The Step 1 and comparison-mode resolver loops walked a hardcoded list
of install paths trying to find scripts/last30days.py. Two problems:
1. The list was never exhaustive — it covered ~/.codex/skills, but not
~/.claude/skills, ~/.cursor/skills, ~/.gemini/skills, ~/.copilot/skills,
~/.hermes/skills/research, etc. PR #406 was about to fix that by
enumerating more paths, but enumeration is the wrong shape.
2. The resolver could pick a different install than the SKILL.md the
model loaded from. Spec-vs-engine divergence is subtle and confusing
when it triggers.
The model already knows the SKILL.md path it loaded (from its Read tool
result). Templating that into the bash block is strictly better than
guessing across an enumerated list:
- Works for every harness without enumeration (Hermes, Cursor, anything
new) because we just use wherever the harness loaded SKILL.md from
- Aligns spec with engine — the engine runs from the same install the
spec was read from
- Deletes ~80 lines of bash across Step 1 + comparison-mode + the
prose preamble describing the resolver
Mechanics:
- SKILL_DIR placeholder in both bash blocks — model substitutes the
absolute path of the directory containing the SKILL.md it just Read
- One-line validation `[ ! -f "$SKILL_DIR/scripts/last30days.py" ]`
catches bad templating with a clear error
- All references to $SKILL_ROOT replaced with $SKILL_DIR (badge
VERSION lookup, prose description in the LAW-7 preamble area)
- STEP 0 unchanged — different concern (marketplaces stale-clone hop)
Version 3.2.3 -> 3.2.4 (behavior change: install paths the resolver
list never enumerated now work; install paths it did enumerate work
the same way they used to but via the SKILL_DIR template).
Closes#403.
The SKILL.md templates instructed the model to invoke the engine with
inline single-quoted JSON: `--plan '$JSON'` and `--competitors-plan '{...}'`.
When any resolved field value contained an apostrophe (common in `context`
strings like "McDonald's", "people's choice", or contracted forms like
"don't", "won't"), the inner `'` closed the outer single-quote and broke
shell parsing before the engine was even invoked.
Observed during PR #400 testing: a Codex run hit the trap and self-healed
by re-encoding, wasting one engine invocation and ~30s of latency.
Fix: switch both templates to the heredoc + tmpfile pattern. The engine's
`parse_plan()` and `parse_competitors_plan()` already check
`os.path.isfile(plan_str)` and read from disk — only the SKILL.md prose
needed to change.
The quoted heredoc marker (<<'PLAN_EOF') is load-bearing: it suppresses
shell interpolation so apostrophes, $, backticks, etc. pass through verbatim.
A trap on EXIT cleans up the tmpfile after the engine call returns.
LAW 7's "MUST contain --plan" self-check guidance and Step 1's invocation
example both updated to reference the file form. Comparison-mode invocation
block updated the same way for --competitors-plan.
Version bump 3.2.2 -> 3.2.3 because this is a behavior change users
running comparison-mode queries will notice (no more "shell quoting error,
retrying" sequences on apostrophe-containing context strings).
12 fixes from the multi-agent code review on PR #400:
Version 3.2.1 -> 3.2.2 across all manifests (SKILL.md frontmatter + body
header, pyproject.toml, .claude-plugin/{plugin,marketplace}.json, sync.sh
cache path). The PR ships observable behavior changes (STEP 0 logic flip,
resolver order change, badge fallback) that should not silently appear
under the same version number — the new fallback reads SKILL.md version
directly so the badge would otherwise be misleading.
render.py::_skill_version:
- `import re` moved to module top
- _VERSION_RE extracted as a module-level compiled pattern that accepts
double-quoted, single-quoted, and unquoted YAML version scalars
- `break` -> `continue` on corrupt manifest, so a corrupt inner manifest
no longer shadows a valid outer one
- Wrap SKILL.md read_text() in try/except for UnicodeDecodeError to keep
badge emission from crashing on mis-encoded SKILL.md
- Docstring clarifies precedence; inline comment marks the fallback boundary
between the manifest walk and the SKILL.md walk
tests/test_skill_version.py (new): 7 unit tests for the fallback paths
(manifest absent, manifest corrupt, corrupt-inner + valid-outer, both
absent, SKILL.md without version, single-quoted, unquoted).
tests/test_plugin_contract.py: tombstone test asserting .codex-plugin/
stays removed (was the only CI guard against accidental reintroduction).
SKILL.md:
- STEP 0 bash echoes CLAUDE_CACHE_LATEST so the model can see the
resolved value when deciding whether to hop
- "Both shapes ship in the wild" comment now names the two cache layouts
(nested {cache}/{version}/skills/last30days/ vs flat {cache}/{version}/)
- Comparison-mode bash invocation gets its own inline SKILL_ROOT resolver
(latent gap: the contract tells the model to skip Step 1 on comparison
queries, so SKILL_ROOT was previously unset there)
CHANGELOG.md: [Unreleased] entries for the resolver rewrite and the
breaking removal of Codex native-plugin support.
All 9 reviewer personas surfaced findings; 3 cross-reviewer corroboration
clusters were promoted (import re, "both shapes" comment, missing fallback
tests). Maintainability follow-up flagged: regex now duplicated across
render.py and 2 test files; could consolidate via shared lib/skill_meta.py
helper in a future PR.
Five provider modules (pinterest, threads, instagram, tiktok, youtube_yt)
and watchlist.py each carried a try/except `requests` import with parallel
urllib + requests branches. The urllib path already used the
stdlib-only wrapper at `lib/http.py` (retries, 429 handling, HTTPError).
This collapses every dual-branch into a single `http.get`/`http.post`
call and removes the `requests` dependency from `pyproject.toml`.
Also drops 4 transitive deps (urllib3, certifi, charset-normalizer, idna)
from the lockfile, leaving the skill stdlib-only at runtime.
Tests for tiktok comments and watchlist delivery were rewritten to mock
`lib.http` directly instead of the now-removed `requests` module.
Out of scope but flagged during review: the 13 surviving SC call sites
share a near-identical scaffold and would benefit from a
`http.scrapecreators_get(url, params, token, ...)` helper. Filed for a
follow-up PR rather than expanding scope here.
Hiten Shah