Checkpoint

library:checkpoint — Session State Capture

Capture everything the next session needs to resume without context loss.

When to Use

• End of a work session
• Before switching to a different task or project
• When the user says "checkpoint", "save state", "wrap up"
• Before context window is likely to expire

Process

Step 1: Gather State

Review the conversation and recent tool calls to collect:

1. Accomplished — concrete deliverables (files created, features implemented, bugs fixed)
2. Changes — run git status and git diff --stat for file-level changes
3. Next Actions — specific, actionable items (not "continue working on X")
4. Open Decisions — unresolved questions with options and impact
5. Key Context — non-obvious information (error messages, gotchas, constraints)
6. Memory Updates — what was saved to Claude memory this session

Step 2: Write Checkpoint

Call library_checkpoint_write with gathered data:

• topic: kebab-case summary (e.g., "auth-middleware-refactor")
• status: current state in one line
• next_session: what to do first next time
• accomplished: semicolon-separated list
• next_actions: semicolon-separated list
• key_context: semicolon-separated list

Step 3: Update Memory

Save a project-type memory pointing to the checkpoint:

name: <Topic> Session Checkpoint
description: <one-line summary>
type: project

Content: path to checkpoint file, resume point, key decisions.

Step 4: Update PM Tasks (if PM configured)

For each PM task touched this session:

• Call library_pm_update with a comment summarizing progress
• If task is complete, transition status (but only if verified — evidence before claims)

Step 5: Check for Orphaned Work

Call library_pm_sync for active projects. Compare accomplished work against open tasks. Flag any session work that doesn't match a PM task (orphaned work — should it be tracked?).

Step 6: Detect gaps + auto-file follow-up tickets

A session almost always discovers follow-ups that are NOT in PM yet — orphaned commits, side-fixes mid-merge, skipped tests with TODOs, drift between spec + code, manual workarounds, deferred validations. These die in operator memory if not filed.

Step 5's orphaned-work check compares accomplished vs open tickets. Step 6 goes further: it scans the session transcript + checkpoint for signals of untracked follow-ups that may not even be visible as "accomplished work" yet.

Procedure

1. Inventory candidate gaps by re-reading the session transcript + the checkpoint just written. Categories to look for:

   Category	Signal
   Test skipped with TODO	describe.skip(.*TODO|skip.*FIXME in committed diff
   Migration number collision	two files with same NNN_ prefix in src/db/migrations/
   Hot-fix mid-merge	commit titled fix(*): ... against own session's merges
   Spec not updated for shipped code	new migration / endpoint / schema change without matching B-spec edit
   Config corrected without audit	account_id / project_id / region value flipped without consumer audit
   Deferred validation	"smoke test pending after X applies" / "blocked on operator click"
   Pre-existing bug papered over	shipped fix with # pre-existing, not introduced here note
   Manual cleanup left for operator	<dir>.old-pre-cleanup / <file>.bak style preserved artifacts
   Secrets near-miss	gitleaks fired on session content (worktree .env, etc.)
   Dep bump that violates pin policy	edited HARD-pin requirement file

2. Cross-check each candidate against open PM:

   • Call library_pm_query with status=To Do + In Progress for active project key
   • For each candidate, grep summaries + descriptions for keyword overlap (table name, file path, ticket-ID reference)
   • If candidate maps to existing ticket → skip (tracked)
   • If candidate has no match → it's a gap

3. Surface the gap list to operator BEFORE filing. Format:

   ## Gap detection — N untracked items
   | # | Gap | Suggested ticket title | Labels |
   |---|---|---|---|
   | 1 | <symptom> | <imperative title> | follow-up,<area> |
   

   Ask: "File these N tickets in <PROJECT_KEY>? (yes / cherry-pick: 1,3,5 / no)"

4. On operator yes / cherry-pick: call library_pm_create_task for each accepted gap. Description must include:

   • Detected during: <this checkpoint slug> YYYY-MM-DD
   • Source signal: the literal grep / commit / file that triggered detection
   • Action: specific repair steps
   • Labels: always include follow-up + 1-2 area tags (e.g. db-migration, tf-audit, test-cleanup, security-audit)

5. Append filed ticket IDs to the checkpoint file under a new section ## Gap follow-ups filed so the next session sees them at the same level as accomplished/blockers.

Hard rules for gap detection

• Never silently skip surfacing a gap. If unsure, surface it + let operator decide.
• Never auto-file without operator approval (default no-on-empty).
• Cherry-pick mode allowed: operator types 1,3,5 to file a subset.
• One ticket per gap — do not stack multiple unrelated fixes in one ticket body.
• If the gap lacks a clear repair path, file it anyway with ## Open question: block — better tracked than forgotten.

Step 7: Confirm

Report to user:

"Checkpoint written to <path>. Memory updated. [N PM tasks updated.] The next session should start by reading that file. Anything to add?"

Wait for user confirmation.

Resume from Checkpoint

When invoked with "resume" intent or when starting a new session:

Step R1: List Checkpoints

Call library_checkpoint_list to show recent checkpoints with dates and status.

Step R2: Select Checkpoint

Present the list and let user pick one to load. If only one recent checkpoint exists, offer to load it directly.

Step R3: Load Checkpoint

Call library_checkpoint_read(path) to display checkpoint content.

Step R4: Resume

Resume from the checkpoint's "What's Next" section. Read referenced files and re-establish context before proceeding with the next actions.

Quality Gates

• No vague actions. "Continue working on X" is not valid. Be specific.
• No assumed context. Write as if the next reader has zero memory.
• Fresh git state. Run git status now, don't reuse earlier observations.
• Absolute paths. No ambiguous references.
• Decisions include rationale. Not just "chose X" but "chose X because Y."

Token Budget

Weight: Medium Estimated context cost: ~1500 tokens Subagent delegation: No

Storage Location (Hard Rule)

Checkpoints always live under the Reading Room. By default the file is written to <reading_room.path>/checkpoints/. An explicit checkpoints.path in library-config.yaml is allowed but must resolve under reading_room.path — anything else is rejected with an error. There is no silent fallback to the MCP server's CWD.

MCP Tools Used

• library_checkpoint_write — write structured checkpoint file
• library_checkpoint_list — list recent checkpoints (for resume flow)
• library_checkpoint_read — read a specific checkpoint (for resume flow)
• library_pm_update — comment on PM tasks (if PM configured)
• library_pm_sync — check for orphaned work (if PM configured)
• library_pm_query — find existing tickets to cross-check candidate gaps (Step 6)
• library_pm_create_task — file new tickets for accepted gaps (Step 6)