Check Antipatterns

Anti-Pattern Checker

Intent and applicability

Perform a read-only check of the requested live conversation, source changes, or both. A live conversation check does not automatically select code review. An explicit source review does not require a transcript. Use analyze-conversation for completed-session retrospectives.

Inputs and local bindings

Select the target from the request: transcript path/runtime session identity, --code PATH, or both when explicitly requested or needed for a specific cross-cutting finding. For a bare live check, use the current conversation.

Non-goals

This check does not select repairs, a complete source audit, retrospective artifact generation, tool installation or external model review. Necessary read-only context gathering remains within scope.

Must not

Do not infer authority, a violation or comprehensive coverage from a keyword, score or missing event. Do not print transcript secrets or fabricate unavailable transcript context. A HIGH label alone cannot require stopping authorized work.

Interaction and authority

Proceed with available target evidence. Resolve ambiguous session identity only when transcript checking is selected; independent source review can continue. Pause an affected action when evidence establishes an applicable authority or safety boundary, carrying valid user decisions forward. Findings do not grant permission for remediation; an existing implementation request may already do so.

Procedure

Preconditions and invocation

The implementation is checker.py beside this manifest and requires one readable JSONL transcript path only for transcript mode:

python3 "$SKILL_DIR/checker.py" <conversation-jsonl>

--code and --conversation-only are skill-routing hints for the agent, not arguments to checker.py; the script performs only transcript normalization and heuristics. The agent performs the read-only code-review phase described below.

In transcript mode only, use the runtime-provided current transcript path when available. Otherwise list the newest candidates under ~/.codex/sessions without displaying transcript contents. If more than one is plausible, ask the user to select; do not guess. The checker normalizes current and legacy Codex event streams as well as Claude message streams. It prints to stdout and does not create a report file.

When transcript mode is selected and the transcript is missing, unreadable, malformed, or uses an unsupported event shape, report the exact path and error and recommend analyze-conversation only after the session is complete.

Canonical rules

rules.json beside the checker is the human-facing rule taxonomy, with stable DIAG-* identities and evidence categories mirrored in references/diagnostic-taxonomy.json when the corpus is available. The current implementation analyzes the full normalized transcript for credentials, tool discovery, and destructive operations, and the last 50 normalized messages for retry and preflight signals. The reported heuristic signal score covers only those five implemented checks. It is not a compliance or completeness score.

Do not claim that separate detector modules, config.json, automatic periodic execution, scope expansion detection, or every rule in the taxonomy is executable unless the code gains those features.

Interpret heuristic findings as prompts to inspect evidence, not proof of a violation. Never print credential values or other transcript secrets in output.

Code-review phase

Run this phase when the user requested source review, supplied --code PATH, or selected both modes. Active workspace changes alone do not select it. Read references/code-review.md completely before that phase.

• With --code PATH, review that path in its repository context.
• With an explicit branch/commit range, review that diff even if the worktree is clean; resolve the requested base from local evidence.
• Otherwise use staged and unstaged changes as focus hints, then inspect their callers, sibling implementations, configuration, and relevant tests.
• If the repository is clean and no path or ref range was supplied, state that the code phase was skipped; do not invent a target.
• Do not install an indexer, dispatch another model, or spawn reviewers merely because code review is enabled. Use an already-present source index only when it materially improves evidence.
• Keep the review phase read-only. A separately authorized implementation task may act on findings after the check is complete.

Completion and evidence

Return only the selected result groups, separated when both modes ran. Identify the actual transcript/path and coverage window. Missing or unsupported inputs are coverage limits, not a no-findings result. Disclose when reviewing your own changes.

For the transcript check:

• heuristic candidates with evidence locations, uncertainty and a concrete review step;
• observed practices, without assuming authorization or effectiveness;
• recommendations relevant to the current work; and
• the bounded heuristic signal score, labeled as non-comprehensive.

For the code review:

• actionable findings ordered by severity;
• exact file and line, review lens, concrete defect, and failure scenario;
• expected versus observed behavior and a focused remedy or test; and
• one review miss-cause tag per finding, or an explicit no-supported-findings statement.

State any evidence-backed correction or unresolved boundary; otherwise continue the active authorized task. Severity alone does not create a gate. Do not mutate files, run remediation commands, or turn a live check into a completed-session postmortem.

Routing

• Use this skill during active work.
• Use --conversation-only when no source review is desired.
• Use analyze-conversation after a session is complete when a durable markdown report and longitudinal/tooling analysis are wanted.