speckit-evidence-audit
Merge-gate audit: synthetic propagation + diff-scan. Hard-blocks on either.
- License 许可
- MIT
- 最近更新
- 2026-06-16
- Stars
- ★ 2
- Forks
- 0
01 前言 Front Matter
- name
- speckit-evidence-audit
- metadata
- { "author": "github-spec-kit", "source": "evidence:commands/speckit.evidence.audit.md" }
- description
- Merge-gate audit: synthetic propagation + diff-scan. Hard-blocks on either.
- compatibility
- Requires spec-kit project structure with .specify/ directory
03 SKILL.md 正文
/speckit.evidence.audit
Produce a merge-readiness verdict for the current feature. Combines two signals:
- Task graph (via
speckit.evidence.graph). Any[S]or[S*]task counts against merge-readiness. - Diff scan — greps
git diff <base>...HEADagainst the default pattern library inaudit-patterns.yml. Block-severity hits count against merge-readiness. Advisory-severity hits print but do not block.
How to invoke
./fake.sh build -t EvidenceAudit
The merge-gate audit computes in-process in compiled F#
(FS.Skia.UI.Build.Evidence.Engine.runAudit); there is no Python or shell audit
runner. The gate reads tasks.md / tasks.deps.yml / readiness/ and the
unified git diff (resolved at the build.fsx edge) and writes the audit
artifacts (seh-audit-summary.json, the scan hit files, diff-scan-hits.json).
The feature-base ref is auto-detected (main / master); the diff-scan reads
the retained audit-patterns.yml data file. --accept-synthetic is logged but
never changes the verdict (Principle V).
When it runs
- Automatically as the
after_implementhook declared in the evidence extension'sextension.yml. - Manually any time the user wants a readiness snapshot.
Exit codes
0— PASS. No synthetic tasks, no blocking diff-scan hits.2— NEEDS-EVIDENCE. At least one blocking signal. (Still the exit code when--accept-syntheticis used.)3— graph compute failed (cycles, dangling refs). Fix the graph first.4— usage error.
Strictness model
The audit is configured block on both: any remaining [S] or [S*]
AND any block-severity diff-scan hit are hard gates. The --accept-synthetic
flag is the only way past; it requires written justification and is logged.
Advisory-severity diff-scan hits are informational only (the synthetic-banner
pattern is intentionally advisory — seeing SYNTHETIC: comments is proof that
Principle V disclosure is happening).
When you see NEEDS-EVIDENCE
Walk the report top to bottom:
- Declared
[S]tasks — can any be upgraded to[X]by swapping in real evidence? If yes, update the task, fix the code, re-run. If no, confirm the Synthetic-Evidence Inventory row is current. - Auto-propagated
[S*]tasks — these clear automatically once their root-cause[S]upstreams clear. Check the root-cause list inreadiness/task-graph.md. - Blocking diff-scan hits — each hit names a file, line, pattern id,
and reason. Either fix the code (preferred) or, if genuinely a false
positive, extend the whitelist in
audit-patterns.ymlwith a targetedfile_globorline_regex. - If merging now is unavoidable (staged rollout, upstream dependency not
ready), use
--accept-synthetic "written reason"— the documented escape hatch, not a bypass. The justification lives inreadiness/synthetic-evidence.jsonand SHOULD be mirrored into the PR description.
Output
specs/<FEATURE_ID>/readiness/task-graph.{json,md}— refreshed by the graph compute step.specs/<FEATURE_ID>/readiness/diff-scan-hits.json— structured diff findings (blocking + advisory).specs/<FEATURE_ID>/readiness/synthetic-evidence.json— written only when--accept-syntheticis used.
Authoritative status region (spec 037, US2)
Machine-readable status values are read only from a fenced code block whose
info string is exactly audit-status. Prose, markdown bullets, and any other
fenced block are never read as status, so a blocker term inside explanatory text
or a negation cannot raise a false block (FR-004, FR-005).
Deterministic resolution rule:
- First region wins — the first
audit-statusregion that declares a key provides its authoritative value. - Duplicate key within the region is a parse error — never silent last-wins.
- Prose never wins — a key in prose/bullets/other blocks is ignored.
- Malformed entry (missing
=, empty key) is a parse error — never silently treated as passing or failing.
Blocking is structured, not substring (FR-006): the audit blocks on explicit
violating values (exact-package-match not in {true,yes},
package-resolution=nu1603, taskbar-only=true, or taskbar-entry=true with
window-visible=false) — never on substring presence of taskbar-only /
mismatch / nu1603 in text. Scanner:
.specify/extensions/evidence/scripts/python/audit-status-scan.py.
Sequential FAKE Commands
FAKE-backed commands (./fake.sh, fake.cmd, or dotnet fake) share
repository .fake state and are not safe to run concurrently. Non-FAKE audit
file reads may run in parallel when they do not depend on .fake, but graph
and audit targets must run sequentially:
./fake.sh build -t EvidenceGraph./fake.sh build -t EvidenceAudit
04 同仓库更多 Skills
fs-skia-ui-widgets
Generated product guidance for Skia-rendered FS.Skia.UI Controls, rich text, chart controls, graph controls, DataGrid, and custom wrappers.
fs-skia-testing
Assert generated-product expectations and evidence in a governed FS.Skia.UI product.
fs-skia-skiaviewer
Wire a generated FS.Skia.UI product to the desktop viewer host.
fs-skia-scene
Build pure scene descriptions in a generated FS.Skia.UI product.
fs-skia-keyboard-input
Map keyboard input to product commands in a generated FS.Skia.UI product.
fs-skia-elmish
Drive a generated FS.Skia.UI product through the pure Elmish adapter.