verify-claudinho
Skill35 starsChanged 6 days agoMITView on GitHub
Drive one surface of the built Claudinho binary in isolation, offline by default, and keep the evidence. Use it to verify a change on the real CLI, the MCP server over stdio, the statusline or the hook without touching the operator's config, cache or network.
In this repository10 other files it gives its agents.
- AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- .cursor/rules/bundle-bracket-pr.mdcCursor rule
- .cursor/rules/change-discipline.mdcCursor rule
- .cursor/rules/release-discipline.mdcCursor rule
- .cursor/rules/surface-parity.mdcCursor rule
- .cursor/rules/trust-boundary.mdcCursor rule
- claudinho-modeSkill
- claudinho-modeSkill
- verify-claudinhoSkill
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 34
Every child the control CLI starts (the bootstrap follow, the main command, the --json twin, the capture, the MCP server, the hook) runs …
≈ “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-claudinho
The inner loop for an agent that changed Claudinho: one command per surface, run on the REAL built binary in an
environment built from scratch, offline unless told otherwise, with the evidence of every child kept. The live
pass stays pnpm release:qa (a person, the real feed); this skill never replaces it.
Launch
pnpm -r build from the repository root. A CLI and a stdio server have no server to keep alive: there is nothing
to start and nothing to stop.
Doctor
node scripts/verify.mjs doctorChecks, each printed as a line ok <name> or FAIL <name> with its detail, the exit nonzero on any FAIL (--json
prints the same as one object, the checks list with each check's ok and detail, and no check lines; the
evidence path still goes to stderr): Node at or above
the engines floor; the CLI and MCP dists present and the CLI dist's --version equal to packages/cli/package.json
(that child is the phase version, with its own evidence and the deadline --timeout sets; a timeout FAILs it
whatever it printed); a temporary config directory made and removed; a temporary cache directory made, written and
removed (the harness's own, never the operator's); CLAUDINHO_COMPETITION and CLAUDINHO_TEAM REPORTED as set or
unset (a report: the children never see them). The feed is not asked.
Drive
Every child the control CLI starts (the bootstrap follow, the main command, the --json twin, the capture, the MCP
server, the hook) runs with an environment BUILT FROM SCRATCH: a temporary HOME, absolute temporary
XDG_CONFIG_HOME and XDG_CACHE_HOME, TZ=UTC, CLAUDINHO_NO_STAR=1, colour off except under capture, the mode's
preload loaded in the child through NODE_OPTIONS, every spawn the product asks for recorded and never started,
every fetch attempt recorded. The main command of run, replay and capture, and the --json twin of run and
replay, are also given --lang en unless the scenario passes --lang; the bootstrap follow and prompt, hook
and ambient are given none. The mode is --offline (the default: every fetch fails at once with a network error,
so the surface renders its degraded path), --replay <corpus> (a recorded feed; an unrecorded URL fails the run) or
--live (the provider, for a person on purpose; never an agent mid-task). --env KEY=VALUE sets only a scenario
key (CLAUDINHO_COMPETITION, CLAUDINHO_TEAM, CLAUDINHO_SOURCE, CLAUDINHO_MARKETS_SOURCE, CLAUDINHO_LANG,
LANG, TZ); every other key is refused. A scenario's CLAUDINHO_LANG reaches every child, which reads it as the
product does: the MCP server for a call with no lang, when it is exactly en, es, pt or fr; a command given
--lang en stays English; follow and ambient --json, given no --lang, take it as the CLI does.
node scripts/verify.mjs run --offline --follow premier-league --twin --json -- today 2026-10-07
node scripts/verify.mjs mcp --offline --follow premier-league --json get_today '{"date":"2026-10-07"}'
node scripts/verify.mjs mcp --json --list
node scripts/verify.mjs seed club --slug eng.1 --cache <dir> --json
node scripts/verify.mjs prompt --seed club --slug eng.1 --follow premier-league --json
node scripts/verify.mjs hook --seed club --slug eng.1 --follow premier-league --json
node scripts/verify.mjs replay <corpus> --follow premier-league --twin --json -- today 2026-10-07
node scripts/verify.mjs capture <label> --follow premier-league --json -- today 2026-10-07run refuses before any child: the install commands (init, init-statusline, init-hook,
init-cursor-statusline, claude, cursor), star, _refresh and any --copy (they write the operator's
settings or the clipboard, outside the sandbox). --follow --team is not offered (pinning a club reads the
provider). Every child has a deadline (--timeout <s>, 60 by default; the MCP session 30) and is killed and reaped
when it passes (TERM, a grace of two seconds, KILL), its partial evidence kept. A capture's deadline reaches the whole
process group and the descendants ps recorded before the first signal; when ps did not answer, or its snapshot
showed the leader already gone (a snapshot vouches only for a leader alive in it), the TERM is withheld so the
group stays whole for the grace and ps is asked again at the KILL step (a leader already reaped is not asked about,
since its pid may be another process's by then); the group and each recorded descendant are signalled only while not yet seen gone (each is retired at its
first ESRCH, the group and the tree probed every 50 ms through the grace), and a descendant the controller could not
see is said in the result with its reason (a descendant may have survived), never silently left.
The feature map, one file per surface with how to drive it and what proves it, lives at
.claude/skills/verify-claudinho/features/ (five surfaces so far: today, live, next, table, statusline,
the hook on the statusline's file). Every offline: line of those files is run by
packages/core/test/verify-cli.test.ts; a replay: line names what a recorded feed must prove (the next PR).
Evidence
Under --out <dir> (default: a timestamped directory under the temporary root, printed by every command): per
phase (follow, main, twin; the MCP session's server; doctor's version) <label>.<phase>.txt (stdout),
.err (stderr), .exit, .spawns (one JSON line per launch the product asked for, recorded and never started:
spawn, exec, execFile, fork and their sync forms alike) and .fetches (one JSON line per fetch attempt: the
URL, the mode and the outcome blocked, replayed:raw, replayed:synthetic, miss, malformed, or, live,
live:sent before the request and live:<status> or live:error after it); <label>.result.json (the aggregate:
ok, the mode, every phase's exit, the fetch entries each naming their phase, the spawn entries in phase order
(the counter's arrays; their phase is their file's), the environment's keys, the paths). The label is reserved in --out when a command starts: a second command with the same label and
directory is refused, so no run's evidence is ever overwritten or mixed. ok is false on any phase's nonzero
exit, the MCP server's own nonzero exit, any miss or malformed recording, a fetch attempt the preload could not
record (it says so on the child's stderr), a timeout, or an MCP failure; the child's own exit is kept beside it. A
capture writes <label>.ansi (the combined pty stream as script wrote it) and <label>.txt (the same with the
escapes and script's own header and footer lines removed, so its first line is the product's). The PNG card is
the maintainer's private demo tool, not this skill's.
Cleanup
The temporary HOME, config and cache directories are removed when the command ends (--keep keeps them). The
evidence directory is never removed.
Selecting the evidence for a change
| Changed behavior | The evidence to select |
| Provider parsing or a trust boundary | the recorded payloads and hostile inputs in packages/core/test/, the live and cache parity (AGENTS.md "Conventions": "Every data vendor implements the ProviderAdapter interface" and "Text has ROLES, not one universal cleaner") |
| A verdict or its presentation | the CLI text and --json twin, the MCP text and data, the share cards, a qualifier surviving a cut (run --twin, mcp; AGENTS.md "A verdict becomes output in ONE place") |
| Selection, configuration, the cache | precedence (--env CLAUDINHO_COMPETITION against --follow), refused inputs, cold, stale and partial state, the next call, concurrency (run, seed, prompt; AGENTS.md "The competition is decided ONCE, at the edge, and then travels as a value") |
| The statusline or the hook | the line, the spawn count, no network and no market work (prompt, hook on a seed) |
| An MCP contract | the base and head tools/list (mcp --list on both), the stdio smoke |
| Distribution or a release | the packaging contents, the manifests, pnpm release:qa (a person, live) |
Replay and live are two different things
A provider outage leaves a live check unverified. A green replay proves nothing about today's feed. Live ESPN is
the canary and pnpm release:qa, single-flight. An agent mid-task uses --offline or --replay.
Validation scope is unchanged
A prose-only change gains no build and no drive (AGENTS.md "Validation scope").
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.

