liza-mas/checkpoint-summary

Purpose

After agents complete a planning or writing phase (epic planning, story writing, spec generation), summarize their output so a human can efficiently review what was decided, what remains open, and where their attention is needed.

This skill answers: "What did the agents produce, what did they decide, and what do I need to weigh in on?"

The agents already did the work — planning, reviewing, approving. This skill reads their outputs and distills them into a checkpoint summary that respects the human's time.

Distinct from spec-review: spec-review audits spec quality. This skill summarizes what was already reviewed and approved, surfacing only what needs human judgment.

Trigger

Use this skill when:

• Epic planning completes and the human needs a checkpoint summary
• Story writing completes across multiple agents
• Any multi-agent phase produces artifacts the human hasn't read
• User asks "what did the agents produce?", "what needs my attention?", or "summarize the plans"
• Orchestrator requests a human checkpoint (Liza mode)

Inputs

The single entry point is .liza/state.yaml — the source of truth for all Liza state.

From state.yaml, the skill reads:

• goal.spec_ref: the upstream source document the agents worked from
• tasks[]: each task with its scope, status, output capabilities, approvals, and history
• Artifact refs (read all that exist, in priority order):
  • tasks[].plan_ref / tasks[].arch_ref — task-level planning and architecture artifacts
  • tasks[].output[].plan_ref / tasks[].output[].arch_ref — output-entry artifacts
  • tasks[].spec_ref — task-level spec (may differ from goal.spec_ref)
• tasks[].approvals[]: review verdicts with provider/diversity context (canonical); fall back to tasks[].approved_by if approvals[] is absent
• tasks[].history[]: full event timeline (claimed, checkpoint, submitted, approved, merged)
• sprint.status and sprint.checkpoint_trigger: why the checkpoint was triggered

No other discovery is needed. Read every artifact file referenced by the ref fields above. Read the upstream source (goal.spec_ref). Everything else is in the state file itself.

Protocol

Phase 1: Inventory

1. Read .liza/state.yaml to understand the full pipeline state: goal, tasks, agents, sprint status, and checkpoint trigger.

2. Read the upstream source (goal.spec_ref) to understand what the agents were working from — entities, decisions, constraints, interactions, scope boundaries.

3. Read all produced artifacts referenced by task-level refs (plan_ref, arch_ref) and output-entry refs (output[].plan_ref, output[].arch_ref). Skip entries where no ref field points to a file. For each artifact read:

   • What was produced (title, scope, capabilities/stories count)
   • What verdict the reviewers gave (tasks[].approvals[]; fall back to approved_by)
   • Key events from tasks[].history[] (rejections, re-reviews, anomalies)

Phase 2: Extract

From the artifacts and agent outputs, extract three categories:

Decisions Made

Choices the agents made that the upstream source left open. For each:

• What was decided: the specific choice, in one sentence
• Where: which artifact and section (e.g., EP-002 CAP-003, or story-007 AC-2)
• Confidence: if the agent flagged it (HIGH/MEDIUM/LOW), report that; if they didn't flag it at all, mark it as unflagged
• Reversible?: can this be changed later, or does downstream work lock it in?
• Departures from upstream: decisions that contradict or reinterpret the upstream text get flagged separately — even if defensible, the human should know

Open Points

Items that remain unresolved after the agent work:

• Open questions: items agents explicitly flagged as needing human input (OQ-tagged)
• Gaps: things no artifact addresses that the upstream source expects
• Cross-artifact inconsistencies: places where sibling artifacts disagree
• Underspecified areas: concepts mentioned across multiple artifacts but never defined in any of them (e.g., "clusters and tags" referenced in filtering, listing, and CLI but never given a creation mechanism)

Risks

Implementation risks the artifacts create or carry forward:

• Reuse assumptions: references to existing components without compatibility assessment
• Atomicity/complexity: operations described as atomic that may be difficult to implement
• Handoff gaps: one artifact's output is another's input, but the interface isn't defined
• Irreversible operations: without undo or rollback being in scope

Phase 3: Prioritize

Not everything needs human attention. Classify each item:

| Priority | Meaning | Action needed | |----------|---------|---------------| | Decide | Human must make a choice before next phase starts | Present the decision with options | | Confirm | Agents made a reasonable choice — human should validate | Present the decision, default is accept | | Note | Worth knowing, no action needed | Include in summary, don't interrupt for it |

Prioritization heuristics:

• Unflagged decisions that depart from upstream text → Decide
• LOW-confidence assumptions → Decide
• MEDIUM-confidence assumptions on irreversible operations → Decide
• MEDIUM-confidence assumptions on reversible operations → Confirm
• HIGH-confidence assumptions → Note (unless they depart from upstream)
• Gaps that block the next phase → Decide
• Gaps that affect only implementation details → Note
• Risks with no mitigation path → Confirm
• Risks with natural mitigation during implementation → Note

Phase 4: Report

Present in this format. Decide items first, then Confirm, then Notes.

# Checkpoint Summary: [Phase Name]

## Status

| Artifact | Scope | Verdict |
|----------|-------|---------|
| [name] | [one-line scope] | [Approved / Rejected / Conditional] |

> N decisions needing input · M items to confirm · K notes

## Decisions Needing Human Input

Items where the human must choose before the next phase starts.

### [Decision Title]

- **Context:** [What the upstream says or doesn't say]
- **Agent decision:** [What the agents chose]
- **Why it needs you:** [What makes this non-obvious — departure from upstream,
  irreversible, low confidence, or gap]
- **Options:** [If applicable — confirm agent choice, override with X, or defer]

## Decisions to Confirm

Agents made reasonable choices. Confirm or override.

### [Decision Title]

- **Agent decision:** [What was chosen]
- **Rationale:** [Why it's reasonable]
- **Override if:** [When the human might want something different]

## Open Points

Unresolved items carried forward.

### [Item Title]

- **What's open:** [Description]
- **Impact:** [What's affected if unresolved]
- **Where it surfaces:** [Which artifacts reference this]

## Risks

### [Risk Title]

- **What could go wrong:** [Description]
- **Which artifacts:** [Where this risk lives]
- **Mitigation:** [If any exists in the plans, or "none specified"]

## Notes

Brief items worth knowing but not requiring action.

- [Item]: [One-line description]

Constraints

• DO NOT modify source artifacts — only the generated checkpoint summary report may be written
• DO NOT re-review or second-guess the reviewer's verdict — summarize it
• DO NOT bury decisions in long prose — one item, one heading, one clear question
• DO surface unflagged decisions — the highest-value findings are choices agents made without marking them as assumptions
• DO state the specific question the human needs to answer, not just "this needs review"
• DO keep the summary scannable — a human should grasp the state in under 2 minutes
• DO report the total count of decisions/open points/risks in the Status section so the human can gauge the review effort before diving in

Anti-Patterns

• Re-reviewing approved work: The agents already reviewed. Don't re-evaluate the architecture, question design choices that were explicitly approved, or propose alternatives. Summarize what was decided.
• Exhaustive listing: Dumping every assumption from every artifact. The human doesn't need to see 40 HIGH-confidence assumptions that are obviously correct. Filter to what matters.
• Missing the silent decisions: Focusing on what agents flagged (assumptions, open questions) and ignoring what they didn't. The most important items are often decisions baked into the plan without being called out.
• Vague attention items: "The CLI design needs review" is not actionable. "The CLI plan doesn't expose project deletion, but the API supports it — should users be able to delete projects in v1?" is actionable.
• Severity inflation: Marking everything as "Decide". Most items are "Confirm" or "Note". Reserve "Decide" for items where the human genuinely has a choice to make that changes the downstream work.

Integration

| Skill | Relationship | |-------|-------------| | epic-writing | Upstream producer. Summarize epic plans at the planning checkpoint. | | user-story-writing | Upstream producer. Summarize stories at the story-writing checkpoint. | | spec-review | Complementary. spec-review finds spec defects; this summarizes agent decisions. Different purposes, can run on the same artifacts. |

Mode-Specific Behavior

Pairing mode: Present the summary interactively. Walk through "Decide" items one at a time, collecting the human's decision before moving to the next. For "Confirm" items, present as a batch — the human can scan and override selectively. End with a count of decisions made and items still open.

Liza mode: Checkpoint Summary operates autonomously within task scope. Write the report to the worktree (e.g. docs/checkpoint-summary.md) and submit for review. If any "Decide" items exist, mark BLOCKED with blocked_reason summarizing the decisions needed and blocked_questions listing each one — the human must resolve them before the next phase starts. If only "Confirm" and "Note" items exist, submit normally.

| Pairing Prompt | Liza Behavior | |----------------|---------------| | "N decisions need your input — walk through them?" | Mark BLOCKED; report in worktree | | "Agents made N decisions — all look reasonable. Confirm?" | Submit for review; report in worktree | | "No open points — ready for next phase" | Submit for review; report in worktree |