legis-workflow

name

legis-workflow

description

This skill should be used when the user asks to explain or evaluate a policy cell, submit a graded override, check the override-rate CI gate, run a governance gate, read git branch/commit context, read the git-rename feed for Loomweave, gate a Filigree closure on verified binding evidence, route Wardline scan findings through governance, read recorded pull-request or CI check outcomes, run the policy-boundary-check, or back-fill SEI-keyed governance records — or when working in a project that uses legis for git/CI governance and graded enforcement.

Legis Workflow

Legis is the git/CI and governance side of the Weft suite. This skill is the depth behind the lean CLAUDE.md block: the full CLI reference, the MCP tool catalogue, the error/recovery table, and the worked patterns an agent actually runs. Keep it faithful to the installed legis — when in doubt, legis --help and legis <command> --help are authoritative.

What legis is

Legis answers what changed, in which branch/commit/PR/check context, and what governance or attestation state exists for that change? It is an SEI consumer (Loomweave remains the identity authority) and the suite's single governed judge — Wardline analyses trust; Legis governs it, one judge not two. It does not own issue state (Filigree) or code identity (Loomweave); it adds branch/commit/PR/check context and a graded enforcement layer on top.

Enforcement is a 2×2 of policy cells, each agent-set, each a distinct override flow:

	Judge OFF	Judge ON
Simple	chill — agent self-reports a recordable override; human reviews async (ACCEPTED_SELF)	coached — an LLM wall evaluates the override before it records; ACCEPTED_BY_JUDGE or BLOCKED (not self-clearable)
Complex	structured — block + escalate; a human operator must sign off before the gate clears (ESCALATED_PENDING)	protected — full machinery: HMAC-signed verdicts, decay sweep, override-rate gate, operator override

The operating invariant is agent-first: humans on the loop, not in the loop. Every cell produces an append-only audit trail keyed on SEI, so the record survives rename/move. The recorded override is the safety mechanism — an attributable audit event, never a silent pass.

Reaching the tools

Prefer the MCP tools (mcp__legis__*) when a Legis MCP server is attached; fall back to the legis CLI otherwise. Each surface maps thinly over the same service layer, so they agree on outcomes.

Identity is launch-bound. The MCP server is started with legis mcp --agent-id <name>; that --agent-id is the actor for every override, sign-off, and audit record the session produces. No tool schema accepts an actor argument — you cannot spoof or override identity from a call. (Contrast the CLI's sei-backfill --actor, which stamps appended backfill events from a one-shot command, not an interactive session.)

The MCP transport is stdio JSON-RPC (one object per line). Tool errors come back as isError results with a structuredContent envelope carrying error_code, message, recoverable, and next_action (see Error handling).

CLI reference

legis <command> [flags]. Most stores fall back to environment variables; flags override.

legis serve — run the Legis API server

• --host (default 127.0.0.1), --port (default 8000) — bind address.
• --governance-db — governance store URL (env LEGIS_GOVERNANCE_DB).
• --check-db — check store URL (env LEGIS_CHECK_DB).
• --protected-policies — comma-separated protected policy list (env LEGIS_PROTECTED_POLICIES).
• --loomweave-url — Loomweave identity API URL (env LOOMWEAVE_API_URL).
• --filigree-url — Filigree issue-tracker API URL (env FILIGREE_API_URL).
• --binding-db — sign-off binding ledger URL (env LEGIS_BINDING_DB).
• Judge flags (shared): --judge-provider (openrouter; omit to keep protected cells fail-closed), --judge-model (env LEGIS_JUDGE_MODEL), --judge-max-tokens (env LEGIS_JUDGE_MAX_TOKENS).

legis mcp — run the MCP stdio server

• --agent-id (required) — launch-bound agent identity; the actor for all records this session.
• --governance-db (env LEGIS_GOVERNANCE_DB), --check-db (env LEGIS_CHECK_DB).
• --policy-cells — policy cell registry TOML path (env LEGIS_POLICY_CELLS).
• --protected-policies (env LEGIS_PROTECTED_POLICIES), --loomweave-url (env LOOMWEAVE_API_URL).
• Judge flags (shared): --judge-provider, --judge-model, --judge-max-tokens.

legis check-override-rate — CI gate

Fails (exit 1) if the override-rate gate is FAIL. For CI use.

• --db — governance store URL (default mirrors the server's LEGIS_GOVERNANCE_DB / DEFAULT_GOVERNANCE_DB).

Prints override-rate gate: <STATUS> (rate=…, sample=…). A missing SQLite DB under CI=true (without LEGIS_ALLOW_MISSING_GOVERNANCE_DB=1) fails; otherwise it prints PASS_WITH_NOTICE and exits 0. A failed hash-chain integrity check exits 1.

legis governance-gate — run governance CI gates

Currently runs the override-rate gate (same implementation and --db semantics as check-override-rate). Use this name for the general CI gate entry point.

legis sei-backfill — resolve legacy locator-keyed records

Resolves legacy locator-keyed governance records through Loomweave batch resolve and emits a JSON report.

• --db — governance store URL (env LEGIS_GOVERNANCE_DB).
• --loomweave-url (required) — Loomweave identity API URL.
• --execute — append backfill events (omit for a dry-run report).
• --actor (default legis-sei-backfill) — actor stamped on appended events.

legis policy-boundary-check — boundary-evidence gate

Fails (exit 1) when @policy_boundary metadata lacks current behavioural evidence.

• --root (default src) — Python source root to scan.
• --repo-root (default .) — repo root for test_ref resolution.
• --format (text | json, default text) — human-readable lines vs machine-readable findings.

Prints policy-boundary-check: PASS (exit 0) when clean; otherwise one path:line: rule_id: qualname: reason per finding (exit 1).

MCP tool catalogue

All tools return a structuredContent JSON payload. Names are exact.

Governance / policy

Tool	Purpose
policy_explain	Explain which governance cell controls a policy/entity pair, whether that cell is enabled here, and which move the agent may make next. Reports matched_rule — the routing pattern that matched, or null when the policy fell through to default_cell (distinguishes a configured-but-disabled policy from an unconfigured name).
policy_list	List the policy-to-cell routing table (default_cell + the configured pattern rules) and every governance cell's real enabled state on this server. The complex tier (structured/protected) reports enabled: false without LEGIS_HMAC_KEY. No arguments.
policy_evaluate	Evaluate a policy against a target without recording an override. Returns outcome, detail, and any provenance_gap.
override_submit	Submit an override as the launch-bound agent. Routes to the governing cell and returns a discriminated outcome envelope (ACCEPTED_SELF / ACCEPTED_BY_JUDGE / BLOCKED / ESCALATED_PENDING / NEED_INPUTS).
signoff_status_get	Poll whether a structured sign-off request (by seq) has been cleared.
override_rate_get	Read the fixed operator force-past override-rate gate (status / rate / sample_size). Measures operator force-pasts; not movable by agent retries.
scan_route	Route Wardline scan findings through one cell, a severity_map, or a cell + fail_on threshold. Returns ROUTED or SKIPPED_DIRTY_TREE (typed amber skip).

Git

Tool	Purpose
git_branch_list	List local git branches and upstream divergence facts.
git_commit_get	Read one git commit by SHA or safe ref.
git_rename_list	List git rename evidence for a revision range (rev_range).
git_rename_feed_get	Loomweave-ready rename feed: committed renames over base..head plus optional uncommitted working-tree renames (include_worktree).

Pulls / checks

Tool	Purpose
pull_request_get	Read recorded pull-request metadata (number) with joined check outcomes.
check_list	Read recorded CI/check outcomes for a target_type of commit, branch, or pr plus a target.

Filigree binding

Tool	Purpose
filigree_closure_gate_get	Read whether legis holds verified binding evidence for closing a Filigree issue (issue_id). Requires the binding ledger to be enabled.

Override-submit outcomes (by cell)

• chill → ACCEPTED_SELF — self-cleared; human reviews asynchronously.
• coached / protected → ACCEPTED_BY_JUDGE (may be re-judged later) or BLOCKED. A BLOCKED verdict carries a blocked_reason_code (RATIONALE_INSUFFICIENT / CODE_VIOLATION / POLICY_HARD_BLOCK / UNCLASSIFIED), self_clearable: false, and next_actions: [REVISE_CODE, REVISE_RATIONALE]. A blocked attempt does not count toward your override-rate — you cannot self-clear past the judge.
• structured → ESCALATED_PENDING — human sign-off required; poll signoff_status_get with the returned seq.
• protected with missing inputs → NEED_INPUTS — supply the listed fields (e.g. file_fingerprint, ast_path) and resubmit.

Pass an idempotency_key on override_submit to make retries safe: a repeat with the same request returns the original outcome; a reused key with a different request is rejected (INVALID_ARGUMENT).

Error handling

Tool errors carry error_code, message, recoverable, and a next_action hint. Branch on error_code, not message text.

error_code	Recoverable	next_action
INVALID_ARGUMENT	yes	Correct the tool arguments and retry.
INVALID_CELL_SPEC	yes	scan_route routing is server-owned and unconfigured by default. The operator sets LEGIS_WARDLINE_CELL (e.g. =surface_only) or LEGIS_WARDLINE_CELL_BY_SEVERITY out-of-band, then relaunches. (Request-side routing requires the LEGIS_UNSAFE_WARDLINE_REQUEST_ROUTING opt-in — discouraged.) The error message names which kind of cell spec was rejected.
CELL_NOT_ENABLED	yes	Two enablement tiers, by cell — both operator-enabled, out-of-band. Simple tier (chill/coached) is reachable WITHOUT a key: the operator maps the policy to a cell via policy/cells.toml or LEGIS_POLICY_CELLS (LEGIS_DEV_DEFAULT_CELLS=1 selects the chill dev default), then relaunches. Complex tier (structured/protected and the binding ledger) additionally needs LEGIS_HMAC_KEY set by the operator out-of-band, then a relaunch. The error message names which cell is unenabled.
NO_SUCH_REQUEST	yes	Poll a known sign-off sequence returned by override_submit.
NOT_FOUND	yes	Refresh the target identifier and retry.
UNKNOWN_TOOL	yes	Call tools/list and use one of the advertised tool names.
GIT_ERROR	yes	Check the git ref or revision range and retry.
SERVICE_ERROR	yes	Inspect the error message before retrying.
AUDIT_INTEGRITY_FAILURE	no	Stop and ask an operator to inspect the governance trail.
INTERNAL_ERROR	no	Inspect the error message before retrying.

AUDIT_INTEGRITY_FAILURE (raised on a failed hash-chain verification or a binding ledger error) and INTERNAL_ERROR are not recoverable — do not retry; surface them to a human. Everything else is recoverable by fixing the input or asking the operator to enable a cell.

Two routing-specific notes for scan_route:

• Wardline routing is server-owned. Passing cell / severity_map / fail_on when the server already configures routing (LEGIS_WARDLINE_CELL / LEGIS_WARDLINE_CELL_BY_SEVERITY) returns INVALID_CELL_SPEC. Request-side routing is only honoured under the explicit LEGIS_UNSAFE_WARDLINE_REQUEST_ROUTING=1 escape hatch.
• An unsigned dirty-tree dev artifact arriving where signed provenance is required is not an error — it returns outcome: SKIPPED_DIRTY_TREE (a typed amber skip; nothing is governed). Commit for a signed artifact, or set LEGIS_WARDLINE_ALLOW_DIRTY=1 to govern it unsigned in dev.

Workflow patterns

Evaluate a policy cell, then submit a graded override

policy_explain {policy, entity}        # which cell governs, is it enabled, what move is next
# read explanation.cell and available_moves (already filtered to agent-callable tools)
override_submit {policy, entity, rationale [, file_fingerprint, ast_path, idempotency_key]}

• chill → ACCEPTED_SELF; you are done, the human reviews the trail async.
• coached/protected → if BLOCKED, do not retry verbatim — REVISE_CODE or REVISE_RATIONALE per next_actions; the judge cannot be talked past and the blocked attempt costs you nothing on the override-rate.
• structured → ESCALATED_PENDING; poll signoff_status_get {seq} until cleared: true. Do not proceed on the gated change until then.
• protected → if NEED_INPUTS, supply file_fingerprint + ast_path (the bytes and AST node the judge binds its verdict to) and resubmit.

Check the override-rate gate in CI

The gate measures operator force-pasts, not agent retries — a high rate means the policy is miscalibrated or an operator is breaking their own rules.

# in-session read:
override_rate_get {}                    # → {status, rate, sample_size}
# CI step (exit 1 on FAIL):
legis check-override-rate --db <governance-db>
#   or the general entry point:
legis governance-gate --db <governance-db>

Read the git-rename feed for Loomweave

Legis is the (contract-locked) rename provider Loomweave's SEI re-binding matcher consumes.

git_rename_feed_get {base, head?, include_worktree?}
#   committed renames over base..head, plus optional uncommitted working-tree renames
# lower-level evidence over an explicit range:
git_rename_list {rev_range}

Gate a Filigree closure on verified binding evidence

Before closing a governed Filigree issue, confirm Legis holds verified, SEI-keyed sign-off binding evidence for it.

filigree_closure_gate_get {issue_id}    # requires the binding ledger to be enabled
# only close in Filigree once this reports verified binding evidence;
# Filigree retains lifecycle authority — Legis only certifies the evidence.

If the ledger is not enabled you get CELL_NOT_ENABLED — ask the operator to wire LEGIS_BINDING_DB / --binding-db.

Route Wardline findings through governance

scan_route {scan}                       # routing is server-owned; pass only the scan
# → ROUTED (governed into the configured cell) or SKIPPED_DIRTY_TREE (commit, or
#   set LEGIS_WARDLINE_ALLOW_DIRTY=1 in dev)

Gate boundary evidence in CI

legis policy-boundary-check --root src --repo-root . --format json
#   exit 1 with findings when @policy_boundary metadata lacks current behavioural evidence