cognitive-signature-ide

Anthropic Claude Advisor pattern — consulted at executor inflection points (low-confidence extraction, ambiguous dimensions, conflicting governance, unexplained drift) for strategic reframe rather than parameter-tuning.

CogSig v2 enforcement layer — convert signature from inject-only to inject+act. PreToolUse hook checks every Edit/Write/MultiEdit against the active signature.json and warns or rejects per the user's selected mode.

Prepend the current signature.json to Claude's context so code suggestions can mirror the user's style (Claude reads the signature each prompt and decides how closely to follow). INJECT-CONTEXT tier — see `skills/enforce/SKILL.md` + README "Architecture: Inject + Act" for the v2 force-mechanic. Auto-fires on every UserPromptSubmit via plugin hook; also invokable manually as `/cogsig inject`.

Slash command to enable, disable, or diff the cognitive signature injection.

Pattern-review surface for pending patterns. After extraction, diff new-signature vs permanent-signature — items at instance_count ≥ 2 not already in permanent become "pending." User approves / rejects / edits via slash commands, OR Claude surfaces them inline via inject.py next-response context.

One-command auto-seed from Claude Code session history. Scans ~/.claude/projects/**/*.jsonl, aggregates user-typed directives, runs extract → signature.json active. The default onboarding path.

Call Opus 4.7 on cached code samples to produce a structured signature.json capturing the user's coding style.

Sample recent code the user has written — most-recent commits + recent file edits — and stage them for signature extraction.

Load an imported signature JSON as the active signature. Sets origin=imported so Historian classifies any drift as EXPECTED.

Write a shareable signature JSON from the current signature.json. Stamps origin=self and optional team_id.