audit-docs
Audit cross-document coherence: docs ↔ roadmap ↔ code ↔ fix index ↔ issues. Finds drift — features in docs/ not in the roadmap (or vice versa), fix-index…
Internal contract: the single owner of the bounded pre-write implementation discovery step. One read-only mapper that closes seven evidence questions, emits one fixed compact map, and routes READY | REPLAN | NEEDS-DESIGN | BLOCKED before any branch, planning, or source write.
$ npx -y skills add gtrabanco/agentic-workflow --skill implementation-discovery --agent claude-codeHow it fires
How this skill gets triggered: by you, by Claude, or both.
/implementation-discoveryContext preview
The summary Claude sees to decide when to auto-load this skill.
Internal contract: the single owner of the bounded pre-write implementation discovery step. One read-only mapper that closes seven evidence questions, emits one fixed compact map, and routes READY | REPLAN | NEEDS-DESIGN | BLOCKED before any branch, planning, or source write.
name: implementation-discovery user-invocable: false version: 1.1.0 author: "Gabriel Trabanco <gtrabanco@users.noreply.github.com>" license: MIT description: > Internal contract: the single owner of the bounded pre-write implementation discovery step. One read-only mapper that closes seven evidence questions, emits one fixed compact map, and routes READY | REPLAN | NEEDS-DESIGN | BLOCKED before any branch, planning, or source write. Consumed only by execute-phase on the pre-write route. Not a menu entry.
The one authoritative owner of the bounded, read-only pre-write mapper consumed by `execute-phase`. A phase does not touch the repository until this step has closed its questions and said `READY`. Nothing else may define what a valid pre-write discovery pass is.
dependency/status/acceptance/phase/feature-28 receipt checks.
orientation pass and never a substitute for planning.
The mapper receives only:
affected decision/obligation, freshness, and declared unknowns;
progress receipt;
Semantic/symbol navigation and Engram/memory tools are optional adapters that may locate evidence faster; they are never authority. Direct repository/Git/search/ read/test fallback must satisfy the identical evidence contract, and only repository bytes answer a map field. The tool that found a citation does not excuse a missing `path:line` or a contradictory claim.
Every pass must answer — each answer citing repository-relative `path:line`, symbol, or test plus exact source revision / content identity:
1. Which entry points and public/internal interfaces own current behaviour? 2. Which callers, adapters, roles, compatibility surfaces, and failure paths can this phase affect? 3. Which helpers, patterns, decisions, and invariants constrain the change? 4. Which tests, fixtures, probes, and production-like scenarios establish current behaviour? 5. Which exact writes are expected, and which reviewed obligation does each serve? 6. Which carried planning-evidence claims and Plan assumptions are directly confirmed, refined, stale, missing, or contradicted by source? 7. Which relevant unknowns remain, who owns them, and what evidence resolves them?
The mapper records conclusions, not every search. Repeating a search/read requires a new question, changed source, or insufficient cited evidence; otherwise it stops (no-progress).
**Inline** is permitted only when behaviour is localized/familiar, targeted evidence answers every question, no public/persistence/security/recovery/ compatibility boundary may change, and the executor can finish the map before writing.
**A fresh read-only mapper** is required for cross-module/layer or unfamiliar work, public/persistence/security/recovery/compatibility impact, competing project patterns, an unproven material Plan assumption, prior failed attempts/review bias, or any case where editing would begin before the map closes. Same-model fresh context is useful but is not model diversity. Manual sequential fresh conversations are the portable fallback.
No file-count threshold participates in the route or the verdict.
IMPLEMENTATION MAP — unit-id phase-id Map revision: opaque single-consumption id Source identity: HEAD + clean-source proof + cited-evidence manifest digest Authority: SPEC receipt + Plan receipt + phase fingerprint Planning evidence: carried row ids + current confirmation/refinement/conflict Obligations: ordered acceptance/fix/obligation ids Entry points: path:line or symbol + role + exact evidence Affected surfaces: callers/adapters/roles/compatibility/failure paths + evidence Current behaviour: evidence-backed summary Reuse and constraints: helper/pattern/decision/invariant/fixture + evidence Expected writes: ordered path + obligation served + reason Validation: falsification probe + TDD target + phase gate Plan assumptions: confirmed items + evidence Contradictions: none or exact Plan/source conflict Unknowns: none or exact missing fact + owner + resolution evidence Decision: READY | REPLAN | NEEDS-DESIGN | BLOCKED
`READY` requires every field, question, and owned obligation covered, no material contradiction/unknown, an observed falsification probe, and expected writes inside phase authority. It also requires every carried planning-evidence row to be confirmed or narrowly refined.
`REPLAN` means source disproves Engineering assumptions while Product remains stable, or Plan-level topology/architecture/obligation/validator evidence is absent.
`NEEDS-DESIGN` means Product, acceptance, authority, or architecture intent is missing or conflicting.
`BLOCKED` names evidence that cannot currently be obtained.
No other verdict exists.
A verdict routes, it never repairs. Missing/contradicted Plan-level topology, architecture, obligations, validators, or assumptions return to `plan-feature`/`plan-fix` plus a fresh `review-plan`. Product/acceptance/ authority/architecture gaps return to `design-feature` then `review-spec`. A verdict with unavailable evidence is one exact `BLOCKED` prerequisite. Mapping does not decide code correctness — candidate review, verification, and audit au
A reusable set of agent skills that run a disciplined, doc-driven workflow for building software with agents — from idea/issue to a reviewed, classified, merge-ready change.
Repo: gtrabanco/agentic-workflow
Audit cross-document coherence: docs ↔ roadmap ↔ code ↔ fix index ↔ issues. Finds drift — features in docs/ not in the roadmap (or vice versa), fix-index…
Audit a whole PR against the delivery contract and return MERGE-READY or evidenced blockers with the full URL. Consumes the current review-change REVIEW-PASS…
Turn a raw idea or existing feature into a designed product SPEC by completing entity, integration, role, and expectation closure. Upserts never destroy…
Discover repository evidence and write a frozen Normalized Repository State. Produces verified repository evidence and keeps facts, decisions, planned work,…
Internal shared owner of evidence-grounded authoring: the fixed claim/authority/evidence/freshness/unknown row, the ordered inventory-evidence-draft-readiness…
Implement all remaining phases of a planned feature/fix by default, or one explicit P<n>, with frozen acceptance, phase-local gates, commits, recovery, and…