| name | stop-checkpoint |
| description | Force a maintainer checkpoint between artefact (audit table / draft / honest ledger / design-doc revision) and action (commit / merge / push / next step). Shannon authors; Ada invokes. Phase 0A 2026-05-16 NEW skill — closes the autopilot failure mode demonstrated when steps 5/6/7 were committed in sequence without surfacing for review. |
Stop-Checkpoint
This skill exists because the 2026-05-16 Phase 0A session produced
three honest design docs (testing strategy, extractors contract,
validation-debt register) and committed all three to the branch
without surfacing them for maintainer review between commits.
The maintainer corrected the pattern: "I do want to check all these
different documents you're writing as we have been doing so far."
This skill is the operational discipline that prevents the same
failure mode from recurring.
When to Invoke
Invoke before producing any artefact that the maintainer should
review before a commit / merge / push, or before moving to the
next dependent step. Triggers include:
- a validation-debt register update with new
covered / partial
/ blocked rows that change the package's honest scope;
- a decisions log entry ratifying a scope or grammar change;
- an after-phase report (the close gate for any phase boundary);
- a design-doc revision that touches public API, formula grammar,
family list, or extractor contract;
- a rule revision in
AGENTS.md, CLAUDE.md, CONTRIBUTING.md,
docs/design/10-after-task-protocol.md, or any skill file;
- a sequence of related PRs / commits that depends on the
maintainer accepting an earlier link in the chain;
- a
/loop autopilot run that has shipped 2+ artefact commits
without an intervening review touchpoint.
If the artefact already exists (because the autopilot has shipped
it), the skill recommends rollback-or-surface: do NOT silently
proceed to the next dependent step; surface the artefact NOW.
The Discipline
Artefact → checkpoint → action. Never artefact → action directly.
The checkpoint has three concrete moves:
- Surface clickable file links + concise summary of what
the artefact contains and what it implies. Bare paths are
not enough — paste full GitHub blob URLs that open in the
maintainer's browser.
- Name the explicit decision the maintainer is being asked
to make ("OK to proceed to Step N?", "Approve the 102-row
honest tally before AGENTS.md edits?", "Confirm Option C
reading before propagation to 04 / 35?"). Don't just say
"let me know what you think."
- Wait for an unambiguous reply before the action. If
the maintainer says "keep moving" or "yes A and B" or
another approval form, the action is authorised. If the
maintainer asks a clarifying question or surfaces a
contradiction, treat that as a hold, not as approval.
Required Output Format
When invoked, the skill produces a chat message with exactly
four components:
1. Title line: "STOP-CHECKPOINT: <one-line summary of artefact>".
2. Artefact links: full GitHub URLs (one per file), with line
counts and net-change stats.
3. Concise summary: 2-5 bullets per artefact naming the key
claims, the implied downstream actions, and any surprises.
4. Explicit question: "OK to proceed to <next action>?", or
the specific binary decision the maintainer should make.
Optionally append: a brief reminder of which AGENTS.md rule or
design-doc principle the checkpoint enforces (Rule #10
convention-cascade, validation-debt scope-boundary, etc.).
Anti-patterns the Skill Prevents
- Silent self-merge of artefact PRs. Even when CI is green
and the PR is "small", an artefact-producing PR triggers a
checkpoint. Self-merge is reserved for truly mechanical
non-artefact PRs (CI tweaks, dev-log appends following an
already-approved decision, etc.).
- Batching multiple unrelated changes past a single
checkpoint moment. If a session produces three honest
artefacts in sequence, each gets its own surface — not one
batch surface at the end.
- Auto-pilot reasoning from "the user said keep moving" to
"every subsequent action is approved".
keep moving is an
approval to take the next step; it is not a license to
skip future checkpoints.
- Treating a clarifying question as approval. When the
maintainer asks "what does X mean here?" mid-stream, that's
a hold, not approval.
- Surfacing bare paths instead of clickable URLs. The
maintainer cannot review files they can't open. Always
paste full GitHub blob URLs (and raw fallbacks if needed).
Anti-patterns the Skill Does NOT Prevent
- It does not slow down truly mechanical work (typo fixes, CI
cadence, dev-log appends following already-ratified decisions).
- It does not require checkpoints inside a single artefact's
authoring loop (the maintainer doesn't want to review every
edit — just the artefact when it's complete).
- It does not replace
rose-pre-publish-audit (content audit)
or shannon-coordination-audit (cross-team coordination). It
sits between artefact production and the next action.
Worked Example (2026-05-16, the original trigger)
During Phase 0A, the autopilot pattern was:
Step 4 done, commit, push → Step 5 done, commit, push → Step 6
done, commit, push → Step 7 (validation-debt register, the
honest ledger) done, commit, push → started Step 8 (AGENTS.md
edits)
The maintainer caught it before Step 8 landed: "I do want to
check all these different documents you're writing as we have
been doing so far. Do you remember this?"
Under this skill, the correct pattern would have been:
Step 4 done → STOP-CHECKPOINT surface → maintainer approves
→ Step 5 done → STOP-CHECKPOINT surface → maintainer approves
→ ... and so on.
Each artefact is a checkpoint moment. The plan called for an
explicit checkpoint after Step 7 (the register) — the skill
makes that explicit, recurring, and unavoidable.
Cross-References
AGENTS.md Standing Review Roles — Shannon authors this
skill; Ada invokes it.
docs/design/10-after-task-protocol.md — the after-task
protocol consumes this skill's checkpoint moments as
required-section content.
.agents/skills/after-task-audit/SKILL.md — the after-task
audit runs AFTER the checkpoint, on the artefact the
maintainer just approved.
.agents/skills/rose-pre-publish-audit/SKILL.md — the
pre-publish audit can run BEFORE or AFTER the checkpoint
depending on whether content is being authored or finalised.