Checkpoint

When to fire (EXP-L21: attention efficiency decays to ~11% late-session; a reset from a checkpoint restores the flat ~50-70k load-bearing line):

• a long task crosses a natural boundary (a commit lands, a phase completes)
• context is about to compact, or the session is degrading (anchor-drift, I0.13)
• the manifest's execution.checkpoints list says so

Checkpoint vs handoff: same task + same session continuation → checkpoint. Ending the session / transferring to a future session with decisions and policy → handoff.

2. Write _meta/checkpoints/{task}-{timestamp}.json per schemas/kernel.checkpoint.v1.schema.json:

   • task: goal + handoff_ref/contract_ref if they exist
   • steps_completed: each with EVIDENCE (commit sha, test output, command result), never intent. Receipts describe intent; files describe reality.
   • current_outputs: files produced so far
   • provenance.dirty_tree_sha256: when dirty, sha256 of git diff --binary HEAD
   • pending_steps: what remains, in order
   • resume: position (exact place in the work), entrypoint (skill to re-enter), next_operation (the LITERAL first action — "retarget tests/run-tests.sh lines ~1250 registration tests", not "continue the migration")
   • context (optional): required selectors + budget if the resume should be bounded

   Quote every sha/commit string (all-digit shas parse as integers and fail validation).

3. Validate — MANDATORY:

   "${CLAUDE_PLUGIN_ROOT:-.}/orchestration/manifest/kernel-manifest" validate _meta/checkpoints/{file}.json
   

4. Record:

   agentdb write-end '{"skill":"checkpoint","saved_to":"_meta/checkpoints/{file}.json","position":"..."}'
   

Resume path: /kernel:ingest discovers the newest manifest automatically (kernel-manifest latest — checkpoints and handoffs share discovery), or accepts an explicit path.