policy-algebra（政策代数）

/policy-algebra

Generates frozen, drift-checkable governance rules from thin intent strings (deep interview) or rich plan inputs (shallow confirmation). Produces fenced Starlark blocks callers can inject into downstream work, and exposes --verify for canonicalized drift detection.

Invocation

/policy-algebra <target>                      # generate rules
/policy-algebra <target> --write <path>       # generate + persist raw block
/policy-algebra --verify <frozen> <candidate> # drift check

<target> is a readable file path (SHALLOW mode) or an intent string (DEEP mode).

Adaptive input detection

if target is a readable file path:
    mode = SHALLOW
    input = read(target)
else:
    mode = DEEP
    input = target

Deep flow

1. Construct framing prompt for /grill-me: "Interview the user to extract rule invariants for: <target prompt> Focus questions on: Scope: what artifacts/events/phases does this rule govern? Invariants: what must always be true? what must never happen? Variable holes: what changes between instances? Combinators: are existing operators (+, diff, intersect, method calls) enough, or do we need new helpers? Failure modes: what happens when a rule can't be satisfied? Return a structured list of predicates in natural language. /policy-algebra will draft Starlark from your output."
2. Invoke /grill-me via the Skill tool with that prompt.
3. Capture /grill-me's output as the invariant list.
4. Continue to drafting (step 5 of shallow flow).

Shallow flow

1. Read file contents.
2. Scan for invariant declarations — headings like "Invariants", "Rules", "Constraints"; bullets with MUST / NEVER / ALWAYS; predicate-looking code blocks.
3. Extract those as the invariant list.
4. If extraction yields fewer than 2 invariants OR content is ambiguous: fall through to DEEP mode with the file contents as target.
5. Draft Starlark from the invariant list per notation.md conventions.
6. Pass draft to /review (no panel hint — /review assembles).
7. /review converges → frozen, OR caps → inherit /review escalation.

Drafting conventions

Group by subject (DEFAULTS, then named functions); keep each function body to 5 lines or fewer; one prose comment per function; no mutation. See notation.md for the full ruleset.

Freeze and return

After /review APPROVE, capture the exact Starlark text as the FROZEN block. Do not canonicalize at freeze time — canonicalization is a compare-time operation only, keeping freeze/return simple.

Output format:

ALGEBRA:

<raw frozen starlark block, as /review approved>

REVIEW_ROUNDS: PANEL: STATUS: CONVERGED FILE: (only when --write was used) NOTE: (optional)

If --write <path> was supplied, write the raw Starlark block to that path and include FILE: <path> in the return block.

--verify flow

1. Read both files.
2. Extract first `starlark fenced block from each file.
3. Normalize both blocks (strip trailing whitespace, collapse blank lines, trim edges).
4. If equal → print MATCH, exit 0. Else → print DRIFT + unified diff, exit 1.

Use the following bash to verify drift — no external dependencies needed:

# policy-algebra verify — inline, no dependencies
# Usage: bash verify.sh <frozen_file> <candidate_file>
normalize() { sed 's/[[:space:]]*$//' | sed '/^$/d'; }
extract_starlark() { sed -n '/^```starlark/,/^```$/p' "$1" | sed '1d;$d'; }

FROZEN=$(extract_starlark "$1" | normalize)
CANDIDATE=$(extract_starlark "$2" | normalize)

if [ "$FROZEN" = "$CANDIDATE" ]; then
  echo "MATCH"
  exit 0
else
  echo "DRIFT"
  diff <(echo "$FROZEN") <(echo "$CANDIDATE")
  exit 1
fi

Multiple-block behavior

When a file contains multiple ``` `starlark `` blocks, extractAlgebra` silently uses the first. This is intentional — wrap your canonical rules in the first fenced block of any document.

Failure modes

Exit codes

Cross-tool notes

• Claude Code: invokes /grill-me (DEEP interview) and /review (panel convergence) via the Skill tool. Both exist as project-local skills under .claude/skills/.
• Codex / Cursor: the Skill tool is Claude-Code-only. Read the relevant skill files directly:

- .claude/skills/grill-me/SKILL.md — for the interview procedure - .claude/skills/review/SKILL.md — for panel convergence Apply their procedures inline rather than invoking them as commands. The drafting, freeze, and --verify logic in this SKILL.md is tool-agnostic and works identically for Claude Code, Codex, and Cursor.

• CLI (--verify path) is runtime-agnostic: use the inline bash snippet in the --verify section above — no Node or pnpm required.

Governance

This skill is self-hosted: its own rules are encoded as a frozen Starlark governance block. Drift from these invariants is a bug. Verify with /policy-algebra --verify.

# policy-algebra governance — self-hosted
# FROZEN — do not edit; regenerate via /policy-algebra

DEFAULTS = {
    "min_invariants": 2,
    "exit_codes": {"match": 0, "drift": 1, "input_error": 2, "internal_error": 3},
    "required_output_fields": ["ALGEBRA", "REVIEW_ROUNDS", "PANEL", "STATUS"],
    "converged_status": "CONVERGED",
}

# SELECT_MODE: file path → SHALLOW; intent string → DEEP.
def SELECT_MODE(target):
    if is_readable_file(target):
        return "SHALLOW"
    return "DEEP"

# SHALLOW_OR_DEEP: fall through to DEEP when extraction yields < min_invariants or is ambiguous.
def SHALLOW_OR_DEEP(invariants, ambiguous):
    if len(invariants) < DEFAULTS["min_invariants"] or ambiguous:
        return "DEEP"
    return "SHALLOW"

# EXTRACT_BLOCK: use first starlark fence only; multiple blocks → first wins.
def EXTRACT_BLOCK(fences):
    if len(fences) == 0:
        return None
    return fences[0]

# OUTPUT_VALID: output must carry all required fields; STATUS=CONVERGED only after review.
def OUTPUT_VALID(output, review_loop_passed):
    fields_present = all([f in output for f in DEFAULTS["required_output_fields"]])
    converged_ok = output.get("STATUS") != DEFAULTS["converged_status"] or review_loop_passed
    return fields_present and converged_ok

# VERIFY: canonicalize both blocks and compare; never modify source files.
def VERIFY(frozen_canon, candidate_canon):
    if frozen_canon == candidate_canon:
        return DEFAULTS["exit_codes"]["match"]
    return DEFAULTS["exit_codes"]["drift"]

# EXIT_CODE: map outcome strings to numeric codes.
def EXIT_CODE(outcome):
    return DEFAULTS["exit_codes"].get(outcome, DEFAULTS["exit_codes"]["internal_error"])