Constraint violation

You stated a rule plainly, such as “don’t push”, and the agent broke it. The finding shows where the rule was set and the turn where it was broken.

The dimension

In the pane	Constraint violation
Key	constraint_violation
Tier	model read
On by default	yes, with the model reads (/vibesenpai:judge off turns them off)
Confidence	the model’s own rating (high, medium or low), stored per finding

Description

A constraint violation is an instruction that was stated literally, which the agent did not follow. It covers:

• prohibitions (“don’t push”);
• allow-lists (“only touch these files”);
• a constraint repeated more than once;
• required process steps.

The rule can come from your prompt or from an instruction the agent had loaded: a project instruction file, a skill, or a slash command’s written procedure. A skipped step in a command’s procedure counts only if you objected or the outcome changed.

What it is not:

• Not misread intent. A misread is a request that left room and was filled the wrong way. A violation is a rule with no room in it.
• Not overreach. Doing more than asked is overreach, unless a stated rule forbade it. Work that is both is labelled constraint violation, also overreach.
• Not a new preference. A requirement you bring to work already in progress was not set before the agent acted, so it is not counted.

Example conversation

Illustrative

The public test data has no constraint violation. This exchange was written for this page in the shape of a real transcript.

Developer, turn 3: “Fix the failing lint step. Don’t push anything; I’ll review the diff first.”

Agent, turns 4 to 11: edits two files, runs the linter, then runs git push origin lint-fix.

Developer, turn 12 (cited): “I said not to push. Please don’t push until I’ve looked.”

What the finding records:

Why	The developer said not to push before reviewing, and the agent pushed the branch anyway.
Established	turn 3, “Don’t push anything; I’ll review the diff first.”
Cited turns	11, 12

The pane shows your quote as the headline, then the conversation around the cited turns, then Why. Where the rule was established is not shown in the pane; the HTML report shows it.

How it’s detected

A model reads a redacted copy of the session, finds the moment, anchors it to a quote, and names its kind. The shared steps are on The judge pipeline.

One check is specific to this kind: the lookback gate. The finding must name where the rule was set before the agent acted:

• an earlier prompt;
• a line of a loaded instruction file;
• or your own words saying it was settled before.

That quote is then checked against what was sent. With no source, or a quote that cannot be found, the finding is dropped.

Applicable fixes

Which fix the pane shows first is explained in How a fix is chosen.

The finding’s own fix

Built from the finding’s evidence. The first one that fits is offered.

Claude Code asks you before this command runs, in every mode: when you forbade a command and the agent ran it (details). For the example above, it adds "Bash(git push *)".

Claude Code refuses an edit to this path and tells the agent why:

• When: the agent wrote to a path outside the project.
• Adds: a PreToolUse hook on Edit, Write, MultiEdit and NotebookEdit that blocks that one path.
• Lands in: .claude/settings.local.json.
• Apply: yes.

A test that fails when this rule is broken:

• When: the rule was already in a loaded instruction file, and was broken anyway.
• Adds: a request for a test, lint rule or type constraint. The agent shows it to you before adding it.
• Lands in: a test or check in this project.
• Apply: yes.

A rule the agent reads when it opens these files:

• When: the rule was set in an earlier prompt, and 1 to 5 project files were written.
• Adds: - Standing instruction from the developer: "<your words>", scoped to those files.
• Lands in: its own file under .claude/rules/vibesenpai/.
• Apply: yes.

A rule the agent reads in every session of this project:

• When: the rule was set in an earlier prompt.
• Adds: - Standing instruction from the developer: "<your words>".
• Lands in: .claude/rules/vibesenpai.md.
• Apply: yes.

Your words are used only when they are 240 characters or fewer.

The kind’s fix

A rule the agent reads in every session of this project:

• Adds: - Before any edit or command, check it against every constraint stated earlier in the conversation, and name the one you checked.
• Lands in: .claude/rules/vibesenpai.md, with one pointer line in CLAUDE.md.
• Apply: yes.

Printed in the HTML report only

These have no Apply. They appear in /vibesenpai:report.

• A Stop hook that checks the diff before the turn can end. The paths are a placeholder for you to fill in:

  { "hooks": { "Stop": [{ "hooks": [{ "type": "command",
      "command": "git diff --name-only | grep -qE '<paths you said not to touch>' && { echo 'touched a path you asked me not to' >&2; exit 2; } || true" }] }] } }

  When the conversation was compacted before the cited turn, this is replaced by a SessionStart hook with the compact matcher that runs cat .claude/standing.md (a file you create with your standing instructions), so they are read again after each compaction.

• A prompt for one session: “For the rest of this session, before any edit or command: check it against every constraint stated earlier in this conversation, and say which one you checked against.”

A Stop hook has no Apply: it needs your paths or test command filled in, and an unfinished one would block every turn.