verify-pi-lens
Skill530 starsChanged 17 days agoMITView on GitHub
Drive a built pi-lens through its real MCP or tool layer, with home and data dirs pinned under the lane TMPDIR, to verify LSP diagnostics with a real server (Pyrefly), didSave (clangd), the formatter, the review graph and rename_file, on the checkout or on any other commit. Use when a fix or review needs a real-path probe, not a mock, and a model run (scripts/run-harness.mjs) is too costly.
In this repository11 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- llms.txtllms.txt
- orchestrateSkill
- release-qaSkill
- retroSkill
- writeSkill
- pi-lens-ast-grepSkill
- pi-lens-lsp-navigationSkill
- pi-lens-write-ast-grep-ruleSkill
- pi-lens-write-tree-sitter-ruleSkill
The file
We split this skill into the rules it gives an agent
1 rule. About 0 are common, 0 are distinctive, and 1 is very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈0
Distinctive · estimate
This author's own choices. Worth a look.
1
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Very specific: line 49
Read-only on the tree. It checks the env pins, that the compiled twins are not older than their sources, that the Pyrefly venv exists, that …
≈ “Common” is an estimate: at least 1 in 5 of the 372,429 skill files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
Verify pi-lens
Procedure only. Role rules (what a reviewer or fixer must prove, and how) live in
docs/pi-lens-reviewer.md,docs/pi-lens-fixer.mdanddocs/pi-lens-subagent.md; on any conflict they win. This skill replaces probes rebuilt per lane, which vanished with the worktree (#4359).
Shape after the create-verification-skill pattern: Launch, Doctor, Drive, Evidence, Cleanup, Helpers, and a features map.
Launch
Linux only (pid tracking reads /proc; verify.mjs exits 2 elsewhere), with
node, git and uv. Everything a run writes lands under
the lane TMPDIR; env.sh refuses /tmp.
export TMPDIR=~/.local/share/pi-lens-orchestrator/tmp/<lane>-tmp
S="$(git rev-parse --show-toplevel)/.claude/skills/verify-pi-lens/scripts" # absolute
npm run build # the tree under test must be built
$S/setup.sh pyrefly # once: Pyrefly 1.3.2 + ruff venv (py-venv)
. $S/env.sh # pins HOME, PI_LENS_HOME, PILENS_DATA_DIR, XDG, uv cacheenv.sh sets RUN (disposable) and EVIDENCE (kept). Repeat the S= and
. $S/env.sh lines in every shell call: agent shells reset their environment
between calls.
The server under test is the real stdio MCP server, mcp/server.js, launched
inside the scenario workspace as a host would. To verify a commit other than the
checkout (the base, or the commit before a fix), build it first and pass it as
--tree:
OLD=$($S/setup.sh tree <rev>) # git archive + build under $VERIFY_ROOT/treesDoctor
node $S/verify.mjs doctor --tree "$PWD" # or --tree "$OLD"Read-only on the tree. It checks the env pins, that the compiled twins are not
older than their sources, that the Pyrefly venv exists, that tools/list serves
the driven tools, and that pilens_health answers with tree-sitter available
and no degradations or broken servers. Alive is not usable: do not drive past a
FAIL.
Drive
node $S/verify.mjs drive <scenario> [variant] --tree "$PWD"Exit 0 when every expectation holds. The verdict line counts only real
expectations: PASS (3/3, 1 KNOWN) means three passed and one residual stayed
KNOWN. Output lines are PASS, FAIL, KNOWN
(a residual an open issue owns: not a failed verdict) and XPASS (the residual
no longer reproduces: say so, the issue's acceptance may hold). Each feature file
below lists its scenario, variants, the user goal and the exact checks. Write
the report from the PASS/FAIL lines; quote them.
To prove a fix, run the scenario on the commit before it and on the fix:
FAIL on the first with the defect's own line, PASS on the second.
Evidence
$EVIDENCE/<scenario>-<variant>-<stamp>-<pid>/, one directory per drive, so a
head drive and an old-tree drive in one shell never share a wire log (a sibling
of $RUN, so cleanup keeps it). The verdict line prints the path:
mcp-wire.jsonl: every JSON-RPC frame, timestamped, plus the server's stderr.lsp-wire.in,lsp-wire.out: both directions of the LSP stream, for the scenarios that launch Pyrefly throughpyrefly-tee.sh.steps.json: each call's tool, arguments, elapsed ms and full result.summary.json: the expectations with their details.pi-lens-home/: the run's pi-lens home with its logs.
Quote paths and lines from these files in the report; copy what the review must
keep to the lane's report directory, since $TMPDIR is the lane's, not the PR's.
Cleanup
drive and doctor stop their server and kill only the pids they recorded in
$RUN/pids.json (pid plus process start time, so a reused pid is never
signalled). When a run died half way:
node $S/verify.mjs cleanup # sweeps every runs/*/pids.json under $VERIFY_ROOT, removes those run dirs, keeps evidence
node $S/verify.mjs cleanup --run "$VERIFY_ROOT/runs/<id>" # one run dir only
node $S/verify.mjs cleanup --trees # also drops trees built by setup.sh tree (unlinks node_modules first)A re-sourced env.sh mints a new RUN id, so the default sweeps every run dir
under this TMPDIR; do not run it while another drive of the same lane is live.
Then pgrep -fa 'pyrefly|mcp/server' must show none of your servers.
Helpers
Executable, under scripts/; invoke them from this body, never copy them.
| Helper | Role |
env.sh | Sourced. Pins HOME, pi-lens, XDG and uv dirs under RUN; rejects /tmp. |
setup.sh pyrefly / tree <rev> | The Pyrefly + ruff venv; a built copy of a commit. |
pyrefly-tee.sh | Registered as the Pyrefly command in the scenario's .pi-lens/lsp.json; copies the LSP wire. |
verify.mjs doctor / drive / cleanup | The driver: wire log, pid tracking, expectations. Its MCP client is scripts/lib/mcp-stdio-client.mjs, shared with scripts/release-qa.mjs. |
scenarios/*.mjs | One module per feature: variants, setup, run. Add a variant here, not a new script. |
Reuse, not duplicated: the shared fake server tests/fixtures/fake-lsp-server.mjs
and the vitest harness tests/mcp/harness.ts stay the unit-level tools;
scripts/smoke-tools.mjs --format is the per-language formatter sweep;
scripts/run-harness.mjs is the model-driven end to end (too costly for routine
verifies).
Features map
| File | Scenario (variants) |
features/lsp-diagnostics.md | lsp-created-module (issue, importer-first, importer-only, replay) |
features/rename-file.md | rename-file (into-missing-path) |
features/review-graph.md | review-graph (mcp-create-after-build, builder-incremental) |
features/formatter.md | formatter (python-ruff) |
features/did-save.md | lsp-did-save (analyze) |
features/fallow-lsp.md | fallow-lsp (off, config, flag, bad) |
Each file has the sections Sub-features, How to get to it (user POV), Driving it with the harness, and Gotchas. Add a feature file when a seam has no recipe, and add the probe shape of any reviewed defect as a variant of its feature.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

