policy

name	policy
description	Defines the contract (policy.schema.json) and read substrate (policy-read.mjs) for the central autonomy policy. The SSOT for the 3 action classes, gate vocabulary, policy.yaml hierarchy, and zero-config defaults. Referenced by other skills and hooks.
user-invocable	false

policy - Central autonomy policy contract and read substrate

A referenced companion (ADR-0028 R3) that consolidates HITL gates — previously scattered across each skill's prose — into a single-file declaration. Provides the SSOT for the contract governing action classification and gate decisions (policy.schema.json), and the all-adapter CLI that reads its effective values (policy-read.mjs). Aggregation and application (on the side of each skill / hook) are built separately on top of this contract.

Principles

• Unify the contract: Declare "which actions require approval" in a single policy file rather than in each skill's prose. Each skill simply classifies its own actions into one of the 3 classes and looks up the policy.
• fail-safe (not fail-open): A broken or adversarial policy must never loosen the level of autonomy. policy-read.mjs does not throw, but always falls back to the stricter side (ask) for values it cannot trust. This is the opposite direction of safe-side fallback from observability's fail-open (pass through even on failure).
• Zero-config = equivalent to current behavior: When policy.yaml is absent, the default is equivalent to current behavior (acceptance criterion). Adding configuration doesn't break things via "configure and forget."

The 3 action classes

Each skill classifies its own actions into one of the following fixed vocabulary terms:

Class	Examples	Zero-config default
reversible-local (reversible/local)	Edits on a branch, safe branch delete, worktree prune	proceed (+ audit trail)
externally-visible (externally visible)	Filing an issue, creating a PR, applying topics, PR comments	batch-confirm (one batch confirmation)
irreversible (irreversible/destructive)	merge, publish, stash drop, force operations	ask (always an Approval Gate)

Gate vocabulary

Only the following 3 values can be assigned to each class / action:

gate	Meaning
proceed	Execute as-is and leave an audit trail (no confirmation)
batch-confirm	Confirm once, in a batch, before execution
ask	Pass through an Approval Gate (explicit approval) for every single operation

policy file hierarchy

~/.agents/policy.yaml          # user default (HOME-anchored)
<repo>/.agents/policy.yaml     # repo override (overrides user)

repo overrides user, and user overrides the zero-config default. File format (policy.schema.json is the SSOT):

schema_version: 1
classes:
  reversible-local: proceed
  externally-visible: batch-confirm
  irreversible: ask
actions:
  merge: ask            # per-action override (takes precedence over the class default)
  issue-create: proceed

classes declares the gate for the 3 classes, and actions overrides the gate per kebab-case action name (takes precedence over the class default). Unknown top-level keys are mechanically rejected by additionalProperties: false. Action names are open, since each skill declares its own; only the value (gate vocabulary) and key shape are validated.

Contract (policy.schema.json is the SSOT)

policy.schema.json (a sibling of this skill) is the sole SSOT. Both policy-read.mjs and the tests validate by reading this file, so doc and code never drift apart. Following the same approach as skill-observability's event.schema.json, additionalProperties: false (at the top level) mechanically prevents "unintended loosening of autonomy" caused by typos or unknown keys. schema_version is fixed at 1.

Read substrate (policy-read.mjs)

policy-read.mjs (sibling, a CLI that runs on all adapters) reads and merges the user + repo policy.yaml, and returns the merged effective values as JSON to stdout. To avoid a new runtime dependency, YAML is parsed with a hand-rolled parser covering only the minimal subset policy.yaml requires (nested mappings + key: value + comments).

# Return the merged effective policy (classes / actions / sources / degraded) as JSON
node policy-read.mjs

# Resolve the effective gate for a single action (per-action override → class default → stricter side)
node policy-read.mjs --action=merge          # => resolved.gate (default ask)
node policy-read.mjs --action=issue-create   # => resolved.gate (default batch-confirm)
node policy-read.mjs --class=reversible-local

# Specify the repo root explicitly (default is cwd)
node policy-read.mjs --repo-root=/path/to/repo

Output (example, under zero-config):

{
  "schema_version": 1,
  "classes": { "reversible-local": "proceed", "externally-visible": "batch-confirm", "irreversible": "ask" },
  "actions": {},
  "sources": { "user": false, "repo": false },
  "degraded": false
}

fail-safe conventions:

• File absent → zero-config default (equivalent to current behavior). degraded: false.
• Invalid gate value (schema mismatch) → the effective value for that class / action falls back to ask. It never falls through to a looser, lower-priority value.
• Unparseable file → ignore that file and set degraded: true. Resolve using the other file + defaults (the default for the dangerous class is already ask).
• Never throws on any failure; always exits 0. Does not break callers chained with &&.

PreToolUse enforcement hook (Claude Code / policy-hook.mjs)

The prose layer (each skill "classifying its own action and looking up the policy") is ultimately a request, which the model could ignore or bypass. policy-hook.mjs (a sibling PreToolUse hook of the same shape as usage-guard-hook.mjs) is gate enforcement on the execution-engine side, and physically closes off that gap. It fires before every tool call (including inside subagents), and only looks up the policy to decide deny/allow when it detects a specific irreversible command.

Detected command	Action looked up (class)	Zero-config default gate	Hook behavior
gh pr merge …	merge (irreversible)	ask	deny (exit 2)
gh release create …	release-create (irreversible)	ask	deny
git push --force / -f / --force-with-lease	force-push (irreversible)	ask	deny
npm / pnpm / yarn publish	publish (irreversible)	ask	deny

• gate → decision: If the resolved gate is ask, deny (exit 2 + reason to stderr); if proceed / batch-confirm, allow. The hook only hard-blocks on ask (batch confirmation is the responsibility of the caller / prose layer).
• Narrow gating (preventing an all-tool-deny accident): The policy is looked up only when it matches an irreversible command in the table above. Everything else (reads, edits, safe git/gh/npm, non-Bash tools) passes through untouched. Even if the matcher has a bug, it can only fail in the direction of "missing a dangerous command" — it never denies every tool and wedges the session.
• Doesn't hard-stop on a transient blip (carried over from usage-guard-hook.mjs):
  • (a) file kill-switch: If ~/.claude/policy-guard/DISABLE exists, it's an immediate no-op allow at the very start. Running touch from a ! shell releases it instantly within the session, with no config edit needed.
  • (b) Policy unreadable / unparseable → allow + stderr warning: When policy-read.mjs returns degraded, or the resolver throws, etc. — i.e., whenever the gate cannot be trusted — allow rather than deny. Even if the gated command can be detected, if the gate value can't be trusted, let it through. Since the resumable prose-layer checkpoint (drive / health / lessons-triage) is the primary gate that looks up the same policy-read.mjs (fail-safe to ask), and the hook is a secondary net, it falls to the side of not stopping the session over a broken policy.
  • (c) proceed override (equivalent to --merge): A caller that has already resolved the gate and been delegated autonomy (e.g., drive --merge overrides merge to proceed in prose) exports POLICY_GUARD_PROCEED=<action>[,<action>…] (all / * also work). The hook allows that action without re-gating it. This ensures legitimate opt-in merges aren't blocked by the enforcement net.
• subagent: Appends the payload's agent_id to the deny message as [origin: subagent <id>] (functions as a mid-unit ceiling for a running worker).

Wiring (recommended: hooks add policy)

The hook script is bundled in every adapter's payload. Wiring into settings is handled by the CLI (#174 PR 3 added support for hooks add policy; same shape as usage-guard / observability):

# Wire the PreToolUse policy gate into ~/.claude/settings.local.json (absolute path resolved automatically)
npx @ozzylabs/skills hooks add policy

# Check wiring status
npx @ozzylabs/skills hooks status

# Unwire (removes only the entry it wrote)
npx @ozzylabs/skills hooks remove policy

The CLI resolves the absolute path of policy-hook.mjs from the installed skill dir, shows a diff, and writes it after confirmation (--yes for non-interactive, --dry-run for plan-only, --scope=user for settings.json). It is idempotent (re-running add is a no-op) and never touches entries other than the one it wrote. The policy that the repo does not distribute settings / hooks remains unchanged — the CLI merely writes to local settings with the user's consent.

A policy.yaml template can be generated with npx @ozzylabs/skills policy init (--scope=repo for <repo>/.agents/policy.yaml); an existing file is skipped, not overwritten.

Fallback (manual wiring): If not using the CLI, manually add one entry to ~/.claude/settings.local.json (settings reload mid-session, so no restart is needed):

{
  "hooks": {
    "PreToolUse": [
      {
        "matcher": "Bash",
        "hooks": [
          {
            "type": "command",
            "command": "node /home/<you>/.claude/skills/policy/policy-hook.mjs"
          }
        ]
      }
    ]
  }
}

Fill in the hook script path as an absolute path by hand (skill-dir-relative references don't work inside settings). The path varies by environment:

• user-scope (npx @ozzylabs/skills install): ~/.claude/skills/policy/policy-hook.mjs (~ is not expanded, so write it out in full as /home/<you>/.claude/…)
• dogfood (inside the skills/commons repo): <repo>/.claude/skills/policy/policy-hook.mjs

Both point to "the policy-hook.mjs at the same level as policy-read.mjs." Since irreversible commands go through Bash, "Bash" is sufficient as the matcher ("*" also works — non-Bash tools have no command and pass through).

Disabling: touch ~/.claude/policy-guard/DISABLE (instant, no config edit needed), or remove the above entry from settings (permanent).

Scope of application

This skill provides the contract (schema) + read substrate + PreToolUse enforcement hook. The following are built in separate PRs on top of this contract (not included in this PR):

• Application by each skill (replacing the gate descriptions in implement / lessons-triage / topics etc. with "classify your own action and look up the policy." Already implemented in R3 PR2/PR3).
• Automatic hook wiring (hooks add policy. Already implemented in #174 PR 3. See "Wiring" above).

Notes

• Does not read .env files.
• This skill is read-only. It does not generate or rewrite policy.yaml (that is handled by policy init in a separate PR).
• fail-safe means "fall back to the stricter side," the opposite direction from observability's fail-open (pass through). Do not confuse the two.