"HARD RULE for developing the Kiro Crew source repo ITSELF (not users' projects): develop, build and verify in a git worktree, never against the live gateway. Covers env setup, main-clone sync, build/dist ordering, flags, previews, cleanup, publish authorization; kirocrew-prepare-pr owns PRs."
Resources
1Install
npx skillscat add kirodotdev/kirocrew/kirocrew-worktree-dev Install via the SkillsCat registry.
Kiro Crew worktree development
This skill applies ONLY to the Kiro Crew source repository or a worktree of it.
Ignore it for other projects. One feature uses one dedicated worktree containing
both src/kiro_crew/ and website/, including backend-only changes.
Never edit the live checkout or start/stop the live gateway from a feature session.
Worktree and environment
# From the main clone; skip creation when already in the assigned worktree.
git fetch origin main
git worktree add ../kirocrew-wt-<name> -b feat/<name> origin/main
# A pod worktree never goes under $KIROCREW_SCRATCH: pod_up can't see it.
cd ../kirocrew-wt-<name>
python3 -m venv .venv
.venv/bin/pip install -e ".[voice]" --group dev
cd website && npm ci && cd ..
git worktree listTo bring the main clone itself current (the first path git worktree list
prints), fast-forward it and nothing else:
git -C <main clone> symbolic-ref --short HEAD # must print main
git -C <main clone> fetch origin main
git -C <main clone> merge --ff-only --no-autostash --no-overwrite-ignore origin/mainIt refuses, changing nothing, when local main has diverged or a local change
would be overwritten; then stop and report, because that state belongs to whoever
left it. Before any rebase, checkout, switch, merge, pull or reset in a clone,
run python3 <this skill>/scripts/live_checkout_guard.py <clone>. Exit 30 (or 2):
a running gateway runs from it, so make a worktree; the operator's dashboard
Update syncs that clone. Never overlay a whole tree with git checkout <ref> -- .
or git restore --source <ref> .: neither moves HEAD, both overwrite local edits
without asking, and the checkout form keeps every file upstream deleted. Nevergit stash: every worktree shares one stash list, so a pop can apply another
session's work.
Working from several worktrees? If uv resolves in your shell (command -v uv
— it is a declared dependency of the package, so any activated Kiro Crew venv
has it on PATH; a bare shell may not), build the venv with it instead: same
deps, but site-packages are copy-on-write clones out of one global cache, so on a
reflink filesystem (XFS with reflink, btrfs, APFS) each extra venv costs ~10 MB
of unique disk and ~10 s instead of ~400 MB and ~1 min; without reflink uv copies,
so you keep the speed and lose the disk saving. kirocrew pod provision runs
the same two steps (without the [voice] extra) when KIROCREW_PROVISION_USE_UV=1 is set
(opt-in for now; design record: the "Shared Dependency Cache for Worktrees" RFC
under docs/request-for-change):
uv venv --seed --allow-existing --link-mode clone --python 3.12 .venv
uv pip install --link-mode clone --python .venv/bin/python --project . -e ".[voice]" --group dev--link-mode clone must be explicit (uv's Linux default has been seen to copy
on a reflink-capable filesystem), --project . makes --group read this
worktree's pyproject.toml from any cwd, and --seed keeps .venv/bin/pip
present for make backend. Do not use --link-mode hardlink: it shares the
inode, so an in-place edit under .venv/lib/.../site-packages would land in every
sibling venv and in the cache. A clone is safe to edit; the write copies the block.
dev is a PEP 735 dependency group, not an extra. --group needs pip >= 25.1;
if unsupported, upgrade the worktree's pip with .venv/bin/pip install -U pip.
On Windows use .venv\Scripts\ in place of .venv/bin/.
Making a worktree live swaps code behind the same URL and REAL data home,
including DB and sessions. Only one can be live at a time. This is not preview
isolation: get explicit authorization for a live cutover or migration, and return
to the clean baseline afterwards. The default home is ~/.kiro/crew;KIROCREW_HOME overrides it. A legacy ~/.kirocrew does not auto-migrate.
Verification
Read .github/workflows/ci.yml AND .github/workflows/fast-gate.yml for the
canonical blocking gates; CI wins over prose. Never weaken or skip a check to
make it green. For test authoring, flakes, isolation or suite speed, loadwriting-tests and the owning section ofdocs/system-specs/common/testing-conventions.md before fixing the test.
During iteration, run python3 scripts/local-gate.py from the worktree. It
runs the tests RELATED to the diff on both surfaces with a bounded worker count
and never escalates to a full suite -- meta, mixed and large diffs all get a
related set; an unreadable diff exits 2 and runs nothing. The full suite is CI's
job; --full exists for a human who asks. --dry-run prints the plan,--base REF selects a base. Before publication, MUST load
kirocrew-prepare-pr and run its complete resolved gates[]
floor from the base-ref profile. That skill alone owns PR preparation, local
reviewers, history, dispositions and push-to-green steps.
Minimum manual checks, not a replacement for that floor:
python3 scripts/local-gate.py
isort --check-only src/kiro_crew test conftest.py xdist_budget.py
flake8 src/kiro_crew test conftest.py xdist_budget.py
mypy src/kiro_crew/
cd website
npx tsc -p tsconfig.app.json
npx vitest run
cd ..Rebuild and stage frontend dist BEFORE backend tests that read static assets.
A bare
python -m pytest(the full suite, for a human who wants it) usessetup.cfg's-n auto --dist loadgroupand--max-worker-restart=2. Keep
grouping forxdist_groupserialization. The root conftest budgets workers by
free memory and concurrent runs; explicit-n <N>bypasses it. Check host
headroom before a full suite. The gate scripts pass their own bounded-n.
A scoped subagent may use serial targeted tests; the parent owns aggregate gates.Coverage is off by default (
setup.cfgaddopts carries no--cov), so no--override-iniis needed to skip it. An override REPLACES the whole addopts
list; if you must use one for a multi-test run, keep--ignore=build/private -p no:platformdirs -n auto --dist loadgroup --max-worker-restart=2 --timeout=120. Never a bare--override-ini=addopts=for the multi-test gate.Capture logs under
$KIROCREW_SCRATCH; check the command's exit code, not a
pipe's last command. A greentaildoes not mean a green test.Reproduce a failure on a clean
origin/mainworktree before calling it
pre-existing or flaky. A branch-only failure is yours. Never fix a flake with
a retry, longer sleep, relaxed assertion or skip; usewriting-tests.To rank suspected flakes, start from the flake ledger: the open issues titled
Flaky: <test name> ..., labelledarea: tests
(gh issue list --state open --search "flaky in:title"), where a red confirmed as a
flake is recorded before any rerun (kirocrew-prepare-pr, Before you rerun a red
test). For a test not on it, mine CI rather than guessing from a local pass:gh run list --workflow=ci.yml --limit 250 --json databaseId,conclusion \ --jq '.[]|select(.conclusion=="failure")|.databaseId' > "$KIROCREW_SCRATCH/ci-ids"Read candidate runs with
gh run view <id> --log-failed, rank recurring test
names, then compare each against main. A ratchet failing on feature branches
may be a true positive. Check Windows timing/process causes as well as Linux.Match mypy's version to the dev-group pin in
pyproject.toml; install the same.[voice]plus--group devdependencies as CI. A venv withfaiss-cpu
makes imports typed where CI seesAny, so it is not a CI mirror. On macOS
runmypy --platform linux src/kiro_crew.
A reusable mypy-only environment can be provisioned from the repo root:
python3 -m venv ~/.kiro/crew/venvs/mypy-ci
~/.kiro/crew/venvs/mypy-ci/bin/pip install -e ".[voice]" --group dev
# Never add faiss. Run from the worktree, which supplies mypy's config.
~/.kiro/crew/venvs/mypy-ci/bin/mypy src/kiro_crew/The served frontend is built dist
The gateway serves src/kiro_crew/static/dist/, not source TSX or a Vite server.
After creating a worktree or changing frontend code, build and stage:
cd website && npm ci && npm run build && cd ..
PYTHONPATH=src python -m kiro_crew.frontend stage .Every Vite build publishes into website/dist atomically
(website/scripts/publish-dist.mjs). The stager points static/dist at it (the
dev link), or for an edition at a fresh private copy, and a running gateway
follows either on its next request. Never rm -rf + cp -R static/dist:
that deletes the tree a running gateway serves. make build builds, installs
and stages. Dist is gitignored and does not transfer with a fetch or new
worktree. To inspect a minified bundle, search surviving route,
API or label strings, not React component names.
Flags and isolated preview
Flags belong in the RUNNING instance's $KIROCREW_HOME/config.json, not in code.
The live gateway uses the shared home; dev-backend.sh uses .kirocrew-dev/;
each pod has its own home. Editing production config cannot enable a preview
flag. Check the right instance's flag before blaming the bundle. Config's live
fingerprint cache picks up edits without a restart, except fields markedrestart=True (the schema's requiresRestart), which need that instance
restarted; live flags persist across live-worktree switches.
Green gates are the floor. Preview is optional where unit coverage suffices;
when verifying changed UI, load web-verify and inspect the rendered change.
Prefer a disposable isolated pod, built first with venv and dist:
kirocrew pod up <worktree-name> --json
kirocrew pod up <worktree-name> --seed minimal --json
kirocrew pod down <worktree-name>kirocrew pod --help is authoritative. --provision runs the build on-ramp.--seed names shipped fixtures (pod scenarios lists them); an unknown name is
refused, while a path is the sanitized config-only form. Other boot flags:
--approval reads|yolo|interactivepersists the pod's mode; omission inheritsagent.approval_mode. No approval flag grants permission to publish or touch live.--cronsenables its scheduler; the default is--no-crons.--no-embeddingsavoids GGUF download and vectors in this pod only; search uses
keyword fallback, useful for ingestion tests. The real home's model is untouched.--ttlbounds the dashboard token (default2h). Re-up a stopped pod to change
boot flags.pod tokenre-mints a token;pod urlprints the base URL without it.
Pods need Linux systemd --user or macOS launchd; run kirocrew pod install
once per host. See pod platform requirements.
Every verb above talks to that per-user service manager, so an agent session
behind an outer sandbox with its own user namespace gets Permission denied from
all of them; systemctl --user is-system-running says which case you are in. Thepod_up / pod_restart / pod_down / pod_status / pod_ls tools do the same work through the
gateway, which holds the host bus, and pod_up returns the {base_url, token, port} handle directly. They need the Dev Fleet app enabled.pod ls, status, logs, provision, prune, exec and api cover inspection,
provisioning, removal of gone worktrees, in-pod commands and HTTP probes.
Without a service manager, or for a foreground debugger, run ./dev-backend.sh
from the worktree. It uses an isolated port/home and PYTHONPATH=src; Ctrl+C
stops it, and rerunning picks up code changes. Reclaim its process and dev home
yourself. Never use the live gateway instead.
A healthy pod proves only boot. For behavior proof, followdocs/guides/worktree-verification-recipes.md: select the owning endpoint from
the feature map, assert seeded state, run the pod-e2e Playwright harness (it
ships with the Dev Fleet app) and preserve screenshots. Treat commands marked forthcoming as absent until help
lists them, and a session-control HTTP 403 as a gap, not successful proof.
Agent/MCP specs are a separate isolation axis
KIROCREW_HOME isolates config, DB and sessions, NOT machine-wide~/.kiro/agents/*.json. An ephemeral gateway refuses to rewrite that shared
agent home so its MCP paths cannot outlive its venv. The warning is protection,
not a failure: preview uses the real install's specs and MCP servers. A preview
therefore does NOT exercise changes to mcp-core, mcp-cron or mcp-computer.
Use unit tests; any temporary change to the real spec requires explicit user
authorization and restoration afterwards.
Do not bypass this with KIRO_HOME: it also moves sessions, settings, skills
and steering, while remaining host-path readers can break resume. It is not yet
a complete preview-isolation switch.
Review repair and publication authorization
For Kiro Crew PR CI AI comments, MUST load
kirocrew-prepare-pr: Review repair routing
BEFORE any repair. Follow that canonical model-pinned delegation procedure;
parent self-fixing is not a substitute. This pointer does not change the profile's
read-only local-review gate.
Committing, pushing and opening a PR each require explicit user authorization.
Permission to commit is not permission to push, and green gates grant neither.
A fix-and-push monitoring scope must be explicit; absent it, confirm each push.
Never push to a protected base branch. For authorized publication, follow
kirocrew-prepare-pr's SHA-pinned force-with-lease protocol and single_commit handling,
and AGENTS.md's at-most-two-commit limit; never use an implicit lease or interactive
git. No direct merge: hand back review-ready work, or let kirocrew-prepare-pr's
Phase 4 arm auto-merge as its Mode table decides. Any hold turns that off,
including "don't merge", "review only", or a conductor's or work brief's hold.
Auto-merge still waits for the required approval and every check.
Cleanup and comments
Keep scratch, PR bodies, logs and QA media under $KIROCREW_SCRATCH, not the
worktree. Capture scripts' gitignored temp-screenshots/ is also permitted;
evidence is uploaded as attachments, never committed — with write access, which
this agent has. (A fork contributor, whose --attach upload GitHub refuses, may
instead git add -f the media there; see kirocrew-prepare-pr's Screenshots.) A directory
that must survive your own processes so a later run can advance it is not scratch
and not /tmp: report that you need one instead of choosing a path. Before
ending:
git status --porcelain
git diff -- <regenerated-file>Inspect every change. Restore/remove only residue YOU produced, not user changes
or another agent's work. Discarding uncommitted content is unrecoverable: ask if
ownership is unclear. Leave requested uncommitted implementation intact and report
it; cleanup is not permission to commit or discard it. A merged worktree must be
clean, including untracked files, before Dev Fleet's Prune merged can reap it.
After merge and cleanup: git worktree remove ../kirocrew-wt-<name>.
Add a genuinely needed ignored scratch pattern to .gitignore only within scope.
Comments and docstrings state current behavior and the non-obvious why: invariants,
edge cases, units and security constraints. No task log, PR/CR numbers, review-round
markers, incident dates, milestone tags, SHAs or historical narration. Do not
paraphrase adjacent code. Leave _vendor/ and semantic pragmas (type: ignore,noqa, pragma, TypeScript/eslint directives) untouched.