Lynx/pi-config
Watch
1
0
Fork
You've already forked pi-config
 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat/doubt-review-hardening #1

Merged
Lynx merged 6 commits from feat/doubt-review-hardening into main 2026-10-01 12:12:24 +02:00
Conversation 0 Commits 6 Files changed 11 +542 -171
Lynx commented 2026-10-01 12:12:20 +02:00
Owner
Copy link
No description provided. 
- Waivers could never pass: the finding scan read every F-N row in the
  document, so Waivers rows were re-parsed as findings with an "invalid
  classification". The scan is now limited to the Findings section.
- Must-fix closure matched /resolved|fixed/ anywhere, so "unresolved",
  "not fixed" and "open (to be fixed)" counted as closed. Closure must now
  be exactly resolved or fixed.
- Scenario coverage used three independent substring checks, so citing
  "Invalid password lockout" also covered "Invalid password". References
  must now be contiguous and are not satisfied by a longer known name.
- quality-ui plans could omit the UI approval gate line (6 of 7 accepted).
- Waiver approval accepted any non-"no/none/pending" word (e.g. "TBD");
  it must now quote the user's words, matching the template's
  "Exact user approval".

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planned-By: Claude Opus 5.5 <noreply@anthropic.com>
A fallback means the orchestrator performed a reviewer's charter itself.
Previously all three reviewers could fall back with no recorded reason
and the gate still passed — the plan author reviewing its own plan.

check-plan now requires each fallback to state its failure in Verdicts
Notes, and fails when two or more reviewers fell back unless Waivers has
a `fallback` row quoting the user's approval of the reduced panel. The
schema retry policy and template comment describe the same rule.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planned-By: Claude Opus 5.5 <noreply@anthropic.com>
Adopted from the OpenSpec reflection harness (qtalen/agentic-ai-playground
15_OpenSpec_Reflection_Harness): separate substantive defects from noise,
and stop review loops at a fixed round count with a human decision.

- Reviewers share one operational must-fix test (divergent builds,
  verification that fails or proves nothing, contradiction, unsafe or
  undecided) and must name the condition; everything else is an
  improvement. A "do not report" list removes cosmetic findings and
  re-litigation of recorded user decisions, which inflated must-fixes
  and revision churn.
- Each reviewer report now states cold-executor readiness, which the
  gate already required but no reviewer produced.
- Closure is limited to two rounds, then the user chooses: waive named
  findings, rework from design, or one more round. Previously
  "iterate once" left the next step undefined.
- check-plan flags open-ended task wording ("e.g.", "etc.", "such as")
  that leaves a cold executor to guess the rest of a list; etc/ paths
  are not affected.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planned-By: Claude Opus 5.5 <noreply@anthropic.com>
The reflection harness freezes artifacts by instruction only; here the
freeze becomes a check. Before this, tasks.md or any other reviewed
artifact could be edited after the doubt review and check-plan still
passed, so apply could run a plan nobody reviewed.

- check-plan --fingerprint prints a short sha256 over proposal, specs,
  design, tasks, verification and (quality-ui) every ui-mocks file.
  Task checkboxes are normalized so apply can tick them.
- doubt-review.md records it as **Reviewed fingerprint** after closure;
  check-plan fails on a missing or mismatched value.
- Schemas: record the fingerprint after closure; apply re-runs
  check-plan when resuming and treats any other artifact edit as needing
  closure confirmation or the user's approval to re-record.

It cannot stop an agent from deliberately re-recording, but every edit
after review now fails the gate and shows up in the doubt-review.md diff.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planned-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planned-By: Claude Opus 5.5 <noreply@anthropic.com>
List the files the tree was missing (subagent model router, its config,
enable-quality, tmux, tests) and fix ask-user-question's path. Drop the
Requirements table since bootstrap covers it, add pi-lsp-extension to the
manual install, and document OPENROUTER_API_KEY and PI_BROWSER_PROFILE.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Planned-By: Claude Opus 5.5 <noreply@anthropic.com>
Lynx merged commit 70abf5c9c2 into main 2026-10-01 12:12:24 +02:00
Sign in to join this conversation.
Reviewers
No reviewers
Labels
Clear labels
No items
No labels
Milestone
Clear milestone
No items
No milestone
Projects
Clear projects
No items
No project
Assignees
Clear assignees
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
Lynx/pi-config!1
No description provided.