git:20260903.2f5affb to git:20260910.a57f970

7 added, 4 removed. Audit A to A.

---

name: spec-fix

description: Streaming light tier for bug fixes and small changes (user-explicit only, never model-initiated). Appends F-N entries to the standing spec/changes/fixes/ batch — locate & confirm before touching code, fix directly when confident or research candidates when not, self-check evidence per entry; the ONE independent audit happens at $spec-ship.

---

# $spec-fix

Task: $ARGUMENTS

The streaming light tier: bug fixes and small changes delivered across conversations, one

F-N entry at a time, into a single standing batch. Tier invariants: a verbatim quote anchor

per entry, honest self-check evidence per entry, and ONE independent verification per

**batch** — at `$spec-ship`, never per entry.

## The standing batch dir

- The batch always lives at `spec/changes/fixes/` with the ledger `fixes/fix.md`

(format → [`../spec-core/references/fix-spec.md`](../spec-core/references/fix-spec.md)).

No naming, no per-batch setup: batch = the dir's lifetime between ships. **spec/ roots at

the directory the session was launched from — never inside a subproject, never in another

worktree's or the main repository's tree** (the ship/archive gates resolve it at that root

only; see the root rule in SKILL).

- No `fixes/` dir → create it with `status: open` and start at F-1.

- **Collision guard**: `spec/changes/fixes/` exists WITHOUT fix.md → that is someone's

normal change dir which happens to be named "fixes" — REFUSE to write into it, report the

collision, and ask how to proceed (rename their change dir, or archive it first). Never

mix a fix ledger into a foreign change.

- - `fixes/` never counts toward the single-active-change slot (gates exempt dirs with fix.md

+ - `fixes/` never counts toward the gates' active-change count (gates exempt dirs with fix.md

present and proposal.md absent) — a full change and the fix stream run in parallel,

neither blocking the other. fix is ungated by design; its protection is the batch audit.

## Size advisory (never a refusal)

Estimate the change first. If it looks full-flow sized (>150 lines / 3+ files / new

dependency / architecture choice), say ONE advisory line — "this looks full-flow sized; a

full proposal is available if you'd rather" — then **proceed anyway**. The size estimate is

reference information for the user, never a gate: routing to the full flow is the user's

call, not the model's.

## Per-entry flow

- 1. **Locate & confirm before touching anything**: find the exact code lines, state the

- problem point and your root-cause reading. The user's ask is quoted **verbatim** into the

- entry (the mini anchor — interpretation happens against the quote, never a paraphrase).

+ 1. **Locate & confirm before touching anything**: consult the knowledge base first — scan

+ `spec/knowledge.md` (the index) and open relevant `[fact]` subdocs before hunting the

+ codebase (a recorded trap or call-chain fact beats a fresh scan; legacy flat file → read

+ whole, declare); then find the exact code lines, state the problem point and your

+ root-cause reading. The user's ask is quoted **verbatim** into the entry (the mini

+ anchor — interpretation happens against the quote, never a paraphrase).

2. **Fork on confidence**:

- Confident and small → implement directly. The Coding Charter binds (Read

`../spec-core/references/code-charter.md` before the first

keystroke — same rule as $spec-apply); the Concerns discipline binds (requirement

silent → most permissive behavior; tightening impulses go to the entry's Concerns,

never into code).

- Uncertain about the right fix → **inline research first**: read the framework source /

official docs / community reports of the same failure, present 2–3 candidates with a

recommendation, implement the chosen one; the chosen rationale lands in the entry's

Root cause. No research.md ceremony at this tier.

3. **Self-check**: run the working checks (compile / focused tests) yourself — commands +

exit codes go into the entry. Anti-Cheating binds: nothing unrun is "success". Do NOT

dispatch a verifier here (terminal-audit doctrine; the user can still demand an immediate

spot-verify on any entry conversationally).

4. **Append the F-N entry** to fix.md (append-only — never renumber or rewrite earlier

entries): verbatim ask, root cause, files touched (+ commit hash when one exists),

self-check evidence, concerns.

5. **Close with the pending count**: end the report with `批次现有 N 条待审` (N = entries

not yet audited by ship). This line is mandatory — it keeps the un-audited backlog

visible every single round.

## Known limitation

One session at a time: two sessions appending fix.md concurrently can lose an entry (the

same exposure every sdd artifact has — single-session operation is the artifact model's

global assumption).

## Anti-patterns

- ❌ Model-initiated fix — user-explicit only; size signals warrant at most a one-line

suggestion (same activation doctrine as the full flow)

- ❌ Editing code before locate-and-confirm (the diagnosis comes first)

- ❌ Paraphrasing the ask in the entry (quote, never paraphrase)

- ❌ Refusing by size (advisory only — routing is the user's decision)

- ❌ Dispatching a verifier per entry (the ONE independent audit is $spec-ship's)

- ❌ Renumbering, rewriting, or deleting existing F-N entries (append-only ledger)

- ❌ Skipping the pending-count closing line