checkpoint

Workflow Checkpoint Skill

Overview

Save and mutate persistent workflow state that survives context auto-summarization.

State files store: task details, worktree locations, PR URLs, and review status. This skill owns the write side of workflow state — initializing a workflow, updating fields, transitioning phases, and capturing a structured handoff. For the read/restore side (resuming after a break, reconciling against git, verifying a workflow exists), see @skills/rehydrate/SKILL.md.

Triggers

Activate this skill when:

• Starting a new workflow (ideate)
• Transitioning between workflow phases
• Saving progress for later continuation (checkpoint)

Phase Transitions

Valid transitions, guards, and prerequisites for all workflow types are documented in references/phase-transitions.md. CRITICAL: Phase mutation is a separate action from field mutation. When a transition has a guard, action: "update" the prerequisite fields first, then action: "transition" — guards read the most recent state, so updates land before guards evaluate. Attempting to mutate phase via action: "update" returns a RESERVED_FIELD error pointing at transition (see "Reserved fields" below).

Schema Discovery

Use exarchos_workflow({ action: "describe", actions: ["update", "init", "get"] }) for parameter schemas and exarchos_workflow({ action: "describe", playbook: "feature" }) for phase transitions, guards, and playbook guidance. For the lightweight oneshot variant (with its implementing → synthesize|completed choice state driven by synthesisPolicy), call exarchos_workflow({ action: "describe", playbook: "oneshot" }) — oneshot is a first-class playbook alongside feature/debug/refactor. Use exarchos_event({ action: "describe", eventTypes: ["workflow.transition", "task.completed"] }) for event data schemas.

State Location

Workflow state lives in the MCP event store, not the filesystem. Use exarchos_workflow get to read state and exarchos_view pipeline to discover active workflows. The pipeline view is repo-scoped by default — only the caller's repo; when the response reports unscopedTotal greater than page.total, workflows in other repos are hidden — re-query with scope: "all" to reveal them. Do not scan ~/.claude/workflow-state/*.state.json — that path is legacy and may be stale or empty.

State Operations

For full MCP tool signatures, error handling, and anti-patterns, see references/mcp-tool-reference.md.

Initialize State

At the start of ideate, use exarchos:exarchos_workflow with action: "init" with:

• featureId: the workflow identifier (e.g., "user-authentication")
• workflowType: one of "feature", "debug", "refactor", "oneshot"
• synthesisPolicy (optional, oneshot only): one of "always", "never", "on-request" (default "on-request") — silently ignored for non-oneshot types

This creates a new workflow state entry. The initial phase depends on workflowType:

• feature → starts in plan
• debug → starts in triage
• refactor → starts in explore
• oneshot → starts in plan

Workflow Types at a Glance

• feature — full plan → plan-review → delegate → review → synthesize for real features with subagent dispatch and review
• debug — triage → investigate → (thorough | hotfix) for bug workflows with track selection
• refactor — explore → brief → (polish | overhaul) for code improvements, polish for small and overhaul for multi-task
• oneshot — plan → implementing → (completed | synthesize) for trivial changes; direct-commit by default with an opt-in PR path resolved via a choice-state guard driven by synthesisPolicy and the synthesize.requested event

See @skills/oneshot/SKILL.md for the lightweight variant's full prose, including the choice-state mechanics and finalize_oneshot trigger.

Update State (fields only)

Use exarchos:exarchos_workflow with action: "update" with featureId and updates. This action mutates non-phase fields only — phase, workflowType, featureId, createdAt, and version are reserved (see "Reserved fields" below).

• Set artifact path: updates: { "artifacts.spec": "docs/specs/2026-01-05-feature.md" }
• Mark task complete (by index): updates: { "tasks[0].status": "complete", "tasks[0].completedAt": "<timestamp>" }
• Add worktree: updates: { "worktrees.wt-001": { "branch": "feature/001-types", "taskId": "001", "status": "active" } }

Worktree status values: 'active' | 'merged' | 'removed'

Transition Phase

Use exarchos:exarchos_workflow with action: "transition" with featureId and target:

• Advance phase: target: "delegate"

Transitions are HSM-validated and emit a workflow.transition event. Guarded transitions read state after the most recent update, so any updates: {...} that the guard depends on must land first.

Editing the tasks array

The dot-path parser used by set updates recognizes only numeric array brackets (tasks[0], tasks[1], …). Keyed forms like tasks[id=T-001] are NOT supported and now throw an INVALID_INPUT error with a clear message — earlier versions silently wrote to a bogus top-level key, returning success: true while the actual task was untouched. Three patterns are supported:

1. Replace the whole array (use this when the plan is being revised wholesale):

   exarchos_workflow({
     action: "update",
     featureId: "<id>",
     updates: { tasks: [
       { id: "T-001", title: "...", status: "pending" },
       { id: "T-002", title: "...", status: "pending" },
     ]},
   })
   

2. Edit one task by its array index:

   exarchos_workflow({
     action: "update",
     featureId: "<id>",
     updates: { "tasks[0].status": "complete", "tasks[0].completedAt": "<ts>" },
   })
   

   First read tasks (action: "get", query: "tasks") to find the index of the task you want to edit, then set by that index.

3. Append a new task by writing to the next-free index. If the array currently has length N, write to tasks[N]:

   // Suppose tasks already contains T-001 and T-002 (length 2). To append:
   exarchos_workflow({
     action: "update",
     featureId: "<id>",
     updates: { "tasks[2]": { id: "T-003", title: "Follow-up", status: "pending" } },
   })
   

   The parser allows writing one slot past the current length (MAX_ARRAY_GAP = 1); writing further out (tasks[5] against a length-2 array) throws INVALID_INPUT. Read the current tasks length before appending.

Integration Points

When to Update State

Event	State Update
ideate starts	init the workflow (initial phase plan)
Design & Rationale authored	update: { "artifacts.spec": "<path>" } (no transition — plan is the initial phase)
Decomposition added	update: { "artifacts.plan": "<path>", "tasks": [...] }, then transition target: "plan-review"
Plan-review gaps found	update: { "planReview.gaps": [...] }, auto-loop to plan
Plan-review approved	update: { "planReview.approved": true }, then transition target: "delegate"
Task dispatched	Set task status = "in_progress", startedAt
Task complete	Set task status = "complete", completedAt
Worktree created	Add to worktrees object
Review complete	Update reviews object
PR created	Set artifacts.pr, synthesis.prUrl
PR feedback	Append to synthesis.prFeedback

Oneshot-specific state updates

Oneshot is a first-class workflow type with a compressed lifecycle and an opt-in PR path. The rows below mirror the feature-workflow table above.

Phase	State updates	Events emitted
plan (oneshot)	oneshot.planSummary, artifacts.plan, optional oneshot.synthesisPolicy	workflow.transition
implementing (oneshot)	tasks[].status, artifacts.tests	task.*, optional synthesize.requested (via request_synthesize)
synthesize (oneshot)	synthesis.prUrl, artifacts.pr	workflow.transition, stack.submitted
completed (oneshot)	—	workflow.transition (to completed)

The implementing → synthesize | completed fork is a choice state resolved by finalize_oneshot, which reads the synthesisOptedIn guard (synthesisPolicy + synthesize.requested events). See @skills/oneshot/SKILL.md for the full opt-in mechanics.

Automatic State Updates

Skills should update state at key moments:

ideate/SKILL.md:

After authoring the Design & Rationale section of the unified docs/specs/ artifact:
- `action: "update"` — `updates: { "artifacts.spec": "<path>" }`
  (no transition — `plan` is the initial phase; continue to decomposition in the same phase)

plan/SKILL.md:

After saving plan:
1. `action: "update"` — `updates: { "artifacts.plan": "<path>", "tasks": [...] }`
2. `action: "transition"` — `target: "plan-review"`

delegate/SKILL.md:

On task dispatch:
- Update task status to "in_progress"
- Add worktree to state if created

On task complete:
- Update task status to "complete"
- Check if all tasks done, suggest checkpoint

State Schema

See docs/schemas/workflow-state.schema.json for full schema.

Key sections:

• version: Schema version (currently "1.1")
• featureId: Unique workflow identifier
• workflowType: Required. One of "feature", "debug", "refactor", or "oneshot"
• phase: Current workflow phase
• artifacts: Paths to design, plan, PR
• tasks: Task list with status
• worktrees: Active git worktrees
• planReview: Plan-review delta analysis results (gaps, approved)
• reviews: Review results
• synthesis: Merge/PR state

Reserved fields

exarchos:exarchos_workflow with action: "update" rejects two classes of paths with RESERVED_FIELD:

1. Top-level immutable keys — phase, workflowType, featureId, createdAt, version. Set once at init; never mutated directly.
2. Underscore-prefixed paths — any dot-path whose top-level key, or any segment, begins with _ (e.g. _version, _checkpoint.summary, _eventHints). These are projection or event-store metadata.

Alternate write paths:

• phase → exarchos:exarchos_workflow with action: "transition" and target: "<phase>". Transitions are HSM-validated and emit transition events.
• Underscore-prefixed paths → emit a typed event via exarchos:exarchos_event with action: "append" (e.g. checkpoint, state.patched). The projection folds the event into the field on the next read.
• workflowType, featureId, createdAt, version → not migratable. If you need a different workflow type, init a new workflow.

A RESERVED_FIELD error envelope now carries a typed data block:

{
  "success": false,
  "error": {
    "code": "RESERVED_FIELD",
    "message": "Cannot update reserved field: phase",
    "data": {
      "rejectedPath": "phase",
      "rule": "`phase` is top-level immutable — set once at init, never directly mutated thereafter.",
      "alternateWritePath": "Use `exarchos:exarchos_workflow` with `action: \"transition\"` and `target: \"<phase>\"` — phase changes are HSM-validated and emit transition events."
    }
  }
}

Read the full descriptor — including the regex catch-all for underscore paths — via exarchos:exarchos_workflow with action: "describe" and actions: ["update"]. The returned reservedFields block is the single source of truth.

Structured Handoff Output

When capturing a handoff (the checkpoint command surface), emit this structured summary so the agent that resumes sees the same contract you operated under before context clears. The ### House Rules block mirrors the rehydrate output (see @skills/rehydrate/SKILL.md) for correctness-signal symmetry — an agent producing a checkpoint sees the same house rules an agent rehydrating would.

## Checkpoint Saved

**Feature:** <feature-id>
**Phase:** <current-phase>
### Progress
- Tasks: X/Y complete
- Current: <what's in progress>
- Next: <suggested next action>

### House Rules (apply every action this turn forward)
**Skill:** <phasePlaybook.skillRef or "(no playbook for this phase)">
**Tools:** <phasePlaybook.tools rendered as bullets>
**Required model-emitted events:** <phasePlaybook.events rendered as bullets — e.g. `task.progressed`, `phase.advanced`>
**Auto-emitted events (runtime fires these):** <phasePlaybook.autoEmittedEvents rendered as bullets>
**Transition:** <phasePlaybook.transitionCriteria> | Guard: <phasePlaybook.guardPrerequisites>
**Validation scripts:** <phasePlaybook.validationScripts joined>

### Event Emission Hints
<_eventHints.missing rendered as bullets, or "(none — phase machinery satisfied)">

### Resume Instructions

To continue this workflow in a new session, run `rehydrate` — or start the harness fresh, and the SessionStart hook will auto-discover active workflows.

> **Discipline reminder:** every task transition this turn forward MUST land on the workflow event stream via `exarchos_event.append` or `delegate` subagent emission. Direct `Edit` / `Bash` / `git` actions on task branches without corresponding events will desync the workflow tracker (see RCA `docs/rca/2026-05-08-rehydrate-behavioral-gap.md`).

Auto-Checkpoint Triggers

Suggest a checkpoint when:

1. After delegate completes — all tasks done, before review
2. After a PR is created — in synthesize, before the feedback loop
3. After 3+ feedback iterations — context-accumulation risk
4. When the user mentions context issues — proactive save

Best Practices

1. Update often - State should reflect reality at all times
2. Use MCP tools - Prefer checkpoint MCP tools over manual JSON editing
3. Reconcile on resume - Always verify state matches git state (see @skills/rehydrate/SKILL.md)
4. Checkpoint at boundaries - Save state before likely context exhaustion
5. Read state, don't remember - After summarization, read from state (see @skills/rehydrate/SKILL.md)

Troubleshooting

MCP Tool Call Failed

If an Exarchos MCP tool returns an error:

1. Check the error message — it usually contains specific guidance
2. Verify the workflow state exists: call exarchos:exarchos_workflow with action: "get" and the featureId
3. If "version mismatch": another process updated state — retry the operation
4. If state is corrupted: call exarchos:exarchos_workflow with action: "cancel" and dryRun: true

Checkpoint Missing

If checkpoint is invoked with no active workflow:

1. Discovery first: call exarchos:exarchos_workflow with action: "list" to enumerate active workflows; if the list is empty the checkpoint command's "no active workflow" report is correct — exit cleanly
2. If list returns a candidate, verify it: call exarchos:exarchos_workflow with action: "get" and that featureId
3. If a workflow exists but checkpoint fails: check disk space and permissions on the event store

Example Workflow

1. Start new workflow: Use exarchos:exarchos_workflow with action: "init" with featureId: "user-authentication", workflowType: "feature"

2. After authoring the Design & Rationale section (plan is already the initial phase — no transition):

   • action: "update", featureId: "user-authentication", updates: { "artifacts.spec": "docs/specs/2026-01-05-user-auth.md" }

3. Save progress before a break: Use checkpoint to capture a structured handoff; resume later with rehydrate (see @skills/rehydrate/SKILL.md).