AGENT SKILL

rebuild-validation

Use when validating that a structured artifact (spec, schema, contract, doc-set, runbook) is sufficient for an independent agent to reconstruct the underlying system from it alone. Triggers on "blind rebuild test", "validate this spec", "is this documentation enough", "spec-driven regression", "contract test", or when shipping any reverse-engineering / spec / schema work. Also covers refactor (the user edited R-/S- entries and asks the agent to splice the diff into an existing codebase - "按新 spec 改", "refactor to spec", "重写这个屎山"). Drives a subagent that has NEVER seen the source artifact to read only the spec and rebuild against a graded machine-tickable two-key grading (functional-checklist.md + test-oracle.md), then surfaces diff-able weaknesses.

License 许可
MIT
最近更新
2026-10-11
Stars
★ 0
Forks
0

01 前言 Front Matter

name
rebuild-validation
author
Hermes Agent
license
MIT
version
1.1.0
metadata
{ "hermes": { "tags": [ "validation", "spec", "contract", "subagent", "regression" ], "related_skills": [ "test-driven-development", "dispatching-parallel-agents", "receiving-code-review" ] } }
description
Use when validating that a structured artifact (spec, schema, contract, doc-set, runbook) is sufficient for an independent agent to reconstruct the underlying system from it alone. Triggers on "blind rebuild test", "validate this spec", "is this documentation enough", "spec-driven regression", "contract test", or when shipping any reverse-engineering / spec / schema work. Also covers refactor (the user edited R-/S- entries and asks the agent to splice the diff into an existing codebase - "按新 spec 改", "refactor to spec", "重写这个屎山"). Drives a subagent that has NEVER seen the source artifact to read only the spec and rebuild against a graded machine-tickable two-key grading (functional-checklist.md + test-oracle.md), then surfaces diff-able weaknesses.

02 资源文件 3

03 SKILL.md 正文

Rebuild Validation

Empirical validation that a structured artifact — a spec, a schema, a contract, a documentation set — is sufficient for an independent agent to reconstruct the underlying thing from scratch.

The signal it produces is binary: PASS / FAIL per checklist item. The output is a bulleted list of items, each marked PASS or FAIL with a one-line reason, plus specific patching suggestions for the author.

When to Use

  • You just wrote a spec / contract / schema and need to know if it's enough.
  • The artifact claims "this is the source of truth for X"; verify the claim.
  • A reverse-engineering pipeline (A → spec) needs a feedback loop.
  • Documentation accuracy: do these docs let a fresh agent complete the task?
  • An API client rebuild task: are these stub docs enough to implement against?
  • Refactor scope: the user edited specs/*.spec.md (added/removed R- or S- entry) and asks you to splice the new behaviour into the existing codebase without rebuilding it from scratch.

Don't use for: unit tests, "smoke tests" of running code, or anything where the artifact isn't a static specification.

Two grading keys, not one

A single functional-checklist.md answers "does the rebuilt code ship the right observable behaviour?". That is not enough.

Every rebuild-grade spec SHOULD also include a test-oracle.md (or equivalent) — a per-symbol white-box grading key:

### `<symbol>`
- input: <literal, byte-exact>
- expects: <literal output> OR <raises ExactException(message)>
- pins: <one-line statement of WHICH invariant this pins>

Why both:

  • functional-checklist = black box. "Does the CLI exit 0 on success, 2 on ValueError?" — visible from outside the binary.
  • test-oracle = white box. "Does greet(" ") raise ValueError before or after .strip()?" — pinned ordering the checklist cannot distinguish. The order is what makes the function correct; without oracle discipline, every rebuild gets it right by coincidence and you only find out at runtime.

What real roundtrips reveal: when both keys are run, the rebuild typically passes 100% of checklist items (the easy grade) but fails 5-15% of oracle items (the order-of-operations, full-width punctuation, error-before-format, dispatch-precedence class). The oracle is what tells you the rebuild got the semantics right, not just the behaviour. A green checklist with a red oracle is a specific failure — list the exact oracle entry and the actual-vs-expected.

This skill consumes the oracle. The companion reverse-engineer-spec skill is responsible for distilling it: one or two fixtures per public surface from the original test suite (NOT copying the whole test file). See references/oracle-distillation.md for the distillation method, and treat that file as the source of truth for "what goes into an oracle entry".

Spec-edit → existing code (refactor)

This skill covers "build a thing from spec" (whole or scoped). The companion edit-an-existing-thing-per-spec path has different discipline and is captured here because the same grading keys apply:

  1. Diff the spec. Identify which R- / S- / oracle entries the user added, removed, or edited. The user's edit is authoritative.
  2. Classify each change into: behaviour change (rewrite affected symbol), conventions change (style / imports / error types), inventory-only (count or wording shifted but no behaviour), or pure clarification (codebase already satisfies it).
  3. Snapshot the source. git worktree add ../snapshot if the tree is clean; otherwise refuse to run. With side-by-side strategy the original directory is the rollback; no separate snapshot needed.
  4. Spawn a leaf rewriter. Constrain its scope to only the affected symbols from step 2. It is forbidden to refactor unrelated code, even ugly code. Patch, do not regenerate.
  5. Run all three grading keys (pytest + functional-checklist + test-oracle). ALL THREE must be green.
  6. Verify the diff matches the symbol list. git diff --stat between snapshot and rewriter output must touch only those files. Anything else is a regression of trust.

Two outcomes worth distinguishing explicitly:

  • All three green AND diff empty → the edit was a pure clarification. Passing outcome, not failure to edit. The user's edit is now a documented intent, pinned by an oracle entry. Show the new oracle entry as proof the clarification is exercised.
  • All three green AND diff non-empty → state the changed symbols and confirm the spec edit is now in the code.

The failure of trust mode: rewriter touches code outside its scope. When git diff shows unrequested changes, roll those back and tighten the rewriter's scope. Otherwise the audit breaks: the user can no longer trust "diff == edit". Diff-vs-symbol-list is the audit backbone of refactor; without it, refactor degrades into vibe-coding that happens to also invoke pytest.

The Five-Step Pattern

1. Write a graded checklist

functional-checklist.md (or equivalent) is plain markdown with bullets:

- [ ] <concrete check 1>
- [ ] <concrete check 2>

Each line MUST be machine-tickable: a concrete command, return value, or observable outcome. "Works" / "looks right" is not enough. Good:

- [ ] greet("Ada", "es") returns "Hola, Ada!"
- [ ] `pytest` exits 0
- [ ] `greeter --help` writes usage to stdout, exit 0

Pre-flight check before you even validate: confirm the artifact itself is ready. The minimum survivable spec has (a) an entry-point document the validator will read first (AGENTS.md or equivalent) and (b) a functional-checklist.md with ≥5 - [ ] items. If either is missing, stop and tell the author — running a validator against a stub spec produces a 0/0 PASS that nobody can act on.

See references/checklist-template.md for a starter.

2. Pick a blind validator

The validator MUST have zero access to the source artifact that produced the spec. Two ways to enforce:

  • Spawn a subagent in its own session; pass the spec path only.
  • Restart the current session so prior context is gone; feed the spec.

The validator MUST NOT use git log, web search, or any debugging aid that could leak the original. If it can leak, the test is invalid.

3. Let the validator pick its own scope

For large artifacts (>30 files or >5k LOC equivalent), the validator may pick:

  • whole: rebuild everything from spec.
  • single-module <name>: rebuild one representative section. Pick the section most users / callers touch.
  • single-file <path>: smallest scoped test.

The validator justifies its pick in one paragraph and proceeds. The author records the choice in AGENTS.md.

Why this matters: forcing whole-artifact validation on a 100k-LOC spec blows budget and produces false negatives. A well-chosen scoped PASS is more useful than a global FAIL with token artifacts.

4. Run the rebuild

The validator works the spec top-down: AGENTS.md → architecture.md → module specs → checklists. Conventions come last. After implementation, the validator ticks every checklist item and reports PASS / FAIL with reasons.

5. Surface diff-able weaknesses

Ask for specific, actionable critique, not narrative review. The prompt format that worked:

Critique the spec: which sections were clear, which were vague, where you had to invent something not in the spec. Name the heading in the spec. Quote the missing text or rule that would have made rebuilding unambiguous.

The author converts each weakness into a patch against the spec (or against the producer skill). Re-run if cheap.

Two distinct signals, not one. Always separate failures from debt:

  • Failures — checklist items the rebuild cannot tick. These mean the spec is missing or wrong on a contract.
  • Spec debt — checklist items the rebuild ticked but only by inventing something not in the spec (LICENSE body, dict-vs-if branch, print() vs sys.stdout.write, etc.). These mean the spec is insufficiently prescriptive. A green build with 4 inventions is a weaker spec than a green build with 0 inventions, even though both report "20/20 PASS".

Always produce the spec-debt list, even on a fully green run.

Strengths and Limits

Strengths:

  • Quantifiable signal: X / N PASS, M with reasons.
  • Catches contract drift, vocabulary mismatches, scope ambiguities, and missing error-message literals that no static review catches.
  • Cheap: a 5-file codebase runs in ~30s of agent time.

Limits:

  • Validator cost scales with spec size — scope must be chosen wisely.
  • Cannot grade non-functional areas (perf, security, aesthetics) without explicit checklist items.
  • A passing rebuild ≠ a correct rebuild. The checklist must be exhaustive; weak checklists pass weak specs.

Common Pitfalls

  1. Leaky validator. If the validator can see the original source (via find /tmp/... or browser history), the test is invalid. Verify isolation by passing only the spec path.
  2. Checklist that grades itself. A checklist with "looks reasonable" or "is well-organized" is unfalsifiable. Every item must be a concrete check.
  3. Forcing whole scope. Letting validators bail to a smaller scope when budget gets tight is correct, not failure. A scoped PASS with reason beats a global FAIL with token noise.
  4. Narrative review. "The spec was good but could be clearer" is not actionable. Demand heading + missing-text-level critique.
  5. No re-run after fix. Patch the spec, but skip re-validation when it's cheap. Same weakness can hide for months otherwise.
  6. Spec drift after the test. If you patch the spec for unrelated reasons, re-run; the build phase will hit drift.
  7. Patching the rebuild until tests pass. The validator's job is to grade the spec, NOT to debug or rewrite the rebuild. A tweaked file that now passes is a false green; the spec still has the gap. The validator must output a graded checklist and stop. If you (the human reading the report) are tempted to fix a failed item by editing the rebuild, that is the bug. Roundtrip tests don't produce fixes; they produce failure signals.
  8. Treating metadata as contract. LICENSE body text, README prose, pyproject.toml description / author fields — none are contract. Let the rebuild invent them. A spec that tries to pin every presentation field produces false-positive spec debt on every run.
  9. Skipping the spec-debt list on a green run. A green run with inventions is a weaker spec than a green run with zero inventions. Always list what the rebuild was forced to invent, even on PASS.
  10. Accepting a green checklist when the oracle is red. The checklist is the easy grade; the oracle is the invariant check. Always run both. A green checklist + red oracle is a specific failure — list the oracle entry, not a vague PASS.
  11. Refactor diffing beyond its scope. When the rewriter outputs a diff that touches more files than the symbol-list predicted, roll the extra changes back and tighten the brief. Otherwise the audit backbone (diff == edit) breaks and refactor degrades into vibe-coding that happens to also pass pytest.

Verification Checklist

  • [ ] Pre-flight passed: entry-point doc exists, functional checklist has ≥5 machine-tickable items.
  • [ ] Validator has no access to the source artifact.
  • [ ] Validator justified its scope choice if it picked <whole>.
  • [ ] Report lists every checklist item with PASS / FAIL / reason.
  • [ ] Spec-debt list is present (not optional), and non-empty when the validator invented anything even on a fully green run.
  • [ ] Critique names spec headings and quotes missing text, not narrative.
  • [ ] No edits were made to the rebuild to mask failures. Validator output is the only output; reading the report, you must not feel tempted to fix the rebuild.
  • [ ] Weaknesses are patched; if cheap, the rebuild was re-run.

One-Shot Recipes

Toy-repo roundtrip (~30s)

# 1. Pick a 5–10 file repo with tests
# 2. Write functional-checklist.md with 10+ items
# 3. Spawn subagent with only spec path; ask it to rebuild at /tmp/rebuild-x
# 4. Subagent runs the checklist, reports PASS / FAIL

Real-world module roundtrip (~5 min)

# 1. Pick a single high-coupling module (e.g. console.py for rich)
# 2. Write specs/<module>.spec.md with SHOULD / MUST + Scenarios
# 3. Subagent picks single-module scope; justifies
# 4. Subagent rebuilds only the chosen surface and grades checklist

References

  • references/checklist-template.md — starter functional-checklist template.
  • references/roundtrip-prompt-template.md — copy-paste subagent goal template, with the whitelist + spec-debt pair that distinguishes a graded rebuild from a vibe-coded one. Pair this with references/checklist-template.md.
  • references/oracle-distillation.md — how to distill a test-oracle.md from an original test suite (one or two fixtures per public surface, NOT a transcription). This is the producer side that reverse-engineer-spec runs; this skill consumes its output.

04 同仓库更多 Skills