Skip to content

Gitleaks: narrow the fixture allowlist and run the guard in CI #2342

Gitleaks: narrow the fixture allowlist and run the guard in CI

Gitleaks: narrow the fixture allowlist and run the guard in CI #2342

Workflow file for this run

name: gates
# Enforce the standards this repo documents. The skill it ships now flags
# "standards documented but no gate enforces them" as a finding — so this repo
# wires its own documented gates (CLAUDE.md pre-commit block + definition of done)
# into CI rather than leaving them advisory. The gate logic lives in one place,
# scripts/ci-gates.sh, exercised by scripts/test-ci-gates.sh; CI runs that same
# helper against the real repo so the enforced check and its tests never diverge.
# PR-only: branch protection requires `gates` with strict up-to-date branches,
# so a PR's run already tested the exact tree that lands on main; a second run
# on the post-merge push would repeat it and burn runner minutes. Superseded
# runs on the same PR are cancelled.
on:
pull_request:
branches: [main]
workflow_dispatch:
concurrency:
group: gates-${{ github.event.pull_request.number || github.ref }}
cancel-in-progress: true
permissions:
contents: read # least privilege (OWASP Scorecard: Token-Permissions)
jobs:
# Heavy steps run as parallel jobs (wall time = slowest job, not the sum); the aggregate `gates` job below keeps the
# required check name. scripts/test-ci-gates.sh shards its cases by SHARD=i/N; `perpr` shards scripts/ci-tests.d with the same SHARD=i/N spelling.
selftests:
strategy:
fail-fast: false
matrix:
shard: [1, 2, 3]
runs-on: ubuntu-latest
timeout-minutes: 35
steps:
# SHA-pinned, not a mutable tag (the skill's own K rule for CI actions).
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 0 # fix-class gate diffs base..head; a depth-1 clone cannot resolve the range
persist-credentials: false # no later step pushes or fetches a private ref; GH_TOKEN env covers gh
# Derived files (SHA256SUMS, size-budgets.tsv rows, INDEX.md, INDEX-scripts.md) are not committed by lane PRs, so two
# PRs never conflict on them. Rebuild them here, before any gate reads them; a land commit (VERSION
# moved) is skipped, so there the committed files are checked byte-for-byte.
- name: Regenerate derived files (unlanded PR)
if: github.event_name == 'pull_request'
run: bash scripts/regen-unlanded.sh origin/main
# Prove the gate helper against its own fixtures (routing/version/privacy
# + the real install path) before trusting it on this repo.
- name: Gate helper self-tests
env:
SHARD: ${{ matrix.shard }}/3
run: bash scripts/test-ci-gates.sh
enforce:
runs-on: ubuntu-latest
timeout-minutes: 35
steps:
# SHA-pinned, not a mutable tag (the skill's own K rule for CI actions).
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 0 # fix-class gate diffs base..head; a depth-1 clone cannot resolve the range
persist-credentials: false # no later step pushes or fetches a private ref; GH_TOKEN env covers gh
# Derived files (SHA256SUMS, size-budgets.tsv rows, INDEX.md, INDEX-scripts.md) are not committed by lane PRs, so two
# PRs never conflict on them. Rebuild them here, before any gate reads them; a land commit (VERSION
# moved) is skipped, so there the committed files are checked byte-for-byte.
- name: Regenerate derived files (unlanded PR)
if: github.event_name == 'pull_request'
run: bash scripts/regen-unlanded.sh origin/main
# Enforce the documented gates on the real repo via the same helper the
# self-tests pin — one authoritative implementation, no inline copies.
- name: Enforce documented gates
run: |
set -euo pipefail
# Routing runs on every shipped skill dir — a reference under an overlay
# cannot ship unrouted (dogfoods "documented but unenforced is a finding").
# Globbed, not enumerated: a new skill adds no line here, so PRs on different skills never conflict.
for d in .claude/skills/*/; do
bash scripts/ci-gates.sh routing --max-bytes 24000 "$d"
done
bash scripts/ci-gates.sh version .
bash scripts/ci-gates.sh privacy --banlist .banlist.txt .
bash scripts/ci-gates.sh enumeration .
bash scripts/check-leader-doc.sh --selftest && bash scripts/check-leader-doc.sh
# Frozen per-file byte-count budgets (scripts/size-budgets.tsv):
# every shipped SKILL.md/references/*.md must stay within its
# budget, which only ratchets down via consolidation, never up.
# A raise is enforced by the "Size budgets ratchet down" step below.
bash scripts/ci-gates.sh size --config scripts/size-budgets.tsv .
# No git-tracked file at a banned image/media/archive/build-output
# extension (scripts/binaries-allowlist.tsv): stops the
# committed-screenshot/artifact bloat class; extension-scoped, not
# size-scoped, so a legitimately large text fixture never trips it.
bash scripts/ci-gates.sh binaries .
# Frozen per-archetype MUST-LOAD token-est ceilings
# (scripts/mustload-budgets.tsv), parsed from SKILL.md's own
# "Archetype -> load map" table: makes the always-paid must-load
# read cost per archetype visible and ratchets it down, never up.
bash scripts/ci-gates.sh mustload --config scripts/mustload-budgets.tsv .
# Rule-enforcement lint: a MUST/NEVER/ALWAYS rule in a SKILL.md must
# name a mechanism or be marked advisory. Existing debt is frozen in
# scripts/rule-enforcement-baseline.tsv; only NEW unenforced rules
# fail. Baseline growth vs the PR base is the next step below.
bash scripts/ci-gates.sh rules --baseline scripts/rule-enforcement-baseline.tsv .
# Four legs, SHARD=i/4 (the i/N spelling selftests uses): each runs only the scripts/ci-tests.d registrations that scripts/ci-tests-shard.sh lists
# for it, so the job takes as long as its slowest slice, not the sum. `gates` needs `perpr`, so it fails unless every leg passes.
perpr:
strategy:
fail-fast: false
matrix:
shard: [1, 2, 3, 4]
runs-on: ubuntu-latest
timeout-minutes: 35
steps:
# SHA-pinned, not a mutable tag (the skill's own K rule for CI actions).
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 0 # fix-class gate diffs base..head; a depth-1 clone cannot resolve the range
persist-credentials: false # no later step pushes or fetches a private ref; GH_TOKEN env covers gh
# Derived files (SHA256SUMS, size-budgets.tsv rows, INDEX.md, INDEX-scripts.md) are not committed by lane PRs, so two
# PRs never conflict on them. Rebuild them here, before any gate reads them; a land commit (VERSION
# moved) is skipped, so there the committed files are checked byte-for-byte.
- name: Regenerate derived files (unlanded PR)
if: github.event_name == 'pull_request'
run: bash scripts/regen-unlanded.sh origin/main
# A slow per-PR test registers as scripts/ci-tests.d/<slug>.bg.sh: it starts here, runs alongside this leg's serial
# tests, and the "Per-PR tests" step waits for it and fails on its exit code (a hang is killed after 15 minutes), so it adds no serial time.
# This leg's registrations (serial and background) are listed once, here, into $RUNNER_TEMP/bg/leg; both steps read that list.
- name: Start background per-PR tests (scripts/ci-tests.d/*.bg.sh)
env:
SHARD: ${{ matrix.shard }}/4
run: |
mkdir -p "$RUNNER_TEMP/bg"
date +%s >"$RUNNER_TEMP/bg/started"
bash scripts/ci-tests-shard.sh >"$RUNNER_TEMP/bg/leg"
unset SHARD # the tests must not inherit it: scripts/test-ci-gates.sh reads SHARD as its own slice
for t in scripts/ci-tests.d/*.bg.sh; do
[ -e "$t" ] || continue
grep -qxF "$t" "$RUNNER_TEMP/bg/leg" || continue
f="$RUNNER_TEMP/bg/$(basename "$t")"
: >"$f.log"
# shellcheck disable=SC2016 # $1 and $2 expand in the child shell, on purpose
nohup bash -c 'timeout 900 bash -e "$1" >"$2.log" 2>&1; echo $? >"$2.rc.tmp"; date +%s >"$2.end"; mv "$2.rc.tmp" "$2.rc"' _ "$t" "$f" >/dev/null 2>&1 &
done
# Per-PR tests: a PR registers its test by adding scripts/ci-tests.d/<slug>.sh (one command line),
# not by appending a line to the step above, so PRs never conflict on this file.
# pyyaml: scripts/test_workflow_persist_credentials.py fails closed without it.
- name: Install hypothesis for the property tests
run: python3 -m pip install hypothesis==6.170.12 pyyaml==6.0.3
- name: Per-PR tests (scripts/ci-tests.d)
run: |
set -euo pipefail
[ -s "$RUNNER_TEMP/bg/leg" ] || { echo "FAIL: the start step wrote no test list for this leg"; exit 1; } # else grep fails and every test is skipped
for t in scripts/ci-tests.d/*.sh; do
case "$t" in *.bg.sh) continue ;; esac
grep -qxF "$t" "$RUNNER_TEMP/bg/leg" || continue # another leg runs it
echo "== $t"
bash -e "$t"
done
for t in scripts/ci-tests.d/*.bg.sh; do
[ -e "$t" ] || continue
grep -qxF "$t" "$RUNNER_TEMP/bg/leg" || continue
f="$RUNNER_TEMP/bg/$(basename "$t")"
echo "== $t (started in the background)"
[ -f "$f.log" ] || { echo "FAIL: $t was never started"; exit 1; }
end=$(($(cat "$RUNNER_TEMP/bg/started") + ${BG_WAIT_S:-960})) # counted from the background start, not from here
while [ ! -f "$f.rc" ] && [ "$(date +%s)" -lt "$end" ]; do sleep 1; done
cat "$f.log"
took='?'; [ -f "$f.end" ] && took=$(($(cat "$f.end") - $(cat "$RUNNER_TEMP/bg/started")))
echo "bg elapsed: ${f##*/} $took s"
[ -f "$f.rc" ] || { echo "FAIL: $t did not finish in time"; exit 1; }
[ "$(cat "$f.rc")" = 0 ]
done
evals:
runs-on: ubuntu-latest
timeout-minutes: 35
steps:
# SHA-pinned, not a mutable tag (the skill's own K rule for CI actions).
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 0 # fix-class gate diffs base..head; a depth-1 clone cannot resolve the range
persist-credentials: false # no later step pushes or fetches a private ref; GH_TOKEN env covers gh
# Derived files (SHA256SUMS, size-budgets.tsv rows, INDEX.md, INDEX-scripts.md) are not committed by lane PRs, so two
# PRs never conflict on them. Rebuild them here, before any gate reads them; a land commit (VERSION
# moved) is skipped, so there the committed files are checked byte-for-byte.
- name: Regenerate derived files (unlanded PR)
if: github.event_name == 'pull_request'
run: bash scripts/regen-unlanded.sh origin/main
# Live harness (stub HTTP server, no key, no spend) and the review-scoring
# instrument (canaries, FP budget, fixture integrity).
- name: Live eval harness and review scorer tests (offline)
run: |
python3 scripts/test_live_evals.py
python3 scripts/test_score_review.py
python3 scripts/test_issue1418.py
python3 scripts/test_install_commands.py
python3 scripts/test_spend_report.py
python3 scripts/test_field_learnings.py
python3 scripts/test_terse_scope_checks.py
python3 scripts/test_reply_budget_prevention.py
python3 scripts/test_handback_copypaste_exempt.py
python3 scripts/test_lane_guard_banlist_seed.py
python3 scripts/test_lane_guard_hotfiles.py
python3 scripts/test_sandbox_preflight.py
python3 scripts/test_learning_to_pr.py
python3 scripts/test_wire_defaults.py
python3 scripts/test_weekly_receipt.py
python3 scripts/test_token_ratchet.py
python3 scripts/test_destructive_fp.py
python3 scripts/test_repeat_poll_and_brief_lint.py
python3 scripts/test_token_report_rollup.py
python3 scripts/test_tokens_per_pr_parity.py
python3 scripts/test_bench_corpus.py
python3 scripts/test_bench_run_isolation.py
python3 scripts/test_resource_policy.py
python3 scripts/test_council.py
python3 scripts/test_veche.py
python3 scripts/test_veche_timebox.py
python3 scripts/test_roadmap_rank.py
python3 scripts/test_policy_ux.py
bash scripts/test-resource-policy-train.sh
python3 scripts/test_review_feedback.py
python3 scripts/test_learned_rules.py
python3 scripts/test_review_lean.py
bash scripts/test-post-review.sh
python3 scripts/test_eval_predicates_defended.py
python3 scripts/test_perun_doctor.py
python3 scripts/test_install_ux.py
python3 scripts/test_extra_hosts_install.py
python3 scripts/test_hosts_parity.py
python3 scripts/test_host_safety_rows.py
python3 scripts/test_auto_update.py
python3 scripts/test_auto_update_global.py
python3 scripts/test_auto_update_pr.py
python3 scripts/test_fleet_hooks.py
python3 scripts/test_field_mos_1009.py
bash scripts/test-land-release-tag.sh
bash scripts/test-host-manifests.sh
bash scripts/test-land-release-publish.sh
python3 scripts/test_scan_tracked_only.py
python3 scripts/test_review_gate.py
python3 scripts/test_max_pace_no_waste.py
cheap:
runs-on: ubuntu-latest
timeout-minutes: 35
steps:
# SHA-pinned, not a mutable tag (the skill's own K rule for CI actions).
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 0 # fix-class gate diffs base..head; a depth-1 clone cannot resolve the range
persist-credentials: false # no later step pushes or fetches a private ref; GH_TOKEN env covers gh
# Derived files (SHA256SUMS, size-budgets.tsv rows, INDEX.md, INDEX-scripts.md) are not committed by lane PRs, so two
# PRs never conflict on them. Rebuild them here, before any gate reads them; a land commit (VERSION
# moved) is skipped, so there the committed files are checked byte-for-byte.
- name: Regenerate derived files (unlanded PR)
if: github.event_name == 'pull_request'
run: bash scripts/regen-unlanded.sh origin/main
- name: review_checks self-test
run: bash scripts/test-review-checks.sh
- name: mktemp TMPDIR regression
run: bash scripts/test-mktemp-tmpdir.sh
# Shipped gate scripts prove they FIRE on planted violations before any
# result of theirs is trusted (independent oracle per script).
# Strict: a new script nothing calls, or that no registered CI test runs, was built but never wired (field learning, #1380).
- name: New scripts are wired and run
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
run: |
cat scripts/ci-tests.d/*.sh scripts/test_*.py scripts/test-*.sh > "$RUNNER_TEMP/runs.txt"
python3 .claude/skills/agentic-delivery/scripts/wired_check.py "$BASE_SHA" "$HEAD_SHA" --strict --runs "$RUNNER_TEMP/runs.txt"
- name: Shipped script self-tests
run: |
set -euo pipefail
python3 .claude/skills/deep-code-review/scripts/parity_differ.py --selftest
python3 .claude/skills/deep-code-review/scripts/token_differ.py --selftest
python3 .claude/skills/deep-code-review/scripts/fix_class_gate.py --selftest
python3 .claude/skills/deep-code-review/scripts/closes_lint.py --selftest
python3 .claude/skills/deep-code-review/scripts/binaries_gate.py --selftest
python3 .claude/skills/deep-code-review/scripts/reaper_lint.py --selftest
python3 .claude/skills/deep-code-review/scripts/source_scan_tests.py --selftest
python3 .claude/skills/deep-code-review/scripts/ci_cost_lint.py --selftest
bash .claude/skills/deep-code-review/templates/integrator-gate.sh --selftest
python3 scripts/eval_citation_lint.py --selftest
python3 .claude/skills/agentic-delivery/scripts/serial_gate.py --selftest
python3 .claude/skills/agentic-delivery/scripts/handback_cap.py --selftest
python3 .claude/skills/agentic-delivery/scripts/lane_guard.py --selftest
python3 scripts/test_done_means_verified.py
python3 .claude/skills/agentic-delivery/scripts/clock_lint.py --selftest
python3 scripts/test_clock_lint.py
python3 scripts/test_clock_lint_bombs.py
python3 .claude/skills/agentic-delivery/scripts/env_lint.py --selftest
python3 scripts/test_env_lint.py
python3 scripts/test_vacuous_static.py
python3 scripts/test_mutation_changed.py
bash scripts/test-peer-pains-4.sh
python3 .claude/skills/agentic-delivery/scripts/board_common.py --selftest
python3 .claude/skills/agentic-delivery/scripts/board_sync.py --selftest
python3 .claude/skills/agentic-delivery/scripts/board_post.py --selftest
python3 .claude/skills/agentic-delivery/scripts/board_state.py --selftest
python3 .claude/skills/agentic-delivery/scripts/claim_probe.py --selftest
python3 .claude/skills/agentic-delivery/scripts/refix_gate.py --selftest
python3 .claude/skills/agentic-delivery/scripts/priority_gate.py --selftest
python3 .claude/skills/agentic-delivery/scripts/queue_guard.py --selftest
python3 .claude/skills/agentic-delivery/scripts/focus_gate.py --selftest
python3 .claude/skills/agentic-delivery/scripts/surface_check.py --selftest
python3 .claude/skills/agentic-delivery/scripts/restore_ledger.py --selftest
python3 .claude/skills/agentic-delivery/scripts/lane_liveness.py --selftest
python3 .claude/skills/agentic-delivery/scripts/review_digest.py --selftest
python3 .claude/skills/agentic-ceo/scripts/task_ledger.py --selftest
python3 .claude/skills/agentic-delivery/scripts/union_overlap.py --selftest
python3 .claude/skills/agentic-ceo/scripts/stop_reminder.py --selftest
python3 .claude/skills/agentic-ceo/scripts/token_report.py --selftest
python3 .claude/skills/agentic-delivery/scripts/feedback_ledger.py --selftest
python3 .claude/skills/agentic-delivery/scripts/merge_train.py --selftest
python3 .claude/skills/agentic-delivery/scripts/tier_gate.py --selftest
python3 .claude/skills/agentic-delivery/scripts/lane_cap.py --selftest
python3 .claude/skills/agentic-delivery/scripts/escaped_defects.py --selftest
python3 .claude/skills/agentic-delivery/scripts/destructive_guard.py --selftest
python3 .claude/skills/agentic-delivery/scripts/outcome_claim_guard.py --selftest
python3 scripts/trigger_lint.py --selftest
# scripts/skill_index.py generates the per-skill and repo-level INDEX.md
# lookup tables (which file answers which trigger, without opening every
# references/*.md or scripts/*.py blindly). Self-test proves the parser
# on synthetic fixtures (never the real repo's prose); --check proves the
# committed INDEX.md files are byte-identical to what the generator
# would write today, so a routed reference added without regenerating
# goes stale loudly instead of silently.
- name: Skill index (INDEX.md lookup tables) selftest and freshness
run: |
set -euo pipefail
python3 scripts/skill_index.py --selftest
python3 scripts/skill_index.py --check
# A "`<file>.md` **<Section>**" pointer must resolve in that file, so a split or
# rename of a reference cannot silently orphan the pointers to it.
- name: Skill section pointers resolve
run: |
set -euo pipefail
python3 scripts/check_section_pointers.py --selftest
python3 scripts/check_section_pointers.py
# A doc number bound to a source command with `<!-- derived: CMD -->N`
# must equal the command's output, so a stale count fails here.
- name: Derived doc numbers match their source commands
run: |
set -euo pipefail
python3 scripts/test_derived_numbers.py
python3 .claude/skills/deep-code-review/scripts/derived_numbers.py .
# Every fix(...) commit in the pushed/PR range must touch a test surface
# or carry a non-empty No-Test-Reason: trailer. Fails closed on an
# unresolvable range (e.g. the all-zeros before-SHA of a new branch).
# A shipped script is deliberately NOT its own test surface: editing the
# script under test would satisfy the gate by itself (fail-open). Known
# limitation: touching ANY skill's evals/evals.json satisfies it, not
# only the fixed skill's.
- name: Fix commits carry a pinned test
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
run: |
set -euo pipefail
python3 .claude/skills/deep-code-review/scripts/fix_class_gate.py \
--base "$BASE_SHA" --head "$HEAD_SHA" \
--test-glob '.claude/skills/*/evals/evals.json' \
--test-glob 'scripts/test-ci-gates.sh' \
--test-glob 'scripts/test_*.py' \
--test-glob 'scripts/test-mktemp-tmpdir.sh'
# Review tiers (agentic-delivery/templates/review-tiers.tsv): a tier-1
# commit needs `Model: opus` plus an independent, receipted
# `Reviewed-By: <id> receipt:<sha|path|url>` trailer. ADVISORY here: it
# prints every finding (merge commits on the first-parent line
# included) and never fails the build, because this repo's history
# predates the trailers and one shared account cannot prove reviewer
# independence. Release squash commits carry both trailers.
- name: Review tiers (advisory)
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
run: python3 .claude/skills/agentic-delivery/scripts/tier_gate.py --advisory "$BASE_SHA..$HEAD_SHA"
# Lesson -> mechanism: a commit that edits skill prose (a SKILL.md or a
# references/*.md) must also land an executable mechanism (an eval, a
# shipped script, or a gate self-test case) or carry a non-empty
# No-Mechanism-Reason: trailer, so a filed lesson cannot stay prose that
# never fires. Same gate implementation, range, and fail-closed exits as
# the fix-class step above, except that merge commits are judged by their
# own changes and a SKILL.md version-stamp-only bump is not a trigger.
# Here a shipped script DOES count: in this mode the script is the
# mechanism being added, not the code under test.
# Known limitation: it proves a mechanism path was touched, not that the
# mechanism pins the lesson (review judges that).
- name: Prose lessons carry a mechanism
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
run: |
set -euo pipefail
python3 .claude/skills/deep-code-review/scripts/fix_class_gate.py \
--base "$BASE_SHA" --head "$HEAD_SHA" \
--trigger-glob '.claude/skills/*/references/*.md' \
--trigger-glob '.claude/skills/*/SKILL.md' \
--test-glob '.claude/skills/*/evals/evals.json' \
--test-glob '.claude/skills/*/scripts/*' \
--test-glob 'scripts/test-ci-gates.sh' \
--test-glob 'scripts/test-*.sh' \
--test-glob 'scripts/test_*.py'
# Lesson replay (ADVISORY signal, not proof): the step above proves a
# mechanism path was touched; this one looks for a tie between the
# mechanism and the lesson. Each prose-lesson commit prints REPLAYS (a
# changed eval case names a term the lesson newly added, or reverting a
# changed script hunk fails an assertion in its selftest), VACUOUS (no
# tie found), or COULD_NOT_CHECK with a reason. Token overlap is
# gameable; a reviewer judges every line. It exits 0 on VACUOUS and
# fails closed (exit 2) only on an unresolvable range, like the gate
# above. Adding `--gate` makes VACUOUS (only VACUOUS) exit 1;
# COULD_NOT_CHECK never blocks.
- name: Lesson mechanisms replay (advisory)
timeout-minutes: 15
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
HEAD_SHA: ${{ github.event.pull_request.head.sha || github.sha }}
run: |
set -euo pipefail
python3 scripts/lesson_replay.py --selftest
python3 scripts/lesson_replay.py --base "$BASE_SHA" --head "$HEAD_SHA"
# A size-budget row may only rise with a `size-budget-raise:` marker
# (commit range or CHANGELOG) — budgets ratchet down by default.
- name: Size budgets ratchet down
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
run: bash scripts/ci-gates.sh size-ratchet --base "$BASE_SHA" --config scripts/size-budgets.tsv .
# The rule-enforcement baseline may only grow with a
# `rule-baseline-raise: <reason>` commit-message line.
- name: Rule baseline ratchets down
env:
BASE_SHA: ${{ github.event.pull_request.base.sha || github.event.before }}
run: bash scripts/ci-gates.sh rules --baseline scripts/rule-enforcement-baseline.tsv --base "$BASE_SHA" .
- name: SKILL.md name matches its directory
run: |
set -euo pipefail
# Every shipped skill's frontmatter name must equal its directory (not
# just deep-code-review) — an overlay cannot ship with a mismatched
# name that the routing/version gates would miss.
for d in .claude/skills/*/; do
skill="$(basename "$d")"
grep -qE "^name: ${skill}\$" "${d}SKILL.md" \
|| { echo "name: does not match directory for ${skill}"; exit 1; }
done
- name: install.sh parses
run: bash -n install.sh
# Execute the fabrication-refusal evals offline: each deterministic
# predicate must separate a fabricated answer (red fixture) from a refusal
# (good fixture). No model, no network, no spend — the discrimination is
# the property under test. The same gate the self-tests pin above.
- name: Fabrication-refusal predicates discriminate (offline, no model)
run: python3 scripts/eval_predicates.py --selftest
# Prove the split-rubric eval RUNNER's own guards, offline: dry-run
# enumerates + classifies every eval (hard = a deterministic predicate is
# bound; soft = needs the live judge), the decorrelation guard aborts on an
# equal-or-missing model id, and the live path refuses without configuration
# (its model-call site is an un-built stub). No model, no network, no spend —
# the guards are the property under test. The live judge itself is owner-gated.
- name: Eval runner guards hold (offline, no model)
run: python3 scripts/run-evals.py --selftest
# Guards .gitleaks.toml: the fixture allowlist must not blind the scanner (full history needed for the history scan).
gitleaks:
runs-on: ubuntu-latest
timeout-minutes: 10
permissions:
contents: read
steps:
- uses: actions/checkout@3d3c42e5aac5ba805825da76410c181273ba90b1 # v7.0.1
with:
fetch-depth: 0
persist-credentials: false
- name: Install gitleaks 8.30.1 (pinned sha256)
run: |
set -euo pipefail
cd "$RUNNER_TEMP"
curl -fsSL -O https://github.com/gitleaks/gitleaks/releases/download/v8.30.1/gitleaks_8.30.1_linux_x64.tar.gz
echo "551f6fc83ea457d62a0d98237cbad105af8d557003051f41f3e7ca7b3f2470eb gitleaks_8.30.1_linux_x64.tar.gz" | sha256sum -c -
mkdir -p "$RUNNER_TEMP/bin"
tar -xzf gitleaks_8.30.1_linux_x64.tar.gz -C "$RUNNER_TEMP/bin" gitleaks
echo "$RUNNER_TEMP/bin" >> "$GITHUB_PATH"
- name: Allowlist guard
run: REQUIRE_GITLEAKS=1 bash scripts/test_gitleaks_allowlist.sh
# Branch protection requires a check named exactly `gates`: it passes only when every shard job succeeded.
gates:
needs: [selftests, enforce, perpr, evals, cheap, gitleaks]
if: always()
runs-on: ubuntu-latest
timeout-minutes: 5
steps:
- run: |
[ "${{ contains(needs.*.result, 'failure') || contains(needs.*.result, 'cancelled') || contains(needs.*.result, 'skipped') }}" = false ]