---
name: ssot-preflight
description: Repository SSOT preflight before substantive code, config, docs, review, debugging, refactoring, or planning work. Use to check SSOT/STATUS.md, open adjudications, documentation language, protocol version, minimal SSOT read routing, and in-task SSOT write triggers. Do not use for pure operations, trivial format-only edits, or non-repository chat.
metadata:
  protocol_version: "2.62"
  bundle: "SSOT Skill"
  semantic_impact: medium
---

# SSOT Preflight

You are at the start of substantive repository work. Your job here is not to
do the work — it is to decide what's safe to read, what's blocked, and what
durable facts you must capture as you go. Do not write code or docs until
this gate clears.

## Clear the gate

Read `SSOT/STATUS.md` first. Treat any substantive change as blocked until
you have answered, in order:

- **Adjudications.** Any `pending` item under `## 开放裁决项` (or legacy
  `## 待裁决项`), or a `deferred` item whose revisit condition has fired,
  blocks ordinary work until resolved or re-deferred.
- **Documentation language.** Read `documentation_language` and its
  evidence; SSOT Markdown you write this task must match it — except for
  paths, commands, identifiers, enum values, API names, and direct quotes.
- **Protocol version.** Compare this file's `metadata.protocol_version`
  against `STATUS.md` `tracked_skill_version`. Project behind -> stop, route
  to `$ssot-audit`. Installed bundle behind project -> report stale install;
  never downgrade the project. If a source checkout under
  `projects/SSOT-SKILL/` is newer than the runtime-installed skill copy
  currently being read (`.agents/skills`, `.codex/skills`, `.claude/skills`,
  etc.), report the install as stale and reinstall from the source checkout
  before relying on new protocol clauses.
- **Freshness floor (conditional).** This binds only where a coverage claim
  exists: any Area Status `covered` / `partial`, or `coverage_result:
  converged`. When it binds, `tracked_commit` must be ancestor-or-equal of
  `HEAD`; a `covered` / `partial` claim whose reviewed baseline is behind
  `HEAD` must be re-confirmed by a scoped self-review within the same task or
  demoted to `stale`. An honest `in_progress` / `bootstrap` / `catching_up`
  repo with no coverage claim owes no freshness proof.

## Route your reads, don't bulk-load

The mandatory floor is `SSOT/STATUS.md` (adjudications, language, version
— cleared above) plus `SSOT/README.md` as the project-specific router. Read
its task-entry map and read only the owners it routes for this task. The
`01-product/README.md` and `02-architecture/README.md` trunks are read only when
the task-entry map routes the task to them — not by default. If the
task-entry map is missing or does not route the current task, fall back to
reading `01-product/README.md` and `02-architecture/README.md` so a thin-router
repo still gets trunk coverage.

Skip this entire skill only for non-repository chat, pure command
execution, or mechanical typo/format edits that cannot alter architecture,
contracts, state, behaviour, workflows, tests, doc truth, or external
surfaces.

## Capture durable facts as they surface

While working, watch for facts that outlive this task — new/removed APIs,
schema, contracts, lifecycle, trust boundaries, deployment policy, test
policy, confirmed root causes, user-locked decisions, product promises,
"do not revive" paths, repeated-failure-derived agent discipline. When one
appears, either update the unique authority immediately (when clear), or
write a short delta and resolve it at `$ssot-closeout` before final
response, `claim_done`, or commit. Detail (including the `Rule / Trigger /
Why / Evidence / Failure-mode` discipline schema) lives in
`references/area-model.md §2.4` and in `$ssot-closeout` references — not
here.

## Surface task-relevant open risks

Before implementation or final planning, scan the owner files the router
actually routed the task to, plus `STATUS.md ## Open Gaps`. Report the
entries whose trigger, path glob, owner, capability, journey, runtime owner,
command, or failure mode overlaps the task. Cold-scanning the full
`04-records/tech-debt/`, `04-records/bugs/`, or `04-records/gotchas/` directories is the agent's call, not a
default obligation — do it when the routed owners or the task shape make it
worthwhile. These are recommendations, not automatic blockers: each one must be
handled during the task by fixing it, explicitly deferring it with the owner and
reason still valid, or naming the next action in closeout.

Classify each overlapping entry before work starts:

- `fix-now`: current facts are wrong, misleading to the agent, or block
  closeout alignment.
- `recommend-now`: low-cost, high-confidence repair that should be handled in
  this batch.
- `defer-visible`: not blocking this batch, but deferral must name the owner or
  owner record, reason, closure condition, revisit signal, verification guard,
  and next concrete action.
- `ignore-for-scope`: no task overlap after inspection; name why it is outside
  this batch.

`defer-visible` is a floor, not an excuse to disappear work. Vague wording such
as "later", "someday", "future work", or a locked-language equivalent of
"handle later" is invalid unless the same record or STATUS row carries the
owner/record pointer and the retrigger/guard that makes the deferral visible to
the next agent.

Do not treat downgraded working docs or historical source material as current
open risk unless STATUS or an owner entry promotes the risk. Do not hide a
fallback, compat shim, temporary workaround, TODO/FIXME/HACK, later-remove note,
or waiver behind ordinary prose; it must have a registered owner, reason,
closure condition, revisit signal, and verification guard in `04-records/tech-debt/`,
`04-records/bugs/`, `04-records/decisions/`, or `STATUS.md` as appropriate.

## Park user-directive signals

Watch the user's prompts for an explicit lift-or-kill directive
("promote this", "never again"), a repeated same-kind correction across
turns, or recent prompts clustering in one domain. When one surfaces
and the rule is not already a clear durable fact you can write into the
unique authority per the section above, park one `CAP-` row
(`signal_source: user-directive`, `status: open`) in
`SSOT/STATUS.md ## Pending Captures` and let `$ssot-closeout` or
`$ssot-audit` route it. Do not promote, do not edit any other SSOT
file. Detail on the two evidence streams and the move schema lives in
`$ssot-closeout references/promotion-rationale.md`.

## When this skill is wrong, route out

| Situation | Skill |
|---|---|
| No `SSOT/` yet, `.bootstrap/` present, or `coverage_result: bootstrap` | `$ssot-bootstrap` |
| About to commit, `claim_done`, or finalise after substantive change | `$ssot-closeout` |
| Catch up commits / sessions, or protocol version is behind | `$ssot-audit` |
| Health check, independent stop review, CORE-REF / ADAPTER / CONSUMPTION | `$ssot-doctor` |
| Legacy `$ssot-skill` mention | route via `ssot-skill` shim back here |

## Load on demand

| When the task hits | Read |
|---|---|
| STATUS coverage states, adjudications, gaps, tracking baselines, stop gates | `references/status-protocol.md` |
| README/docs/ADR/runbook/PRD/core-ref classification and absorption | `references/source-material.md` |
| Candidate / hypothesis / source-backed knowledge | `references/knowledge-integrity.md` |
| Architecture root/views/domains, Reader Map, decomposition, coverage depth | `references/architecture.md` |
| Top-level area responsibilities and task-entry mapping rules | `references/area-model.md` |
| Writing or reviewing any reader-facing SSOT body; product/architecture use full task-based acceptance and process/records/glossary/root/STATUS use exact-scope acceptance | `references/reader-quality.md` |

## Things this gate must not be used for

- A second code/schema/router source. Current implementation comes from
  code, config, schema, tests, and actual behaviour — not SSOT.
- Copying full source documents into SSOT. Absorb durable facts into one
  authority with evidence; do not mirror.
- Carrying candidate or hypothesis claims inside authoritative architecture
  bodies.
- Self-advancing `tracked_commit`, `tracked_session`,
  `tracked_skill_version`, language lock, bootstrap `passed`, or
  `converged` without the required independent review.
- Re-expanding this preflight file with bootstrap, doctor, audit, or
  closeout procedures. Route to those skills.