-review-prompt.md and the dev-plans.md Agent Rotation
guidance per allod/memory#9: classify each finding's origin, count only
findings new to the pass, replace the zero/zero/zero stop rule with a
convergence heuristic, add a scoped-diff verification pass by a different
model after structural fixes, require an explicit per-pass SIMPLIFY sweep,
and track per-model fix stability to steer the rotation.

allod-agent commented

2026-07-04 14:36:52 +01:00

Author

Member

Follow-up requested: the review-pass model-selection guidance has no actor while passes are launched manually — the finishing agent cannot pick or start the next pass's model. Tightening it so the handoff is explicit: the finishing pass writes a Next pass: line into the prompt's Focus Areas (scoped-diff vs full re-review, the commit(s) to scope to, and a model recommendation — not the fix's author, preferably the most fix-stable on record), and dev-plans.md states plainly that these are recommendations for whoever launches the next pass, since the agent cannot select the model itself. Applying as a follow-up commit.

Follow-up requested: the review-pass model-selection guidance has no actor while passes are launched manually — the finishing agent cannot pick or start the next pass's model. Tightening it so the handoff is explicit: the finishing pass writes a `Next pass:` line into the prompt's Focus Areas (scoped-diff vs full re-review, the commit(s) to scope to, and a model recommendation — not the fix's author, preferably the most fix-stable on record), and dev-plans.md states plainly that these are recommendations for whoever launches the next pass, since the agent cannot select the model itself. Applying as a follow-up commit.

allod-agent added 1 commit

2026-07-04 14:37:26 +01:00

Plan-review prompt: make next-pass model handoff explicit bab868f5e8

Review passes are launched manually, so the finishing agent cannot pick or
start the next pass's model. Spell out the handoff instead: the finishing
pass writes a Next pass: line into Focus Areas (scoped-diff vs full, the
commit(s) to scope to, and a model recommendation other than the fix's
author, preferably the most fix-stable on record), and dev-plans.md states
these are recommendations for whoever starts the next pass, not something
the agent selects itself.