Local CI mode

September 16, 2026 · View on GitHub

Run the same review pipeline that .github/workflows/review.yml + maintenance.yml + auto-fix-pr.yml run in GH-Actions — locally, without consuming GH-Actions minutes. Two deliverables, both additive (the existing workflows are unchanged):

bin/review-local.sh — local equivalent of the review + maintenance workflow orchestration.
/dev-kit:babysit-pr-local — 0-arg babysit loop driven by the local judge, with an always-on pre-push test gate.

When to use

The repo has hit its GH-Actions minute cap on private plans.
The operator wants to iterate on a review verdict locally (faster feedback loop than waiting for CI to spin up).
The operator is testing a provider switch (bin/set-provider.sh <provider>) or a new --local-test-cmd ahead of pushing.

When NOT to use

The branch's PR requires a reviewer bot or org-level MCP that's only available via anthropics/claude-code-action. Local claude -p does not have access to the same MCP servers; inline comments go through gh pr comment instead of mcp__github_inline_comment__create_inline_comment.
The PR requires gh pr merge — merging is always a human action. bin/review-local.sh's --auto-approve casts gh pr review --approve only; the operator merges the PR themselves.

Local review: `bin/review-local.sh`

A direct shell port of the orchestration half of .github/workflows/review.yml + maintenance.yml. The LLM-judge skills (/dev-kit:review, /dev-kit:security, /dev-kit:maintenance) are reused verbatim via the local claude -p invocation.

Usage

# Dry-run (no LLM call, no PR mutation): preview env + planned commands.
bin/review-local.sh --pr 123 --dry-run

# Full review + auto-approve on clean verdict.
bin/review-local.sh --pr 123 --auto-approve

# Force a specific provider (overrides .env:CI_REVIEW_PROVIDER).
bin/review-local.sh --pr 123 --provider anthropic --auto-approve

# Run only /dev-kit:review (skip security + maintenance).
bin/review-local.sh --pr 123 --review-only

# Force-anthropic, only security, no auto-approve (dry-run).
bin/review-local.sh --pr 123 --security-only --provider anthropic --dry-run

Slash command

/dev-kit:review-local --pr 123 --auto-approve

The slash command is a thin wrapper over bin/review-local.sh. Both paths apply the same provider switching + gate logic.

Dynamic gate skip (v1.1.0)

--dynamic-skip opt-in flag that runs lib/gate_dynamic.select_gates BEFORE invoking the LLM judges. The judge inspects the diff + previous verdicts + the gate catalog and pre-flips the per-judge RUN_<NAME> booleans for any gate it recommends skipping. The 4 hard rules (first-push-deterministic, forced_run, critical-gate-in-scope, low-confidence-veto) are applied AFTER the LLM call so the judge can't override them.

bin/review-local.sh --pr 123 --dynamic-skip

Propagate from bin/babysit-pr-local.sh via BABYSIT_DYNAMIC_SKIP=1 (opt-in; default OFF). The audit trail lives at .dev-kit/gate-dynamic/<head_sha>.json. See docs/skills/gate-dynamic.md for the full reference.

Provider setup

The script reads CI_REVIEW_PROVIDER from the process env, then .env (matches bin/set-provider.sh resolution). Switch via:

bin/set-provider.sh anthropic
bin/set-provider.sh deepseek
bin/set-provider.sh minimax

The matching *_API_KEY must be in .env (or the process env):

# .env (gitignored)
ANTHROPIC_API_KEY=sk-ant-...
MINIMAX_API_KEY=sk-cp-...
DEEPSEEK_API_KEY=sk-...

bin/review-local.sh reads the key via lib/ci_setup.read_env_key() and passes the same ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY / ANTHROPIC_MODEL* block as review.yml:120-131 to the claude -p invocation via the env KEY=... claude -p ... prefix. The key never enters the parent shell's persistent environment, so subsequent gh / shell calls cannot leak it via /proc/<pid>/environ or core dumps.

No provider configured? It still works.

bin/set-provider.sh + a *_API_KEY are for CI runners (GH-Actions has no interactive login, so it needs an explicit key injected). A local interactive session almost always already has an authenticated claude CLI (a claude login session or a keychain-stored key) — the script does not require .env, CI_REVIEW_PROVIDER, or any *_API_KEY to be set at all.

If neither --provider, CI_REVIEW_PROVIDER (env or a real .env), nor a matching *_API_KEY is found, the script logs:

no provider explicitly configured and no API key found; falling back to local claude CLI auth (no key/base-url injection)

and calls claude -p ... without any ANTHROPIC_BASE_URL / ANTHROPIC_API_KEY / ANTHROPIC_AUTH_TOKEN override — claude -p inherits the parent shell's existing auth exactly as if you'd run it by hand. This is the common case for a local bin/babysit-pr-local.sh run; explicit provider configuration is only needed when you want a specific non-default provider (e.g. testing minimax while your claude CLI's own login points at anthropic).

An explicit ask (--provider X, or CI_REVIEW_PROVIDER set) with no matching key still fails loudly — that's a real misconfiguration, not the "just use my session" case.

What it does (mirrors `review.yml` step-by-step)

Step	Source step name	Local equivalent
Resolve provider	review.yml `Resolve PR + provider` step	`lib/ci_setup.read_provider()`
Set ANTHROPIC_* env	review.yml `Run /dev-kit:review via …` env block	`provider_env_for "$PROVIDER"` (sourced from `lib/review_local_lib.sh`)
Run `/dev-kit:review`	review.yml `Run /dev-kit:review via …` step	`claude -p "/dev-kit:review --diff ..."`
Run `/dev-kit:security`	review.yml `Run /dev-kit:security via …` step	`claude -p "/dev-kit:security --diff ..."`
Run `/dev-kit:maintenance`	maintenance.yml `Run /dev-kit:maintenance via …` step	`claude -p "/dev-kit:maintenance --diff ..."`
Extract verdict	review.yml `Extract <skill> verdict` step	capture `claude -p` stdout per skill, pipe to `python3 -m lib.maintenance_gate --extract-verdict-from-stdin`
Bump-PR skip	review.yml job-level `if:` filter	`is_bump_pr "$PR_TITLE"` (sourced from `lib/review_local_lib.sh`)
Combined verdict gate	review.yml `Combined verdict gate` step	`rank()` (from `lib/review_local_lib.sh`) + worst-of wins
L3-evidence gate	review.yml `L3 evidence gate (PR body must quote test count)` step	`extract_pytest_tail "$PR_BODY"` (from `lib/review_local_lib.sh`)
Auto-approve	review.yml `Auto-approve on clean verdict` step	`gh pr review --approve --body "..."` (only with `--auto-approve`)
Audit comment	review.yml `Extract <skill> verdict` audit line	`gh pr comment --body "<!-- dev-kit-verdict-audit --> ..."`

Caveats

No MCP inline-comment server: the workflow has access to mcp__github_inline_comment__create_inline_comment via claude-code-action. Local claude -p does not. The skill body falls back to gh pr comment (per skills/review/SKILL.md).
Local API key exposure: the script scopes the key to the claude -p invocation only (via env KEY=... claude -p ...). It does NOT enter the parent shell's persistent env. Do NOT run it on a shared host regardless -- the agent still processes PR content with operator credentials.
Cannot gh pr merge: the script never merges. The operator runs gh pr merge manually after --auto-approve lands.
No provider fallback: --provider is strict; an unknown provider exits 1. The script does not auto-switch to minimax. This matches bin/set-provider.sh behavior.

Local babysit: the pre-push test gate

/dev-kit:babysit-pr already runs locally (the skill body lives in the current shell), and its §Algorithm step 8 re-runs the specific failing check before every commit. What it deliberately does not offer is a broad pre-push run of the whole suite: that is the defining behavior of /dev-kit:babysit-pr-local, where it is always on.

A --local-verify opt-in used to exist on /dev-kit:babysit-pr. It was removed — it gave two ways to ask for one behavior, and the operator had to remember which skill carried which half of the gate. If you want the broad gate, run the local-mode skill:

# Runs pytest -q inside the worktree before every commit + push.
/dev-kit:babysit-pr-local

What the gate does

/dev-kit:babysit-pr-local's §Algorithm step 7.5 sits between APPLY FIX (step 7) and VERIFY LOCAL (step 8):

7.5. LOCAL VERIFY (always on in local mode)
     - lib.babysit_pr_cli.run_local_verify(cmd=--local-test-cmd,
                                          cwd=<worktree>)
       executes the command via `bash -c "$cmd"` and returns a
       LocalVerifyResult. The iteration proceeds only when
       passed=True AND tail_line is the quoted pytest tail line.
     - non-zero exit OR missing tail line OR timeout -> abort iteration
       BEFORE git add / commit / push (MUST-L3 enforcement).

Step 8 (re-run the specific failing check) is preserved in both skills; step 7.5 is the broader pre-commit check, not a replacement.

Why this matters

Without the gate, the babysit loop's typical flow is:

fix → git add → git commit → git push → wait for GH-Actions CI

A known-failing local test consumes one GH-Actions run per iteration. With the gate:

fix → pytest -q (LOCAL) → fix re-iteration → ... → git add → git commit → git push

Failing iterations abort before the push, so no GH-Actions run is consumed until the iteration actually passes the gate. The user still verifies locally, but the GH-Actions budget is preserved for genuinely green PRs.

Implementation

Helper: lib/babysit_pr_cli.py::run_local_verify() runs the command and parses the MUST-L3 tail line. run_babysit_once() is unchanged (the helper stays pure).
Parser: --local-test-cmd is a hidden (argparse.SUPPRESS) override for non-pytest projects; there is no --local-verify flag.
Skill body: skills/babysit-pr-local/SKILL.md §Algorithm step 7.5. The Bash invocation lives in the orchestrator script, not in lib/.
Tests: tests/test_babysit_pr_cli.py::TestParseBabysitArgs pins that --local-verify is rejected and --local-test-cmd still parses; TestRunLocalVerify pins the gate itself.

Caveats

Local test suite must be sane: the gate trusts the local test result. If the local test suite is itself broken or stale (e.g. missing fixture), the gate refuses to push. Operators should run pytest -q once by hand to confirm the local baseline before relying on the gate.
MUST-L3 is enforced by the skill body, not by the helper: if the test command exits 0 but its stdout lacks a pytest-style tail line, the skill refuses to flip to "ready to push". The operator must either pick a test command that emits the tail line or paste the evidence manually.
No fallback to GH-Actions: refusing to push means the iteration aborts. The operator can re-run /dev-kit:babysit-pr to fall back to the default push-and-wait-CI flow.

`/dev-kit:babysit-pr-local` — local-mode babysit (additive sibling)

When GH-Actions minutes are exhausted AND the operator wants the same iterative repair loop as /dev-kit:babysit-pr, but driven entirely by the local LLM-judge verdict instead of gh pr checks --watch, run this skill:

# No flags exposed to operators. Slash invocation is always 0-arg.
/dev-kit:babysit-pr-local   # babysit current branch's PR

The skill is a 0-arg orchestrator. Behind the scenes, the §Algorithm step 4L invokes bin/babysit-pr-local.sh <PR> (refuses --auto-appearing), which execs bin/review-local.sh --pr <PR> to run /dev-kit:review + /dev-kit:security + /dev-kit:maintenance locally. The local audit comment ( ... verdict=Approve source=bin_review_local) is the iteration's "green" signal; the local judge's stdout line (combined verdict: <Word>) is the MUST-L3 evidence quote.

Why a separate skill instead of a flag

The user-facing UX contract — "operate only via skill, no options to remember" — drives the split:

	`/dev-kit:babysit-pr`	`/dev-kit:babysit-pr-local`
Review verdict source	GH-Actions CI	local `bin/review-local.sh`
`gh pr checks --watch`	yes	no
`--auto-approve` semantically possible	yes (the operator may pass it through babysit)	forbidden (refused with exit 2)
Broad pre-push test gate	no (step 8 re-runs the failing check only)	always on

The two skills are additive siblings — they share the lock-file protocol + lib/babysit_pr_cli helpers + worktree-detect plumbing, but the SKILL.md §Algorithm differs in five steps (see skills/babysit-pr-local/SKILL.md §"Step diff vs /dev-kit:babysit-pr"). Operators who want local mode type exactly one command; operators who want CI mode keep using /dev-kit:babysit-pr unchanged.

Hidden flags (not in the slash description; power users + tests)

--pr N — babysit explicit PR number instead of current-branch PR discovery. Use this when the PR was opened from another worktree (e.g. git worktree add .worktrees/<branch> then push from there).
--local-test-cmd CMD — override the pre-push pytest default (pytest -q) with a project-specific runner. The command's stdout+stderr MUST emit a pytest-style tail line (<N> passed in <Ns>s or <N> failed in <Ns>s) per MUST-L3. For Make / tox / nox / Go / JS projects, supply the runner that emits that shape; otherwise the iteration refuses to push.
--local-mode — internal routing flag (already implied by the slash invocation; kept so the parser doesn't double-parse).

None appear in --help, description, or argument-hint. Operators always run /dev-kit:babysit-pr-local with no arguments.

Lock file

<worktree>/.dev-kit/babysit.lock — shared with /dev-kit:babysit-pr. If both skills race on the same worktree, the second arrival sees a fresh lock and refuses with already running. The lock body appends source=babysit-pr-local so a post-mortem can tell which skill held the lock. The stale-lock TTL is the same 30 minutes (lib/babysit_pr_reliability.LOCK_TTL_SECONDS).

Implementation

bin/babysit-pr-local.sh — single-call wrapper (≈30 lines) that validates args (refuses --auto-appearing) and execs bin/review-local.sh --pr $PR_NUMBER.
lib/babysit_pr_cli.py::is_local_mode(argv) + the hidden --local-mode argparse field — routing helpers (the parser's flag is suppressed from --help for L5 compliance).
skills/babysit-pr-local/SKILL.md — the algorithm body (≈365 lines); §Algorithm mirrors /dev-kit:babysit-pr with five surgical substitutions (steps 3 / 4 / 4L / 7.5 / 11 / 12 in the parent).
skills/babysit-pr-local/recipes/canonical-wiring.md — parent preflight block + sub-agent prompt body.
skills/babysit-pr-local/SKILL.md — slash command description (mirrors skills/review-local/SKILL.md).

bin/review-local.sh — local equivalent of the GH-Actions review workflow.
bin/babysit-pr-local.sh — local-mode babysit wrapper (executable; refuses --auto-appearing).
skills/review-local/SKILL.md — slash command wrapper (one-shot local review).
skills/babysit-pr-local/SKILL.md — slash command wrapper (local-mode babysit).
skills/babysit-pr/SKILL.md — babysit-pr skill (GH-Actions-driven; no local-test flag).
skills/babysit-pr-local/SKILL.md — local-mode babysit skill (additive sibling; replaces gh pr checks --watch with bin/review-local.sh).
skills/babysit-pr-local/recipes/canonical-wiring.md — local-mode sub-agent prompt + parent preflight.
lib/maintenance_gate.py — verdict-extraction + combined-gate helper.
lib/ci_setup.py — provider resolution + secret name lookup.
lib/babysit_pr_cli.py — is_local_mode, parse_babysit_args, run_local_verify, run_babysit_once.
bin/set-provider.sh — local provider switch (bin/set-provider.sh anthropic).
.github/workflows/review.yml — GH-Actions equivalent (unchanged).
.github/workflows/maintenance.yml — GH-Actions equivalent (unchanged).
scripts/ci-local.sh — pre-existing local validator runner (no LLM review).
tests/test_review_local_sh.py — shell-level tests for bin/review-local.sh.
tests/test_review_local_lib.py — unit tests for lib/review_local_lib.sh.
tests/test_babysit_pr_cli.py — parser + orchestrator tests for babysit-pr.
tests/test_babysit_pr_local_cli.py — parser + is_local_mode tests for babysit-pr-local.
tests/test_babysit_pr_local_sh.py — shell-level tests for bin/babysit-pr-local.sh.
tests/test_commands_install.py — slash-command install governance.