| name | lint |
| description | Run health checks on the vault for inconsistencies, stale data, orphaned notes, and missing connections. Use periodically or when user says "lint", "health check", or "audit". |
| model | opus |
| effort | max |
| allowed-tools | Read Grep Glob Bash(find * wc * date * grep * sed * cat * printf * rm * mkdir * ls * python3 *) |
Comprehensive vault health audit. Read-only — never deletes, never edits. Report only.
Design rationale in .claude/skills/lint/RATIONALE.md (§N.M anchors).
Architecture (2026-07-08, script-first): the deterministic majority of the check registry (~40 of 55 checks — set-diffs, date arithmetic, regex, schema probes) is implemented in .claude/skills/lint/lint.py, following the generate_graph.py precedent. The script is the authoritative spec for those checks; this file specifies only the LLM-judgment checks and the merge procedure. Main-context cost drops from a full-vault read to one script call + targeted reads of flagged files.
Step 0: Pre-flight (MANDATORY — runs before Scope Resolution)
0.1: Acquire vault lock (mode-dependent)
Per .claude/skills/_shared/preflight.md Procedure 1. Even though /lint never writes to vault content files, a lock is still acquired so that concurrent writers (e.g., a mid-flip /sync manifest, an in-flight /prune sector edit) do not cause /lint to see transient inconsistent state and raise false-positive flags.
- Full mode (
/lint with no arguments): acquire a vault-wide scope lock. Timeout budget: 5 minutes.
- Scoped mode (
/lint TICKER): acquire a .vault-lock.readonly lock. Timeout budget: 2 minutes.
Capture the token emitted at Step 0.1. Verify ownership (Procedure 1.5) before subsequent Bash blocks. Release the lock explicitly in the final reporting Bash block via rm -f "$LOCK_FILE".
0.2: No rename-marker check
/lint is a read-only diagnostic skill and is an explicit exception in .claude/skills/_shared/preflight.md §2.1. lint.py #37 surfaces active .rename_incomplete.* markers as check output — hard-blocking would defeat lint's purpose.
Step 1: Deterministic pass — run lint.py
python3 .claude/skills/lint/lint.py
python3 .claude/skills/lint/lint.py --ticker $T
- Covers (authoritative in-script): #1–#6 (empty-only for #5), #8, #10, #11, #12/#13 (mechanical flags), #14 (heading presence), #15, #16, #17–#27, #28 (candidates), #29 (presence checks a/b/c), #30, #32–#39, #41–#56.
- Exit codes: 0 = pass/nice-to-have · 1 = Important found · 2 = Critical found · 3 = self-validation failed (do NOT report partial results — fix the script invocation or escalate; the script's stderr states the cause).
- The script's report contains three severity sections plus a Judgment candidates section — those candidates are inputs to Step 2, not confirmed findings. Never copy candidates into the final report unreviewed.
Step 2: LLM-judgment pass
Read ONLY the files the script flagged or that these checks name. Do not re-run any check the script covers; do not full-read the vault.
-
Thin critical sections — for each #5 judgment candidate (section exists but under ~25 words), read the section and judge whether it is substantively empty (placeholder prose, restated heading) vs. legitimately terse. Substantively empty → Important.
-
Old financial data — for theses flagged by #6 or with conviction: high, read ## Key Metrics and judge whether the data vintage (as-of dates, quarter references) is older than 6 months. Stale → Nice to Have; stale AND driving a conviction claim → Important.
-
Unlinked mentions — using Grep (word-bounded ticker tokens across Theses/ Sectors/ Macro*/), identify notes that discuss a ticker with a thesis but never wikilink it. Judgment: distinguish substantive discussion (flag, Nice to Have) from passing mention (skip).
12/13. Conviction-evidence and Bull/Bear interpretation — for each #12 candidate and #13 flag from the script, read the thesis and judge whether the mechanical count/ratio reflects a real evidence gap or bias. Upgrade to Important only with a specific stated reason.
-
Partial-write confirmation — for each #28 candidate, read the flagged section and confirm whether it is genuinely truncated (mid-sentence, broken table) or a false positive (stylistic fragment, list item ending in a preposition-like token). Confirmed → Important, and name the likely matching snapshot in _Archive/Snapshots/ if one exists.
-
Registry reverse-check + semantic drift — the script verified registry→skill presence. Do the reverse: scan each producer SKILL.md's quoted Log-prefix strings (Drift-coupling blocks, Log-format fences) for prefixes NOT in _shared/log-prefixes.md → Important (orphan coupling). Also judge any #29 script finding for semantic (not just literal) drift before reporting.
-
Graph-primer anti-pattern review — for each #55 candidate, read the surrounding SKILL.md context and judge whether the matched phrase actually instructs primer-as-filter behavior (violation → Important) or is a negation/anti-pattern warning quoting the phrase (skip).
Step 3: Merged report
Merge Step 1 findings (verbatim) with Step 2 confirmations into the standard format. Attribute each finding to its check ID. Drop unconfirmed candidates silently.
Critical (breaks research quality)
Important (gaps in coverage)
Nice to Have (optimization)
Stats
(reproduce the script's Stats block, plus: judgment candidates reviewed / confirmed / dismissed)
Release the lock in the same final Bash block.
Maintenance contract
lint.py is the spec of record for the deterministic checks. When a check's rules change, edit the script (and its docstring COVERED list), not this file.- New deterministic checks go in the script; new judgment checks go in Step 2 here.
- The script never writes vault state and reads a point-in-time snapshot in <1s, which also shrinks the lock-contention window that motivated Step 0.1's 5-minute budget.
- Cross-skill contracts enforced:
_shared/log-prefixes.md (#29), _shared/sector-resolution.md (#30/#34), _shared/hot-md-contract.md (#35/#42), _shared/graph-primer.md (#54/#55), _shared/preflight.md (#43).
