Using Harness Anchor

Using harness-anchor

You have the harness-anchor runtime layer, companion to superpowers (process methodology): it anchors state persistence, scope boundaries, verification gates, and C/C++ engineering.

Windows session (Git-Bash): plugin scripts already handle path and JSON-engine portability — do not hand-"fix" MSYS paths or shell out to PowerShell. C/C++ caveat: TSan/LSan are unavailable on this platform; read the cpp-sanitizers skill's Windows platform notes before trusting sanitizer coverage.

Priority Order

1. User's explicit instructions (CLAUDE.md, direct requests) — highest
2. superpowers skills (process)
3. harness-anchor skills (state/scope/verification/cpp)
4. Default system prompt — lowest

On apparent conflict, prefer superpowers' process and use harness-anchor to anchor state/evidence around it.

Project Navigation (read these BEFORE acting)

If they exist in the project root, read in this order:

1. AGENTS.md or CLAUDE.md — operating manual
2. feature_list.json — active feature + scope + done criteria
3. PROJECT-TOC.md — file-level index, one line per file
4. golden-rules.md — project taste / anti-pattern rules to honor
5. session-handoff.md — what previous session left

If none exist, the project is un-anchored. Suggest the user run /anchor to scaffold.

Sibling Skills (load when triggered)

• project-indexing — need to find files; PROJECT-TOC.md before Glob
• feature-state-keeper — feature work starts/finishes; updates feature_list.json + progress.md
• init-verification — at start of work; run init.sh health check
• self-correction-loop — after edits that fail lint/build; iterate with evidence
• anti-hallucination-gates — before claiming "done"; Default-FAIL evidence contract
• test-coverage-design — what to test / is coverage real before "done"; dispatches coverage-analyst
• capturing-golden-rules — a mistake recurs / a review comment is really a convention; encode it in golden-rules.md
• context-budget-discipline — context fills up; SELECT/WRITE/COMPRESS/ISOLATE
• docs-lookup — unfamiliar tools/APIs/errors; Context7 → WebSearch → calibrated-uncertainty chain

• cpp-build-systems — CMake/Meson/Make/Bazel project
• cpp-static-analysis — clang-tidy / cppcheck / IWYU
• cpp-formatting — clang-format
• cpp-sanitizers — ASan / UBSan / TSan / valgrind

Hard Rules

1. No "done" without evidence — and no "absent" without scope. "pass" requires compile + tests + lint with concrete output paths; lacking evidence → status stays in-progress. The same rule runs in reverse: a claim that something is missing (tool not installed, no such function) must state what you searched and when — searched <scope>, not found (as of <date>). Express uncertainty: "I am uncertain whether X passes because ."
2. One active feature at a time unless explicit multi-feature plan. Unrelated mid-feature request → surface the scope-jump and confirm before pivoting.
3. State lives on disk, not in chat. Update feature_list.json / progress.md / session-handoff.md.
4. PROJECT-TOC.md before Glob.
5. Subagents are single-level. Never invoke a subagent from within a subagent — a dispatched worker must not run the subagent-backed gates /verify · /test-plan · /gc; the parent runs them after workers integrate.
6. When stuck, follow docs-lookup. Never guess at unfamiliar tools/APIs/errors.

Interop seams with superpowers (contract in the named skill): parallel/subagent dispatch → feature-state-keeper; worktree setup → init-verification; sensor-finding triage → self-correction-loop (+ receiving-code-review).

Commands — recommend them at the right moment

Commands never auto-trigger (unlike skills). Surfacing the right one — with a one-line why — is your job.

• /anchor — project un-anchored (no feature_list.json): scaffolds the state files.

• /cpp-init — C/C++ project with no .clang-format/.clang-tidy (right after /anchor): drops C/C++ config + sanitizer build.

• /index-project — PROJECT-TOC.md absent or stale, or before a broad file search: (re)builds the index.
• /verify — before claiming a feature passes or flipping status to pass: build + lint + tests in a fresh-context subagent (--fix = bounded ≤2-cycle auto-fix).
• /test-plan — after implementing, before marking pass: fresh-context coverage-gap analysis. Complements /verify.
• /gc — after a batch of generated code, before /session-end: fresh-context drift/entropy scan vs golden-rules.md + slop heuristics. Read-only (not git gc).

• /sanitize — after a C/C++ source change, or before merging C/C++: ASan+UBSan run (TSan separately) with an evidence log.

• /status — "where am I / what's the state": read-only snapshot, writes nothing.
• /session-end — at a stopping point: writes handoff, appends progress.md, offers over-budget state archival + a commit.

When to NOT use this

• Inside a subagent dispatched for a narrow task (its own prompt governs)
• When the user has explicitly disabled harness-anchor (CLAUDE.md override)