pair-process-implement composes checkpoint + publish-pr via subagent (resume from checkpoint)

[rucka]

Story Statement

As a developer
I want pair-process-implement slimmed: at the end of the task cycle it writes a checkpoint and invokes publish-pr in a subagent whose prompt is the handoff only, resuming from checkpoint when interrupted
So that the PR is always built on clean context (guaranteed reset inside a single execution) and interrupted work never repeats completed tasks (R4.1, R2.10, R4.2)

Where: implement session; PR produced by the subagent; checkpoint in .pair/working/checkpoints/

Epic Context

Parent Epic: Split implement: checkpoint + publish-pr #206
Status: Refined
Priority: P0 (Must-Have)

Status Workflow

• Refined: Story is detailed, estimated, and ready for development
• In Progress: Story is actively being developed
• Done: Story delivered and accepted

Acceptance Criteria

Functional Requirements

Given-When-Then Format:

1. Given all story tasks completed in an implement session
   When the task cycle ends
   Then implement writes the checkpoint and spawns publish-pr in a subagent passing ONLY the handoff document in the prompt — the subagent's context is the handoff, nothing else

2. Given an implement session interrupted mid-story
   When implement is relaunched on the same story
   Then it detects the checkpoint and resumes from the first pending task, without repeating completed tasks

3. Given a project/tool where subagent spawning is unavailable
   When the PR phase is reached
   Then implement degrades gracefully: checkpoint written, publish-pr invoked inline (current one-shot flow), degradation noted in output

4. Given the existing one-shot flow (no interruption, no subagent constraint)
   When a story is implemented end-to-end
   Then behavior is equivalent to today plus checkpoint/PR split — no regression in the 5-step task cycle

Business Rules

• Two AI macro-phases (R4.2): refinement → implementation; the boundary artifact is a checkpoint
• The subagent boundary is mechanical isolation (D23): anonymous subagent, no named role
• implement never re-does gate/PR logic — it composes publish-pr only

Edge Cases and Error Handling

• Checkpoint exists but branch missing: HALT, report divergence (checkpoint says branch X, repo has none)
• Subagent fails mid-PR: checkpoint still valid; rerun invokes publish-pr again (idempotent: updates existing PR)
• Stale checkpoint (story already Done): warn and require explicit confirmation before reuse

Definition of Done Checklist

Development Completion

• All acceptance criteria implemented and verified
• Skill follows authoring conventions (composition, HALT rules)
• Code review completed and approved
• Documentation updated — docs site (apps/website) implement page updated (new flow)
• CLI (apps/pair-cli) unaffected verified

Quality Assurance

• Interruption/resume test (kill mid-story, relaunch, no task repeated)
• Subagent context audit: prompt contains handoff only
• Regression run of current one-shot flow (AC4)

Story Sizing and Sprint Readiness

Refined Story Points

Final Story Points: 5 (L)
Confidence Level: Medium
Sizing Justification: rework of the largest process skill + subagent orchestration + resume logic; confirmed L(5) — composition-heavy but both composed capabilities are siblings in this epic

Sprint Capacity Validation

Sprint Fit Assessment: yes, given siblings land first
Total Effort Assessment: fits — Yes

Dependencies and Coordination

Story Dependencies

Prerequisite Stories: checkpoint story, publish-pr story (both in this epic)
Dependent Stories: supervised automation (#212) reuses the same primitives
Shared Components: checkpoint format (contract), handoff-as-prompt pattern

Validation and Testing Strategy

Acceptance Testing Approach

Testing Methods: dogfood story on pair repo: full run, interrupted run, degraded (inline) run
Test Data Requirements: story with ≥3 tasks for meaningful interruption points

Notes and Additional Context

Refinement Session Insights: reset-by-subagent chosen because a skill cannot /clear its own context (D7); manual resume post-/clear covered by the same checkpoint
Documentation Links: R4.1, R4.2, R2.10 · Spec G5 · D7, D8, D23

Technical Analysis

Implementation Approach

Technical Strategy: modify .skills/process/implement/SKILL.md: task cycle unchanged; new closing phase = checkpoint(write) → spawn anonymous subagent with handoff-only prompt → subagent runs publish-pr; opening phase = checkpoint(resume) probe; degradation path inline
Key Components: implement SKILL.md phases, subagent invocation instructions, resume gate
Integration Points: checkpoint (both modes), publish-pr, PM tool state via mapping

Technical Risks and Mitigation

Risk	Impact	Probability	Mitigation Strategy
Handoff misses info the subagent needs	High	Medium	publish-pr contract lists required fields; checkpoint template enforces them; E2E dogfood
Tool-specific subagent APIs (agnosticism R2.13)	Medium	Medium	instructions phrased tool-neutrally ("fresh context execution"), degraded inline path always available

Task Breakdown

Checklist

• T1 — Closing phase: checkpoint write + subagent spawn with handoff-only prompt
• T2 — Opening phase: checkpoint probe + resume from first pending task
• T3 — Degraded inline path + divergence/stale edge cases
• T4 — Regression + interruption dogfood + docs site update

Dependency Graph

T1 → T2 → T3 → T4

AC Coverage

Task	AC1	AC2	AC3	AC4
T1	✓
T2		✓
T3			✓
T4	✓	✓	✓	✓