checkpoint

name	checkpoint
description	Writes and resumes a self-contained progress checkpoint (story, branch, tasks done, decisions, remaining todos) so work survives a context reset. Invoke directly to save or resume progress ('save our progress on #313', 'resume story #256'); composed by a future closing phase of /implement and consumed by /publish-pr (via $mode=resume).
version	0.4.1
author	Foomakers

/checkpoint — Resumable Progress State

Write and resume a self-contained progress checkpoint so a fresh session (or a subagent) can continue a story exactly where the previous one stopped. Follows the checkpoint template — five sections: story, branch, tasks done, key decisions, remaining todos.

Arguments

Argument	Required	Description
$mode	Yes	write — persist/return current state. resume — locate and parse the existing checkpoint for a story.
$story	No	Story ID. If omitted, detected from session context or the current branch (feature/#<id>-*). If it cannot be detected → HALT.
$persist	No	Write mode only. true (default) — write/update the file at the default location. false — write-free: synthesize and return the checkpoint text without touching the filesystem; the composer decides where it lands.

Core Rules

• One file per story: default location .pair/working/checkpoints/<story-id>.md. Write mode always updates it in place — never a second file for the same story. The file is the source of truth; an optional issue-body mirror must be explicitly marked as a copy.
• Task-scoped: a downstream handoff consumed only by a session or subagent resuming or continuing that story — NEVER loaded as ambient context into unrelated activity.
• Write freely, clean up on completion: writing — including automatically, e.g. between tasks — is fine; the constraint is on consumption scope, not write timing. When the story is Done (closing phase or PR merge), the checkpoint is removed so finished-story state never lingers.

Algorithm

Step 1: Resolve Story Context

1. Check: Is $story provided?
2. Skip: If yes, proceed to Step 2.
3. Act: Attempt detection, in order:
   • Active story already loaded in the current session.
   • Current branch name matching feature/#<id>-* (git branch --show-current).
   • Resume mode only: a single checkpoint file present under .pair/working/checkpoints/.
4. Verify: Story ID resolved. If not → HALT: "Cannot detect story context — pass $story explicitly."

Step 2: Route by Mode

• $mode = write → Step 3.
• $mode = resume → Step 6.

Step 3: Gather State (write mode)

1. Check: Is the full state (story, branch, tasks done/pending, decisions) already known from the current session context?
2. Skip: If yes, use it directly. Proceed to Step 4.
3. Act: If not (e.g., invoked by a subagent with no prior context), reconstruct from artifacts:
   • Branch: git branch --show-current.
   • Tasks done/pending: read the story's Task Breakdown checklist (per way-of-working.md), cross-referenced with commits on the branch — the same technique /implement uses for idempotent resume.
   • Decisions: scan adoption/tech/adr/ and adoption/decision-log/ for files touched since the branch diverged from main.
4. Verify: All five state elements resolved. Anything that cannot be reconstructed confidently is marked [unknown — needs confirmation] — never guessed.

Step 4: Detect Existing Checkpoint

1. Check: Does .pair/working/checkpoints/<story-id>.md already exist?
2. Skip: If not, proceed to Step 5 to create it.
3. Act: If it exists, read its current content — the new state overwrites it in place (Core Rule).
4. Check: Do other checkpoint files also resolve to this story ID (e.g. a stray duplicate from a prior manual copy)?
5. Act: If duplicates are found, use the most recently modified one as the base and flag the others to the caller as duplicates to remove.

Step 5: Write / Return Checkpoint

1. Act: Render the checkpoint following the checkpoint template, filling all five sections.
2. Check: Is $persist false?
3. Skip: If $persist is true (default) — create .pair/working/checkpoints/ if missing, write the rendered content to <story-id>.md (overwrite in place, per Core Rule).
4. Act: If $persist is false — do not touch the filesystem. Synthesize the text only.
5. Verify: Regardless of $persist, the full rendered checkpoint text is returned to the caller (Output Format). If an issue-body mirror is requested, mark it explicitly: "Mirror of .pair/working/checkpoints/<story-id>.md — file is the source of truth."

Step 6: Locate Checkpoint (resume mode)

1. Check: Does exactly one .pair/working/checkpoints/<story-id>.md exist?
2. Skip: If exactly one, proceed to Step 7.
3. Act: If none exists → HALT: "No checkpoint found for #$story — start fresh, or provide checkpoint text directly."
4. Act: If more than one candidate resolves to the same story (edge case), use the most recently modified file and flag the others as duplicates to the caller.

Step 7: Parse Checkpoint

1. Act: Parse the five sections in order: Story, Branch, Tasks Done, Key Decisions, Remaining Todos. Ignore a trailing "Template Notes" (or any 6th) section if one is present — it is authoring guidance, not checkpoint state.
2. Check: Are all five sections present and structurally well-formed?
3. Skip: If yes, proceed to Step 8 with full confidence.
4. Act: If any section is missing, empty, or malformed — report exactly what WAS parsed, list what's missing or ambiguous, and ask the caller to confirm before proceeding. Never fill gaps by guessing.

Step 8: Report Parsed State

1. Act: Present the parsed state to the caller (Output Format) — story, branch, tasks done, tasks pending, key decisions.
2. Verify: All five sections from Step 7 (or their explicit gap markers) appear in the presented state — the caller can identify the next pending task without re-deriving Tasks Done.

Output Format

Write mode:

CHECKPOINT WRITTEN:
├── Story:    [#ID: Title]
├── Branch:   [feature/#ID-description]
├── Path:     [.pair/working/checkpoints/<story-id>.md | write-free — not persisted]
├── Mode:     [Created | Updated in place]
└── Tasks:    [N done / M total]

Followed by the full rendered checkpoint text (per the template) — always returned, regardless of $persist.

Resume mode:

CHECKPOINT RESUMED:
├── Story:      [#ID: Title]
├── Branch:     [feature/#ID-description]
├── Tasks done: [list T-N]
├── Tasks left: [list T-N]
├── Decisions:  [N recorded | None yet]
└── Confidence: [High — all sections parsed cleanly | Needs confirmation — see gaps below]

Example: Write Mode Mid-Story

Input — invoked with $mode: write from a session with story #313 loaded, 2 of 5 tasks done, no explicit $story:

$mode: write

(Story detected from branch chore/#313-t6-completion-criteria; state gathered from Step 3: Task Breakdown checklist shows T1-T2 checked, an ADL was recorded for the shared-reference naming convention.)

Output — .pair/working/checkpoints/313.md is created with:

# Checkpoint: #313 — Skill corpus effectiveness pass

**Last updated:** 2026-07-16 14:30
**Written by:** implement session

## 1. Story
**ID:** #313
**Title:** Skill corpus effectiveness: trigger descriptions, shared references, ...
**Goal:** Optimize the 35-skill corpus for LLM-executor effectiveness.
**Source:** github.com/foomakers/pair/issues/313

## 2. Branch
**Branch:** `chore/#313-t6-completion-criteria`
**Commit strategy:** commit-per-task
**Commits so far:** 2 — "sharpen soft Verify beats across assess-* skills"

## 3. Tasks Done
- [x] T1 — Router/catalog + correctness fixes — PR #325
- [x] T2 — Authoring effectiveness standard — PR #319

## 4. Key Decisions
- Shared-reference naming: `skill-conventions/<topic>.md` — see 2026-07-13-shared-reference-naming.md

## 5. Remaining Todos
- [ ] T6 — Completion criteria + steering pass
**Next immediate action:** finish the negation sweep on record-decision, then run the T2 checklist on 4-5 sampled skills.

Followed by the CHECKPOINT WRITTEN summary block (Output Format above).

Composition Interface

When composed by a future closing phase of /implement:

• Input: /implement would invoke /checkpoint with $mode=write between tasks (or on developer request) to persist progress, $mode=resume at Phase 0 when re-invoked on a story that may have been interrupted, and remove the checkpoint on story completion (cleanup).
• Output: Write mode's returned text becomes the session's record of state. Resume mode's parsed state lets /implement skip re-analysis and jump straight to the first pending task.

When composed by /publish-pr (the companion capability from the same epic split, wired in #256):

• Input: /publish-pr invokes /checkpoint with $mode=resume to read the story's handoff (branch, tasks done, decisions) before composing the PR — it is a read-only consumer. It never calls $mode=write; when no checkpoint exists, /publish-pr gathers state directly from the branch + story instead.
• Output: Resume mode's parsed state feeds the PR body. /publish-pr owns none of the checkpoint's lifecycle (no write, no cleanup).

When invoked independently:

• Full interactive flow. The developer or a subagent passes $mode and, if needed, $story explicitly.

Edge Cases

• No story context detectable: HALT and ask the caller to pass $story explicitly (Step 1).
• Corrupted or incomplete checkpoint on resume: report exactly what was parsed, list what's missing/ambiguous, and ask for confirmation before the caller proceeds (Step 7).
• Multiple checkpoints found for one story: use the most recently modified file; flag the duplicates to the caller for cleanup (Steps 4 and 6).

Graceful Degradation

See graceful degradation (guideline/template missing → use a minimal structure directly; PM tool not accessible during state reconstruction, Step 3 → ask the developer to confirm tasks done/pending directly) for the standard scenarios. Additional cases:

• If the checkpoint template is not found, use the minimal five-section structure directly: Story, Branch, Tasks Done, Key Decisions, Remaining Todos.
• If .pair/working/ does not exist yet, create it (and checkpoints/ under it) on first write.
• If git is not available or the branch cannot be determined, mark the Branch section [unknown — needs confirmation] rather than guessing.

Notes

• This skill writes at most one file — .pair/working/checkpoints/<story-id>.md — and only in write mode with $persist=true (default).
• Idempotent — see idempotency convention. This skill's check: write mode updates the same file in place (never duplicates); resume mode is read-only and safe to repeat.
• .pair/working/ holds operational, per-project runtime state — never touched by install or update (D14). It is not part of the distributed KB defaults.
• Checkpoints complement, not replace, git/PM-tool state. Even when state is reliably reconstructible from git and the PM tool (as /implement does today), a checkpoint adds an explicit, fast-to-read summary — most valuable across context resets and subagent handoffs, where reconstruction from scratch is expensive or impossible.
• The write-free ($persist=false) option serves composers that own their own persistence (e.g., embedding the handoff directly into a PR body) rather than writing a separate file.