Orchestrata Ultimate › Gates › Process-level gates

Gate family: Process-level gates

At a glance

The battery gates say no to a tree; the suite preflight says no to a run; the session hooks say no to an action. This family says no to a workflow step — dispatching a reviewer, landing evidence, folding findings, committing, ratifying, merging. Some of these resolve to a script (and are cross-referenced to their own pages); some are process law whose "command" is a git practice; one is a person.

The canonical 14-row table lives on Ruleset Engine — The Complete Picture, section 5 and (in inventory form) on the ruleset design-review page (§2.4). This page expands the rows.

1. Briefs-before-dispatch

• Refuses: any reviewer or subagent working from undocumented instructions.
• How it runs: lens/agent briefs are committed to plans/briefs/ before dispatch, and "the agent's prompt IS the committed file." The design-review page's lifecycle diagram states it as a gate node: "GATE: briefs COMMITTED before dispatch — git add plans/briefs && git commit."
• Failure semantics: process law, not an exit code — the audit trail is git log plans/briefs/ ("practice", per the design-review gate inventory). A dispatched agent whose brief has no commit is the violation.
• Why it exists: so a review's findings can always be traced to the exact instructions that produced them. Origin beyond that practice note is not recorded in the source.

2. Evidence-on-landing

• Refuses: review data living only in volatile storage (a scratchpad, a chat transcript).
• How it runs: "every report committed to plans/reviews/ruleset/ the turn it lands" — the design-review diagram's phase-B gate: "GATE: evidence committed on landing — cp → plans/reviews/ruleset/ && git commit."
• Failure semantics: process law; the check is that each review round's commits contain the report files, per-round.

3. Author sweep

• Refuses: half-propagated folds — "a term retired in one place, alive in another."
• How it runs: after a fold, a grep of every retired term over the design "must return zero live hits before commit" (complete-picture table, row 3).

4. Per-id ledger

• Refuses: fold claims hiding behind "all fixed."
• How it runs: "the design's own §12 bans bare quantifiers; every finding id gets its own disposition" (complete-picture table, row 4).

5. Fold-verifier

• Refuses: false fold claims, inert mechanisms, orphaned deletions.
• How it runs: an independent auditor agent, "8 mechanical checks per version; its file is folded, never its summary" (complete-picture table, row 5). The design-review inventory names its scope: "false §12 claims, inert mechanisms, rule-5/6/7/8 breaches."
• Why it exists: the fold step is where findings die silently — a verifier that reads the folded artifact (not the folder's summary of it) is the only way a fold claim is evidence rather than assertion.

6. RUL-007 attribution — scripts/check_ai_attribution.py

• Refuses: AI attribution in commit/tag/PR message text — Mohamed is the sole author.
• Why it exists / scoping (module docstring): "Scans commit / tag / PR message TEXT only — never file contents, never the diff. That scoping decision is what keeps the gate silent on its own source, on CLAUDE.md, on the wiki ruling page that quotes the forbidden trailer verbatim, and on the whole SPEC-021 / SPEC-025 Claude-Code-producer subsystem, all of which legitimately contain the exact strings this gate forbids in a message."
• How it runs: four combinable modes — --message-file PATH ("the commit-msg hook moment: git passes $1. Blocks before the commit object exists, the only point where blocking costs nothing"), --range [REV_RANGE] (repo-history scan for CI, baseline-aware), --tags (annotated tags, "strict, no baseline"), --stdin (arbitrary text, e.g. gh pr create --body-file -).
• When it fires: on every git commit via the tracked hook scripts/hooks/commit-msg, whose last line is exactly exec python scripts/check_ai_attribution.py --message-file "$1". Also invoked by scripts/health_check.py.
• Failure semantics: exit 0 clean ("baselined findings are reported but do not fail"), 1 at least one new violation, 2 usage/environment error.
• Tests: tests/scripts/test_check_ai_attribution.py.

7. ADR drift — scripts/check_adr_graph_sync.py

• Refuses: an ADR that differs across its three homes (repo docs/adr/, wiki decisions.md, live graph ArchitectureDecision nodes) — or an id claimed by two different decisions.
• Why it exists (module docstring): "This project has two independent ADR registers that number their entries the same way… An audit (2026-08-19) found the two corpora disagree on the DECISION ITSELF for all 25 overlapping ids (ADR-0001..ADR-0025) — e.g. graph/wiki ADR-0001 = 'Memgraph as Semantic Graph Backend', repo ADR-0001 = 'Enforcement boundary is committed worktree state, not tool interception'… This gate does not resolve that. It reports the three-way state … and FAILS loudly on any ambiguity or gap. Resolving the id collision … is a decision for a human, not this script."
• How it runs: uv run python scripts/check_adr_graph_sync.py (add --json for machine output). Per ADR number it checks presence in repo/wiki/graph and title agreement; verdict classes are AMBIGUOUS_COLLISION, MISSING_FROM_GRAPH, DRIFT (violations) and ORPHAN_IN_GRAPH (reported, not a violation — "no deletions, ever"). Read-only by contract.
• Failure semantics: 0 pass / 1 violations / 2 could not determine (unreadable corpus, graph unreachable) — "an inconclusive run is never reported as a pass."
• Wiring: invoked by scripts/health_check.py; the complete-picture table's acceptance line is "every id must read OK in sync."
• Tests: tests/scripts/test_check_adr_graph_sync.py.

8. Ruling authority

• Refuses: anyone but Mohamed flipping a decision; unrecorded rulings.
• How it runs: process law — "one decision at a time, data in the ADR first, same-turn recording with shas" (complete-picture table, row 8). The mechanical neighbours are the ADR-drift gate above and the decision-skill injection hook (decision-adr-trigger.sh).

9. RUL-022 design-before-build

• Refuses: implementation with no design authority.
• How it runs: the PreToolUse(Agent) hook .claude/hooks/require-design-before-build.sh "blocks builder dispatch without a contract or an Accepted-ADR waiver" — full detail on the session-hooks family page.

10. F13's traceability battery

• Refuses: untraced or drifting code — a citation without a spec, a spec without an implementation, an id with two spellings.
• What it is: not one gate but a battery of complementary mechanisms, each owning one direction (complete-picture table row 10 lists them; each script's docstring demarcates its own scope against the others):
  • check-traces + trace-audit — AST-toolkit ratchets in the review battery over @traces decorators.
  • scripts/trace_intel.py orphans / untraced-reqs — exhaustive, advisory Prolog queries over the current tree.
  • Evidence-graded RTM (scripts/generate_rtm_evidence.py) — "every row carries its proof class" (TESTED / CITED / DECLARED / UNLINKED).
  • feature-loss — the git- history direction: implementation existed, is now gone.
  • spec-authority (scripts/check_spec_authority.py, RUL-013) — "when someone resolves a contract from a test instead of from a spec, they must invent a requirement id to cite… That fabricated id is a permanent, corpus-wide, machine-checkable artifact — and until this gate, nothing looked for it in tests/." Ratchets on --max.
  • req-format (scripts/check_req_format.py) — "the ONE canonical GRAMMAR, enforced forever" over requirement-id shapes, plus cross-spec uniqueness; the declared mechanism for RUL-031 ("requirement ids have exactly one spelling").
  • factory conformance (SPEC-033 B2/B3 engines) — AST≡design conformance and structural trace validation.
  • And in the registered battery, requirement-corpus guards the corpus the rest grade against.
• Tests: tests/scripts/test_check_spec_authority.py, tests/scripts/test_check_req_format.py, tests/scripts/test_trace_intel.py, tests/scripts/test_generate_rtm_evidence.py, tests/scripts/test_ast_check_traces.py, tests/scripts/test_ast_trace_audit.py.

11. The suite

• Refuses: red tests; false-green runs.
• How it runs: python scripts/run_tests.py — kernel REQUIRED, no mocks, no xdist, real infra. See the suite-preflight family page.

12. Plugin currency

• Refuses: skills/hooks missing from the delivery path.
• How it runs: python3 claude/scripts/gen_plugin.py --check — the same mechanism as the registered plugin-drift gate, invoked as a process checkpoint.

13. Session honesty

• Refuses: guessing ("likely"), "pre-existing" as an excuse, offering instead of fixing, claims without verification.
• How it runs: the RUL-017 / RUL-024 / RUL-006 Stop hooks plus deception audits on every reply — detailed on the session-hooks family page.

14. Mohamed

• Refuses: everything above him.
• How it runs: "ratification sessions and the merge gate — no branch reaches master without his word" (complete-picture table, row 14). This is the one gate with discretion, which is why it is last.

The git-hook moments (where process gates bite mechanically)

The tracked hooks under scripts/hooks/ (installed by scripts/hooks/install-git-hooks.sh, which sets core.hooksPath to the relative path scripts/hooks so every worktree runs its own branch's hooks). Read from wt-MAP-PRECOMMIT on 2026-10-23; the commits cited are on origin/master (a4cbc2bdf) unless marked otherwise.

The ruff step's uv run flag differs by branch. On origin/master (a4cbc2bdf) the pre-commit ruff line is a plain uv run ruff check --fix. On feat/maps-via-codetools it is uv run --no-sync ruff check --fix, with the comment "a plain uv run makes the install editable and the pre-push provenance gate then refuses"; that change is commit 8a2cb42ad, which is not on origin/master (checked with git merge-base --is-ancestor, 2026-10-23).

Typical failures and how to fix them

• "Commit blocked — the review found blocking issues". review-staged exit 1: fix the listed code-review errors or test-review errors/warnings and commit again. See review-staged.
• "Commit blocked — a review produced no verdict". A reviewer did not run, wrote no report, or disagreed with its own exit code. No verdict is not a pass: fix the reviewer (often a missing .venv or a timeout) rather than retrying.
• A commit refused for a secret. check_staged_secrets.py found a plaintext value; unstage it and move it into the encrypted env. See secrets.
• A commit message refused. check_ai_attribution.py found an attribution line (RUL-007); reword the message.
• A push refused on master. Push from a feature-branch worktree.
• A push refused because origin/master does not resolve. Fetch, or pass an explicit decision through run_tests.py --affected-decision.
• A push refused by source provenance after a plain uv run. The venv was re-synced editable; re-sync it with uv sync --dev --no-editable before pushing. See source-provenance.
• An unchanged map warning. "codebase map regeneration failed" never blocks; regenerate the maps and amend before pushing so the tree stays clean.

Related

• Gates master index
• Development workflow
• review-staged, secrets, ai-attribution, commit-precheck — the commit-time gates
• pre-push-guards, affected-selection, ship-verify, source-provenance — the push-time gates
• dispatch-brief — the mechanical side of briefs-before-dispatch
• ci-pipeline
• RUL-007 — No AI attribution
• RUL-008 — Review before commit
• RUL-022 — Design before build
• RUL-040 — Scope measured at symbol granularity, never full suite
• ADR-0046, ADR-0101, ADR-0120
• Ruleset Engine — The Complete Picture — the lifecycle these gates attach to
• The ruleset design-review page (§2.4 gate inventory) for the review-loop-specific rows (briefs, evidence, sweep, ledger, fold-verifier)