The per-tool judge compares each mutating tool call against the classifier's
distilled task summary read from router-state. That summary is lossy and
frequently "(unknown)" even for a perfectly explicit user request — and with an
unknown task the judge has nothing to compare against, so "Сомнения → NO"
blocked every real edit. Reproduced repeatedly this session: an explicit
"реализуй ... main() ..." prompt still classified unknown → all edits blocked,
including the judge's own fix. Calibration 2 (allow on unknown) was rejected by
the owner as a discipline hole.
Calibration 4 (soft, scope-preserving): when — and only when — the classifier
summary is "(unknown)"/empty, fall back to judging against the user's actual
last prompt (the ground-truth request) instead of nothing. The judge still runs
and still blocks on doubt; it just uses better evidence. When the summary is
meaningful, behaviour is unchanged (the user-prompt reader is not consulted).
When both summary and prompt are unavailable, the task stays "(unknown)" and
doubt→block is preserved.
NOT calibration 2: this does not blindly allow on unknown — it re-grounds the
judge in the literal user request, which the controller cannot fabricate (the
user writes it; it is read locally from the session transcript).
- tools/llm-judge-per-tool.mjs: resolveEffectiveTask(declaredTask, lastUserPrompt).
- tools/enforce-llm-judge-per-tool.mjs: runPerTool reads the last user prompt
(helpers.lastUserPromptText + readTranscript) only on an unknown summary;
main() binds it.
Regression: judge tests 57/57 GREEN; full tools suite 1951 passed | 2 skipped.
The 6 remaining failures are uncommitted point-2 WIP in
enforce-parallel-session-lock.test.mjs — not part of this change, not committed.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Layer-4 per-tool judge over-blocked: it judged every Skill/Edit/Write/
Bash/Task against the declared task and blocked on doubt. A vague prompt
classifies as unknown/ambiguous, so the judge then blocked essentially all
artifact-producing tools — including the prescribed §17 skill entry and the
mandatory TDD test run — making legitimate, owner-mandated work impossible
and blocking its own fix (3 reproduced blocks this session).
Calibration 1 (scope fix, NOT a discipline drop): remove `Skill` from
MUTATING_TOOLS in tools/llm-judge-per-tool.mjs. Invoking a skill mutates no
state and is the §17-mandated entry into work; the real mutations it leads to
(Edit/Write/MultiEdit/Bash/PowerShell/Task/commit/push) stay fully judged.
Calibration 3 (scope fix, NOT a discipline drop): add isTestRunnerBashEvent to
tools/enforce-llm-judge-per-tool.mjs and skip it in runPerTool, mirroring the
existing readonly-Bash exemption. A test run (vitest/pest/phpunit/php artisan
test/composer test/npm test) only inspects + reports and is a mandatory TDD
step; commands chaining to a mutation (&& ; | backtick $() are NOT exempt.
doubt→block on real mutations against a known task is unchanged (covered by the
"mutating Bash (git commit) STILL judged" test). Calibration 2 (allow on
unknown task) was rejected by the owner as a discipline hole and not added.
Regression: vitest tools-only 1945 passed | 2 skipped (+18 calibration tests).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Дмитрий