Skip to content

Add repo-pincer skill (codebase reverse-engineering)#27

Merged
Hefrock merged 1 commit into
mainfrom
claude/festive-cori-wd84yr
Jul 14, 2026
Merged

Add repo-pincer skill (codebase reverse-engineering)#27
Hefrock merged 1 commit into
mainfrom
claude/festive-cori-wd84yr

Conversation

@Hefrock

@Hefrock Hefrock commented Jul 14, 2026

Copy link
Copy Markdown
Owner

New skill: repo-pincer — reverse-engineers a codebase by reconciling top-down claims against bottom-up reality, closing on the discrepancies like a pincer.

Why a separate skill, not folded into agent-eval

agent-eval grades an agent's output; repo-pincer audits a codebase's internal consistency (docs vs. implementation) — different subject matter, different trigger phrases. It also has less structural coupling to agent-eval than agent-redteam does (no shared scoring engine, no JSONL output), and agent-redteam stayed a sibling despite that coupling. Folding this in would have muddied agent-eval's domain-neutral trigger-matching and pushed it well past its ~500-line budget for no real benefit.

The methodology

Two independent passes that must reconcile — the discrepancy between them is the actual finding, not either pass alone:

  1. Top-down (claims) — README → docs → manifests → API/CLI surface. Builds a claims ledger: what's asserted, where, how falsifiable.
  2. Bottom-up (reality) — starts from entry points (exported functions, CLI commands, error paths), reads actual behavior, traces the call graph only as far as needed to check a specific claim.
  3. Reconciliation — every claim classified Confirmed / Drift (was true, isn't anymore) / Aspirational (documented, never built) / Silent (implemented, undocumented), ranked High/Medium/Low by whether it sits on a write path or safety boundary. The ranked list is the deliverable.

Scope

/pincer [path or repo] [--depth quick|standard|thorough] — defaults to a single module, not a whole-repo scan.

No scripts — methodology only

Same shape as the wiki skills, not agent-eval/agent-redteam/deid-reid-harness: its tools (Glob/Grep/Read, the Explore agent) are already general-purpose, so there's nothing to compute or ship as code.

Integration — zero changes to existing skills

If a vault is connected, output optionally writes to Sources/raw/<repo>-pincer-<date>.md — the existing raw-capture convention. wiki-synthesizer/wiki-operator /source picks it up unchanged. Runs fully standalone otherwise; no MCP dependency.

Files

  • skills/repo-pincer/SKILL.md — the skill (77 lines)
  • .claude-plugin/marketplace.json — registered under agent-design, grouped with agent-eval/agent-redteam/deid-reid-harness
  • README.md — skill table + repo structure tree

Docs-only skill, no runnable code — marketplace JSON validated (8 plugins, correct category grouping).

🤖 Generated with Claude Code


Generated by Claude Code

New skill under agent-design, sibling to agent-eval/agent-redteam —
methodology-only, no scripts, since its tools (Glob/Grep/Read, the Explore
agent) are already general-purpose. Same shape as the wiki skills: pure
SKILL.md, no computation to ship.

Two independent passes that must reconcile, closing like a pincer:
- Pass 1 (top-down): claims ledger from README/docs/manifests/API surface
- Pass 2 (bottom-up): as-built model from entry points through the actual
  implementation, traced only as far as needed to check a specific claim
- Pass 3 (reconciliation): every claim classified Confirmed / Drift /
  Aspirational / Silent, ranked High/Medium/Low by whether it sits on a
  write path or safety boundary. The ranked discrepancy list is the
  deliverable, not either summary alone.

/pincer [path or repo] [--depth quick|standard|thorough] — scoped to a
single module by default, not the whole repo.

Runs standalone (no MCP/wiki dependency). If a vault is connected, output
optionally compiles into Sources/raw/ via the *existing* wiki-synthesizer /
wiki-operator /source pipeline — zero changes needed to either skill.

Registered in marketplace.json (agent-design category) and README (skill
table + repo structure tree).

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013H1oBwjtpzDLGreGTTjoP7
@Hefrock
Hefrock merged commit 69a5531 into main Jul 14, 2026
4 checks passed
@Hefrock
Hefrock deleted the claude/festive-cori-wd84yr branch July 14, 2026 18:01
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants