Audit Report — 2026-07-10

[seankoji]

Full audit run of 2026-07-10 (skill: /seankoji:audit) against HEAD 4753ad9. Eight domain agents (code, security, CI, docs, tests, dev, claude-config, usage-analysis); every high-severity and every measurable claim re-verified in the main loop before filing. Prior audit (#13–17, closed 2026-07-02) fully dedup'd — nothing here restates it.

Severity breakdown

Severity	Count	Filed as
High	1	#61
Medium	13	#61–#66
Low	7	#61–#66
Info / confirmed-clean	6	this report only

Domain breakdown

Domain	Findings	Highlights
security	1 + 2 clean	version-bump.yml fork guard; shell quoting + fail-closed discipline confirmed clean
ci	4 + 2 clean	no shellcheck; no frontmatter validation; 5-item checklist fully CI-enforced ✅; CI green ✅
docs	4	elephant.md drift (3 verified instances); CHANGELOG missing #55–59
tests	6 + 1 clean	2,545 LOC of JS/Python with zero coverage; 16/16 existing tests pass ✅
dev	3	ape README prerequisites; CONTRIBUTING grep false-positives
usage	2	audit.jsonl telemetry never fires in production; ollama-sidecar diagnostic gap

Top 5 priorities

1. Audit findings: marketplace-level / cross-plugin (2026-07-10) #61-1 — audit.jsonl telemetry never fires. The cross-plugin scheme AGENTS.md centers on has produced zero lines in production, ever, despite heavy real-world use of all three participating commands (115+ /imps runs alone). The script works; the call sites are skippable tail prose. Everything AGENTS.md promises downstream (the meta-command) is blocked on this.
2. Audit findings: marketplace-level / cross-plugin (2026-07-10) #61-2 — version-bump.yml same-repo guard (xs effort). A one-line if: closes a write-capable path that currently depends on a mutable GitHub org setting staying at its default.
3. Audit findings: marketplace-level / cross-plugin (2026-07-10) #61-6 — regenerate elephant.md. Three verified drift clusters (plugin count 4→6, agy→gemini CLI rename, dropped acronym frameworks). Dogfood /elephant-goldfish:elephant reconcile.
4. Audit findings: elephant-goldfish (2026-07-10) #64-1 — goldfish-judge.sh end-to-end fixtures. The marketplace's fail-closed reference implementation has its exit-code contract untested; the harness was built for exactly this.
5. Audit findings: marketplace-level / cross-plugin (2026-07-10) #61-5 — CHANGELOG for PRs docs(ollama-sidecar): reframe README/tool description as "when jq can't" #55–59 (xs). Five refactors invisible to users.

Findings checklist

🔴 High

• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (1) Cross-plugin audit.jsonl telemetry never fires in production

🟡 Medium

• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (2) version-bump.yml lacks same-repo guard on contents:write
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (3) No shellcheck in CI
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (4) No commands/agents frontmatter validation
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (5) CHANGELOG missing PRs docs(ollama-sidecar): reframe README/tool description as "when jq can't" #55–59
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (6) elephant.md drift (regenerate)
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (7) tests/run.sh hardcodes audit-log consistency plugin list
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (8) CI executable check misses *.py; modes inconsistent
• Audit findings: imps (2026-07-10) #62 (1) imps-run.workflow.js (787 LOC) zero coverage
• Audit findings: ape (2026-07-10) #63 (1) ape README missing Prerequisites
• Audit findings: ape (2026-07-10) #63 (2) ape-forage.workflow.js (326 LOC) zero coverage
• Audit findings: elephant-goldfish (2026-07-10) #64 (1) goldfish-judge.sh no end-to-end test
• Audit findings: ollama-sidecar (2026-07-10) #65 (1) ollama_sidecar.py (1432 LOC) zero coverage
• Audit findings: ollama-sidecar (2026-07-10) #65 (2) ollama status/keep-warm diagnostic (usage-driven)
• Audit findings: claude-tuneup (2026-07-10) #66 (1) scan_perms.py (190 LOC) zero coverage

🔵 Low

• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (9) CONTRIBUTING grep check → 52 false positives
• Audit findings: marketplace-level / cross-plugin (2026-07-10) #61 (10) Actions not SHA-pinned
• Audit findings: imps (2026-07-10) #62 (2) imps-intro.py exec bit
• Audit findings: imps (2026-07-10) #62 (3) imps.md 687 lines (operator decision)
• Audit findings: ape (2026-07-10) #63 (3) exact-match fixtures fragile
• Audit findings: ollama-sidecar (2026-07-10) #65 (3) ollama-sidecar README missing ## Usage header
• Audit findings: claude-tuneup (2026-07-10) #66 (2) settings.json.bak pruning

ℹ️ Info / confirmed clean (no issues filed)

• Shell scripts across all plugins: consistent quoting, no eval, jq via --arg — no injection vector found.
• Fail-closed (goldfish-judge.sh) / fail-soft (audit-log.sh) split exactly matches the AGENTS.md invariant, including malformed-args-still-exit-1.
• CI fully enforces the 5-item add-a-plugin checklist; recent runs green; tests 16/16.
• Two locked worktrees under .claude/worktrees/ — left untouched; prune manually if truly stale (locked may mean in use).

(The audit also produced an operator-private usage analysis — permission-scope and installed-plugin recommendations — delivered directly to the operator, not published here.)

---

Gate: none (full pipeline). Swarmable set #61–#66 handed to /imps after this post. Issues bundled per-plugin per operator preference.