Evan-Daruwalla/claude-skill-suite

landing-check — did the work land, and are the claims true?

Three questions nothing else in the toolchain asks. /code-review reads the diff for correctness. /audit sweeps the whole project, cold. coding-conventions makes the author re-run their own code. All three can pass while the change went to a file nothing executes, the report states a number that is wrong, and half the copies were never touched.

The failure class this exists for, observed in practice: a security fix was deployed by looping over every .git/hooks/pre-commit file in a set of repos. But several of those repos set core.hooksPath, and git then ignores .git/hooks entirely — so in some of them the fix landed on a dead file while the live hook kept the old path. The loop confirmed every write, and the report asserted the fix was universal for six days. The diff was correct. The verification was real. Both were about the wrong file.

Every check below is one a real run needed. None is speculative — see Provenance at the end. Ten runs in: the sweep has found zero code bugs and roughly forty false claims, mislanded fixes, dead guards, and corrections that did not propagate. That is the shape of what it catches.

Trigger

/landing-check, or: "did that land", "sweep the changes", "verify what I just did", "check my work before I commit". Invoked only — it costs an agent, so it never fires on its own.

Not this skill

• Code correctness in the diff → /code-review (built-in; has multi-agent and --fix). Do not re-review logic here.
• Whole-project sweep → /audit (cold, expensive).
• Re-running what you just wrote → coding-conventions (same context, author's own eyes, fires from the PostToolUse hook). If a finding belongs to one of those, name it and hand it off — don't grow this skill into them.

How to run it

Spawn ONE fresh agent (Agent tool, general-purpose). Moderate context: this is scoped to what changed, not to the project.

What the agent gets — artifacts only:

• the repo paths touched, and one line on what the project is;
• git status + git diff (and git diff --stat per repo) — or, for non-repo trees, the list of paths modified this session;
• the WRITTEN account of the work: commit messages, the record entry, the final report text, any doc edited to describe the change;
• which of the touched trees are PUBLIC (pushed anywhere), if any;
• steps 1–3 and the output format below, and the instruction to mark anything it does not reach NOT SWEPT.

What it must NOT get: your account of what you did. The claims are the specimen. If you hand the agent your summary, you hand it the same blind spot that produced the error — you become both defendant and witness. It re-derives every claim from the written artifacts and from disk. This is the one rule that makes the skill worth its cost; drop it and this is self-grading in a second window.

The agent re-reads state live. git status, ahead/behind, what is pushed: derived at sweep time, not from the capture it was handed — other sessions commit, push and edit in between (a concurrent session has committed a fix inside an unrelated commit; six repos were pushed minutes before a sweep that was told "zero pushes"). ahead N proves nothing without a fetch first — ask the remote (git ls-remote, then rev-list --left-right --count @{u}...HEAD).

1. LANDING — did the edit reach the file that runs?

For every changed file, ask whether another copy exists and whether something chooses between them. The signature is: the edited file and the executed file have the same name and different paths.

• Shadowing: git config core.hooksPath (set ⇒ .git/hooks/* is dead code; resolve with git rev-parse --git-path hooks/<name>, the resolver git itself uses), PATH order for scripts, symlinks, vendored copies under node_modules, a venv copy vs a repo copy.
• Multi-tree systems: if a file exists in more than one tree (installed / lab / public, dev / dist, src / build), which one does the runtime load? Was the changed one it?
• Registration: if the change adds or edits something that must be wired up (a hook, a scheduled task, a route, an entry point), read the registration file and confirm the path there resolves to the file you changed.
• Activation — the right file is necessary, not sufficient. For anything that is a control — a guard, deny rule, hook matcher, regex, permission pattern, canary assertion, ignore rule — plant the positive it exists to catch and watch it fire, then a negative and watch it stay quiet. A control never observed firing is UNVERIFIED. Controls that read as protection and could not fire, all found by runs: deny rules that were prefix matchers and could never block a mid-command token; a leak regex whose \R compiled to X:Repo and could not match a real path; a case-sensitive name filter that let an all-caps spelling through; an assertion /findings/.test(src) that was true forever.
• Verdict per changed file: LANDED / MISLANDED / UNVERIFIABLE (say why). A file whose executing copy you could not determine is UNVERIFIABLE, not LANDED. A dead copy left behind is not itself a finding — say which copy is live and that the edit is in it.

2. CLAIMS — does the written account match disk?

Extract every checkable assertion from the artifacts, then re-derive each one independently — a different command than the one that produced it, where a different one exists. Priority order, highest yield first:

1. Universals — "all 7", "every copy", "byte-identical", "none remain", "0 findings", "nothing leaked across all 17 repos". One miss falsifies the whole statement, and this is the class that produced the 08-05 error. Enumerate the full set and check each member; never accept a spot check as proof of a universal. A universal you could only spot-check is UNVERIFIED.
2. Counts — "17 rules", "29 skills", "21/21". Count from the source of truth, not from the sentence. Watch the tense: a count of record entries written inside the entry being counted is off by one the moment it lands; a "the only dirty files left" claim made from inside a dirty file is self-falsifying; a citation of an entry that does not exist yet points at nothing until it is written.
3. States — "passes", "clean", "regenerated", "pushed". Run it, or read the artifact that proves it. Quoted output is a claim about the program: if an artifact quotes a line of tool output as evidence, grep the tool for that literal — a line printed only on the failure branch cannot evidence a pass, and a line the program never emits is a fabricated quote.
4. Numbers — a value with more significant figures than the output it traces to is invented (13 share counts published at 4 dp from a 2 dp table: 12 wrong). A statistic from a LIVE corpus (transcripts, logs, a growing record) is not reproducible without its corpus state pinned (file count + timestamp); report it as UNVERIFIED-AS-STATED and name what does reproduce. Check the arithmetic inside one sentence ("27 plus 3 … 29 today"). For comparisons: same sample set in every arm, estimator named, n stated — a 2×2 whose arms do not share batches is not a comparison. An absent field is not false. An == on a float that round-trips through float32 never matched, and the committed JSON held 0.0 for a true 0.605.
5. Internal contradiction — a sentence falsified by its own paragraph's caveat five lines later; a claim and its refutation both present with no dated strike; the same figure quoted three times and wrong in all three.

A claim you cannot check is reported as UNVERIFIED, never as true. A claim that is true but for a different reason than stated is a finding.

3. COLLATERAL — what should have moved and didn't?

The diff shows what changed. This asks what the change obligated and left undone. Run the inverse of the diff: for each changed value or concept, grep the tree for other places still asserting the old one.

• Sibling copies — every other tree/branch/package holding the same file. The fix for a mislanding has itself mislanded in one of two copies; a wholesale copy of a private file over its public twin has re-introduced what an earlier commit deliberately removed — compare the public copy to its own git show HEAD:, not to the private source.
• Correction propagation — if the work corrects a claim, find every other copy of the OLD claim: HANDOFF, PRD, memory bins, README, code comments, the public twin, the skill's own description. "Corrected in the record, left HANDOFF and the bin stating the old numbers" and "the third copy of this claim was still unstruck" are the most repeated finding of every run.
• Docs quoting the changed value — READMEs, skill descriptions, HANDOFF, PRD, comments, MUST print N/N pins. A count changed in code and unchanged in prose is a finding.
• Publication leak — landed where it must NOT. If any touched tree is public: sweep its diff (and the files it touched, against HEAD) for private identifiers — names, emails, absolute paths, private project names — case-insensitively. Say whether this change introduced the string or inherited it; the two need different fixes and the author has called the first the second before.
• Generated twins — rendered HTML of a record, TOC rows, memory bins, anchors: regenerated? An anchor that resolves but lands one heading early is not "broken", so broken: 0 does not prove it.
• New artifact types — did the work create a kind of file that did not exist before (cache dirs, build output, logs)? Is there an ignore rule, and is it in the right repo?
• Byproducts — files created and left behind: untracked files, temp directories, __pycache__, .tmp, backups. List them with full paths; do not delete anything.
• The obligation the docs state — if the project's own rules say a change of this kind requires something (a record entry, a regenerated twin, a synced snapshot, a test run, a pin bump), check that it happened.

Probe — the mechanical half, run before the agent judges

landing-probe.js (portable Node, zero deps, read-only) enumerates what steps 1–3 would otherwise re-derive by hand, so the agent reads a census instead of guessing one. Give it every touched tree; hand its output to the agent as an artifact.

node landing-probe.js probe <root>... [--file <p>]... [--since <min>]
                           [--settings <json>] [--remote] [--all]
node landing-probe.js --canary

• HOOK — per repo, per hook: the file git will RUN (git rev-parse --git-path, honours core.hooksPath) and any DEAD copy in the other place; a registered path with no file is MISSING.
• CHANGED — staged / modified / untracked per repo; --file names files outside any repo; --since N adds files modified in the last N minutes under NON-repo roots (an installed tree is not a repo).
• TWIN — for each changed file, every same-name file under any root, SAME or DIFF (sha256, CRLF-normalized). Conventional names (SKILL.md, README.md — any basename with more than 6 copies) match on parent-dir/name, not name alone. --all lists every twin group, not just the changed ones.
• REG — --settings <json>: every path-shaped string (after ~/$HOME/ %USERPROFILE% expansion) and whether it RESOLVES.
• REMOTE — ahead/behind per repo, labelled CACHED ref unless --remote fetched first. ahead N off a cached ref proves nothing.

Exit 1 when there is something to look at (a DEAD hook, a DIFF twin of a changed file, a MISSING registration); 0 for a clean census. It judges nothing — DIFF between a private and its genericized public copy is expected, and the agent says so. --canary builds throwaway repos (hooksPath set / unset / pointing at nothing, a bare remote and a clone one commit ahead, CRLF twins, a conventional-name forest, a settings file with one missing path) and MUST print CANARY PASS 23/23 before you trust a census.

Output

Short. This is a pre-commit sanity sweep, not an audit report.

• Verdict line: SAFE TO COMMIT / FIX FIRST / UNVERIFIABLE, one sentence.
• Coverage line: which artifacts were read, which trees swept, and what was NOT SWEPT — a short report must never read as a clean one.
• Landing table: file | copies found | which one executes | fired on a planted positive? | verdict. Only rows that are MISLANDED or UNVERIFIABLE need detail.
• Claims table: claim (quoted) | source artifact | re-derived result | TRUE / FALSE / UNVERIFIED.
• Collateral list: what should have moved and didn't, each with a path.
• Confirmed: the load-bearing negatives and the numbers that re-derived exactly (no misland, no leak, 25/25 reproduced) — a sweep that reports only faults teaches nothing about what is solid, and the author's record needs them.
• Handoffs: anything belonging to /code-review, /audit, or coding-conventions, named and passed on rather than half-done here.

After the report

• Re-verify every finding by hand before acting. The verifier's claims are claims: one run's re-derived integers were off because the corpus had grown under it; the invariants held. Accept what reproduces.
• The fix is subject to the same three questions. Fixes to FIX FIRST findings go to every copy and every doc quoting the old value — the run that fixed a one-of-two-copies mislanding shipped the fix to one of two copies.
• Write the record entry from the re-derived values, after the last verification run, not before it: an entry written first records the state you expected, and the append-only record makes that permanent.

Rules

• Findings only. Change nothing, delete nothing — including byproducts and files that look like junk. Report paths; the author decides.
• Evidence or it isn't a finding: a command and its output, a file:line, or a reproduced state. No speculation.
• A zero result is a claim. If a search comes back empty and something load-bearing rests on it, re-run it a second way before believing it — and if a whole-repo search returns nothing, grep the ignore files for {/} first (one malformed glob silently hides an entire repo from ripgrep-backed tools; git is unaffected). A backslash pattern that matches nothing proves nothing: retry with grep -F, or build the path from a char code.
• Never fabricate a count, path, or command output. "Could not determine" is a valid and expected verdict.
• Cut short is not clean. If the agent dies or is resumed mid-sweep, it marks what it did not reach NOT SWEPT; a partial report that says which parts are partial is worth more than a complete-looking one that does not.
• The probe has a canary; the judgment does not. Steps 2–3 and the verdict are prose run by an agent, and are only as good as that agent — a clean probe is a clean census, not a clean sweep. Say so rather than implying a gate.

Provenance

Ten runs over eleven days across four projects (a skills suite, an ML/robotics project, a web app, a trading system). Among what they found: four false claims on the first run, one inside this skill's own anecdote; a fix broken by its canary's blind spot; a "pre-existing" leak the fix pass had introduced; a fix landed in one of two copies; a dead leak regex; every record anchor one heading off; a personal identifier in a public repo; 27 stale doc pins; four wrong numbers in a sealed entry; quoted output the program never prints; a false universal in a public repo; an unmatched 2×2; a float == that wrote 0.0 for 0.605; an unrendered status; a route enumeration wrong in three places including the source comment; prefix-matching deny rules that blocked nothing; 4 dp share counts from 2 dp output. Zero of these were code bugs.