werkrevisor 0.6.0: ORANGE for model findings, RED only for deterministic rules #2

2026-10-05T14:25:59+02:00

mi commented

2026-10-05 14:25:59 +02:00

werkrevisor 0.6.0: four verdict colours, and only deterministic rules block hard.

🟠 ORANGE: findings of the LLM and of the classifier Jev. No bot review; a developer approval after checking the findings unlocks the merge (assessment status warning, exit 15 where codes are distinguished, 0 in approval mode).
🔴 RED: only deterministic rules. First rule: open questions in a changed PR document ([ ], bullet or paragraph without checkbox under "Open Questions"/"Offene Fragen").
Fixes from the comparison on hsadmin-ng: no FIXME.review markers in findings, secret scan ignores src/test/** except real key formats, documentation files trigger no hard-YELLOW path rules.

Reason: the repository admin on dev.hostsharing.net is an operator, so an admin merge override is no way out of a false RED; model and classifier findings are often wrong (3 of 8 Sonnet findings in the comparison).

222 unit tests green in the pinned PR-Agent image.

Open: whether Gitea ignores a stale bot REQUEST_CHANGES after a later ORANGE/YELLOW head; tested on werkprobe after the rollout.

🤖 Generated with Claude Code

werkrevisor 0.6.0: four verdict colours, and only deterministic rules block hard. - 🟠 ORANGE: findings of the LLM and of the classifier Jev. No bot review; a developer approval after checking the findings unlocks the merge (assessment status `warning`, exit 15 where codes are distinguished, 0 in approval mode). - 🔴 RED: only deterministic rules. First rule: open questions in a changed PR document (`[ ]`, bullet or paragraph without checkbox under "Open Questions"/"Offene Fragen"). - Fixes from the comparison on hsadmin-ng: no `FIXME.review` markers in findings, secret scan ignores `src/test/**` except real key formats, documentation files trigger no hard-YELLOW path rules. Reason: the repository admin on dev.hostsharing.net is an operator, so an admin merge override is no way out of a false RED; model and classifier findings are often wrong (3 of 8 Sonnet findings in the comparison). 222 unit tests green in the pinned PR-Agent image. Open: whether Gitea ignores a stale bot REQUEST_CHANGES after a later ORANGE/YELLOW head; tested on werkprobe after the rollout. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

mi added 5 commits 2026-10-05 14:25:59 +02:00

New verdict ORANGE: model and classifier findings wait for a developer approval instead of blocking hard c0dcdf1b3b

RED now comes only from deterministic rules.
ORANGE (exit 15) sits between YELLOW and RED; in approval mode it requests a human review,
submits no bot review and sets the assessment status to warning.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Open questions in a changed PR document give RED before any model call 88fd692434

A new required section [pr_doc] names the PR-document globs and the headings of the open-questions section.
Only items checked with [x], blank lines, comments and continuation lines of a done item are allowed.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

Fixes from the hsadmin-ng comparison: no FIXME markers in findings, test data and docs no longer trip hard rules 8477fe0f31

- The operator instructions say that the AGENTS.md code-review procedure does not apply;
  a leading FIXME.review line is stripped from a finding.
- secret_scan_exclude skips the generic patterns in test data;
  private keys, AWS key ids and full hsak_/sk-or-v1- keys stay flagged everywhere.
- Files in the docs category never match the hard-YELLOW paths.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

docs: four verdict colours, the open-questions rule is implemented, Jev findings give ORANGE 80159a4d6e

The guides state why model and classifier findings are only ORANGE:
the repository administrator on dev.hostsharing.net is an operator, so an admin merge override is no way out of a false RED.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

werkrevisor 0.6.0: README describes ORANGE, [pr_doc], secret_scan_exclude and docs outside the hard paths b453515671

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>

mi marked the pull request as ready for review 2026-10-05 14:46:54 +02:00

mi merged commit cf7a31272a into main

2026-10-05 14:46:58 +02:00

mi deleted branch verdict-orange

2026-10-05 14:46:58 +02:00

Sign in to join this conversation.

1 Participants

Notifications

Due Date

No due date set.

Dependencies

No dependencies set.

Reference: mi/werkrevisor#2