-
Load stack.toml. From the plugin install dir
($CLAUDE_PLUGIN_ROOT/stack.toml) or, if running standalone, from the
repo root. If neither exists, STOP and report.
-
For each [[host.tool]] and [[deps.tool]]:
- Run
check_cmd. Capture stdout+stderr.
- Parse the version with a simple regex (
(\d+\.\d+(\.\d+)?) —
accept "X.Y" and "X.Y.Z" both). If parse fails or command not
found, mark "missing" (and "outdated" if optional = false).
- Compare against
min_version using a tuple compare (split on .).
- Mark:
ok | outdated | missing | unknown (parse failed).
- Resolve update hint per-OS: detect platform via
uname -s (or
PowerShell $env:OS / $IsWindows). Prefer update_hint_macos /
update_hint_linux / update_hint_windows when present and the
platform matches; otherwise fall back to the generic update_hint
field. This is what stack.toml manifest_version = 2 introduced.
-
For the plugin itself:
- Read
version from the installed .claude-plugin/plugin.json.
- WebFetch
https://raw.githubusercontent.com/Filip-Podstavec/claude-leverage/main/.claude-plugin/plugin.json
and compare to the installed version. Cache the network result for
24h in ~/.claude/claude-leverage/.stack-check-cache.json to avoid
hammering GitHub if the user runs the skill repeatedly.
-
For Claude Code itself:
claude --version is the installed version.- We do NOT fetch the latest Claude Code version from the network —
Anthropic doesn't expose a stable scrapable endpoint. Instead, the
min_version in stack.toml is hand-maintained: bump it when this
plugin starts depending on a feature only present in a newer CC.
-
Walk the current repo for AIDEV anchor health (if cwd is inside
a git repo). Grep git rev-parse --show-toplevel for
AIDEV-(TODO|QUESTION) matches and parse each line.
Two flavors of anchor:
- Deadline-bearing:
AIDEV-TODO(by: 2026-08-01): or
AIDEV-QUESTION(by: YYYY-MM-DD):. Extract the ISO-8601 date and
compare against today.
due-soon (deadline in the next 14 days)overdue (deadline already passed)
- Age-based (no deadline): use
git log -1 --format=%cI -- <file>
for the file's last-modified timestamp.
fresh (≤30 days)aging (30–90 days)stale (>90 days)
Cap walk at 5000 files; skip the bench archive, vendor dirs,
node_modules, pycache, .git.
Reported after the version table:
## AIDEV anchors (current repo: <name>)
- 14 AIDEV-TODO total: 3 fresh, 8 aging, 2 stale (>90d), **1 overdue (by deadline)**
- 5 AIDEV-QUESTION total: 1 fresh, 2 aging, 1 stale, 1 due-soon
**Overdue (deadline passed):**
- `src/billing/charge.py:47` — AIDEV-TODO due 2026-04-01 (44 days overdue)
"replace the polling loop with webhooks"
**Due soon (next 14 days):**
- `src/auth/middleware.py:89` — AIDEV-QUESTION due 2026-06-05 (in 12 days)
**Stale (no deadline, last touched >90 days ago):**
- `src/legacy/handlers.py:120` — AIDEV-TODO (last touched 2025-12-03)
"migrate to v2 API"
If not in a git repo, skip this section silently.
Deadline parsing tolerates a few formats. The distinguishing rule
for implementors: the parenthesized content matches one of these
regexes (anchored, case-insensitive on the keyword):
^by:\s*(\d{4}-\d{2}-\d{2})$ — (by: 2026-08-01)^(\d{4}-\d{2}-\d{2})$ — (2026-08-01) (bare ISO date)^deadline:\s*(\d{4}-\d{2}-\d{2})$ — (deadline: 2026-08-01)
Anything else in the parens (free-form notes like (Q3 2026) or
(after the migration)) is NOT a deadline; that anchor falls under
age-based stale tracking. This prevents misclassifying ambiguous
parenthetical text as a date.
-
Sanity-check AGENTS.md (if present in cwd or repo root):
- File size, two tiers (report the size either way):
- warn over 8 KiB — lean target exceeded; an always-loaded file
this large dilutes attention. Suggest extracting topic depth to
docs/ behind a when-to-read link.
- flag over 32 KiB — Codex caps the assembled project doc at
project_doc.max_bytes (default 32768) and silently drops the rest,
so content past that byte is invisible to Codex agents.
Both tiers are advisory here and do not stop the timestamp reset (step
9). /repo-doctor escalates the 32 KiB tier to a hard ❌ — that split
is intentional (see ADR 0009): /stack-check informs, /repo-doctor
gates.
- Broken
@<path> imports: grep for ^@ lines, verify each
referenced file exists relative to the importer.
- Stale file references: extract
path/to/file.ext-shaped strings
from the body and check existence (best-effort; lots of false
positives, so report only the obvious ones — e.g. when AGENTS.md
mentions scripts/foo.sh and scripts/foo.sh does not exist).
Per-directory AGENTS.md files (**/AGENTS.md, depth ≤ 3) get the same
two-tier size check — and they share the same ~32 KiB Codex budget as the
root file, so a large root plus several per-dir files can truncate even
when no single file exceeds the cap.
Reported after the anchors section:
## AGENTS.md sanity
- `AGENTS.md` — 11.3 KiB, **over 8 KiB lean target** (extract topic
depth to `docs/` behind a when-to-read link)
- `src/billing/AGENTS.md` — 1.1 KiB, ok
- `src/api/AGENTS.md` — **38.4 KiB, over Codex 32 KiB cap** (Codex
will silently drop content beyond byte 32768; split it)
- Broken imports: _none_
- Possibly stale references: 1 (`scripts/old_runner.sh` mentioned
but not found)
-
Walk every tracked *.md file for stale path references. This
is the generalization of step 6's "Possibly stale references" check
beyond AGENTS.md — the failure mode "doc references a file that was
moved or deleted N versions ago" is repo-wide, not AGENTS.md-only.
Scope:
- Use
git ls-files '*.md' so untracked / .gitignored markdown is
not scanned. Skip files under the bench archive, node_modules/,
vendor/, __pycache__/, .git/, and anything inside a
.gitignored tree (relying on ls-files filters most of this).
- Cap walk at the first 200 markdown files (sorted) — repos with
larger doc trees get a "(truncated)" footer; honest scope, not a
silent half-job.
Per file:
- Strip fenced code blocks (
``` ... ```) and inline backticks
before regex-walking the body. Path-like tokens inside code blocks
are usually placeholders / examples, not live links.
- Extract candidates with a deliberately conservative regex (rough
shape, implementor adjusts):
(?:^|[\s(\[>])([a-zA-Z][a-zA-Z0-9_.-]*(?:/[a-zA-Z0-9_.-]+)+.[a-zA-Z0-9]+)(?:$|[\s.,;:)\]])
— must contain at least one /, must end in a file extension, must
not look like a URL (skip if preceded by :// or starting with
http).
- Skip tokens that look like commit hashes, version strings
(
v1.2.3), or domain names (anything matching \.(com|org|io|dev|gov)$).
- Resolve relative to the repo root (
git rev-parse --show-toplevel).
If the file does not exist, record <md-file>:<line> → <token>.
Output budget: hard cap of 20 broken refs in the report; if more,
show the first 20 plus a "(N more not shown)" footer so the section
does not dominate the report.
Reported as a new section:
## Markdown link audit
Scanned 47 tracked `*.md` files; 3 broken path references.
- `README.md:118` → `scripts/legacy/runner.sh` (not found)
- `docs/adr/0002-...md:42` → `agents/old-reviewer.md` (not found)
- `workflows/security-first-feature.md:88` → `templates/old-logging.md` (not found)
If zero broken refs, write _All links resolve._ and move on. If
the scan was skipped (not in a git repo, or CLAUDE_LEVERAGE_SKIP_MD_LINK_AUDIT=1),
omit the section entirely.
This check exists because the v1.0.0 → v1.3.3 pivot in this repo
left three stale tests/README.md / AGENTS.md / ci.yml refs
that survived 4–7 version bumps before being caught by a human-led
pre-push audit. Catching that class of drift automatically is the
point.
-
Emit the Markdown report combining version table + anchors +
AGENTS.md sanity + markdown link audit. Tier the version rows:
required first, then optional. Required-failing rows in bold.
-
Reset the timestamp. Only if no row failed with an error
(process crashed, network exception). A failure status (outdated /
missing / stale anchors / oversized AGENTS.md) is information, not
an error — reset the timestamp.
Run exactly this command — do NOT write a literal number into the
file:
date +%s > "$STATE_DIR/.last-stack-check"
where $STATE_DIR is ${CLAUDE_LEVERAGE_STATE_DIR:-${XDG_STATE_HOME:-$HOME/.local/state}/claude-leverage}
(matching the fallback chain in scripts/hooks/stack-freshness.sh).
The SessionStart hook reads the epoch from the file body — not the
mtime — so a stale or hallucinated number suppresses nudges for
weeks. Capturing the value via date +%s is the only reliable way
to avoid the v1.4.0 field-feedback #7 bug where the body and mtime
disagreed by ~2 months.
