Commands

Install from the marketplace, then add the learned scorer once. All you need is Claude Code.

# Install
$ claude plugin marketplace add tanjalshukla/HedwigCLI
$ claude plugin install hedwig@hedwig-marketplace

# Turn on the online classifier (once per machine)
$ /hedwig-setup

# Pull the latest release later
$ claude plugin update hedwig@hedwig-marketplace

Installed copies are pinned to a release, so run update to get new versions. Without /hedwig-setup the plugin still runs the deterministic heuristic scorer; the update just adds the learning classifier.

Shows how many edits were auto-applied vs. surfaced this session, with plain-English reasons for each surfaced edit. The primary dashboard. Run it after any task to see what Hedwig did and why.

# Example output
Hedwig · trust runtime

  Auto-applied     8   ████████████████░░░░  80%
  Surfaced         2   for your review

  8 of 10 edits auto-applied this session.

  Why it surfaced these:
    · api/routes.py: interface change, multi-file blast radius
    · auth/tokens.py is security-sensitive, no prior history

Shows all-time totals across every session in this repo.

Shows how the learned classifier has drifted from its cold-start baseline, per feature, with direction. Every delta comes from real decisions in this repo, not synthetic data.

# Example output
Hedwig · learned classifier drift (42 real decisions; scorer ACTIVE)

  ▼ blast radius          -8.60   (more cautious)
  ▲ change pattern risk   +8.22   (more trusting)
  ▲ new file              +5.07   (more trusting)
  ▼ prior denials         -3.31   (more cautious)

  ▲ toward auto-apply   ▼ toward check-in
  all drift from real decisions, no weight hand-tuned

Shows "classifier not active" until /hedwig-setup has been run and at least 10 real decisions have accumulated.

Lists self-corrections: edits Hedwig auto-applied that were later reverted by the agent or failed end-of-turn verification. Each one tightened the next similar edit and fed a corrective signal to the classifier. An empty list means nothing auto-applied had to be walked back.

# Example output
Hedwig · 1 self-correction (auto-applied, then walked back)

  • api/routes.py: reverted after auto-apply

  This tightened Hedwig's next decision on that file
  (and, via the classifier, on risk-similar edits elsewhere).

Reviews and confirms a behavioral pattern Hedwig has noticed. Patterns accumulate evidence silently; nothing changes until you confirm. Confirmed patterns fire in the governance cascade for future edits.

# Example output
Hedwig noticed a pattern:

  Always surface api_change edits in this repo for review?

  Why: you reverted 4 of the last 5 auto-applied api_change edits.

  Backed by 5 decisions.

Confirm with  /hedwig-learn confirm
Decline with  /hedwig-learn reject

Declined patterns stay in the bank for transparency; they're not discarded. Confirmed patterns can be reviewed and revoked via /hedwig-rules.

Asks Claude to review this session's decisions and propose standing rules based on what it observed. Proposed rules are candidates only; nothing changes until you confirm them via /hedwig-learn.

# Propose rules from this session's traces
/hedwig-notice

# See the raw trace digest first
/hedwig-notice traces

Every proposed rule cites the real trace IDs behind it. Uncited proposals are rejected automatically.

Views or sets hard constraints: non-negotiable path rules that override all scoring. Use deny to block a path entirely, check_in to always surface it, or allow to always auto-apply.

# List current constraints
/hedwig-rules

# Block all writes under config/prod/
/hedwig-rules add deny "config/prod/**"

# Always surface changes to the auth module
/hedwig-rules add check_in "auth/**"

# Remove a constraint
/hedwig-rules remove "config/prod/**"

Hard constraints are enforced before any scoring; they override leases, the classifier, and confirmed preferences. Path patterns use shell-glob matching against repo-relative paths.

Asks Claude to scan the repo for security-sensitive files that keyword matching alone would miss (config loaders, secret parsers, token handlers) and add them to Hedwig's security floor. Run once per repo.

# Scan and add findings to the security floor
/hedwig-scan

Files added here are treated like auth.py: always surfaced for review, regardless of what the classifier learns. Nothing changes until you confirm the scan's proposals.

Shows what Hedwig has stored about this repo: style guidelines, logic notes, and security facts. This is the memory layer that surfaces into the agent's context before each task.

# View stored guidelines
/hedwig-memory guidelines

# View logic notes
/hedwig-memory notes

# View security facts
/hedwig-memory security

Shows which files tend to change together in this repo, based on past sessions. Useful for understanding blast radius and catching edits that touch only half of a coupled pair.

# Show co-change pairs
/hedwig-cochange

Computed from session history in this repo. More sessions = better signal.