[v0.4.0] Task Resumption Research Spike

[dean0x]

Task Resumption Research Spike

Version: v0.4.0
Priority: High
Complexity: Research (unknown feasibility)
Timeline: 1-2 weeks
Dependencies: None
Blocks: Task Resumption full implementation

Problem Statement

Before implementing Task Resumption, we need to validate whether Claude API/CLI supports session continuation. The implementation approach depends on this research:

• Full resumption: Restore complete conversation context (requires API support)
• Context injection: Inject summary of previous attempt into new session (fallback)
• Not feasible: Document limitations and defer feature

Research Questions

Q1: Claude API Session Continuation

• Does Claude API support resuming from a conversation ID?
• Can conversation history be serialized and restored?
• Is there a --continue or --session-id flag in Claude CLI?

Q2: Conversation State

• What format is conversation history stored in?
• Can tool call history be replayed?
• What are token limits for restored conversations?

Q3: Context Injection Fallback

• If full resumption isn't possible, can we inject context?
• What format works best? ("Previous attempt failed at step X...")
• How much context can we inject before hitting limits?

Q4: Checkpoint Overhead

• How long does checkpoint serialization take?
• What's the storage size per checkpoint?
• Can checkpoints be created asynchronously?

Deliverables

1. Technical Feasibility Document

Location: docs/research/TASK-RESUMPTION-FEASIBILITY.md

Contents:

• Claude API capabilities discovered
• Session continuation support (yes/no/partial)
• Context injection viability
• Recommended implementation approach
• Risk assessment

2. Proof of Concept Code

• PoC A: Full session continuation (if supported)
• PoC B: Context injection fallback
• Test coverage for basic scenarios

3. Go/No-Go Decision

Outcome	Action
Full resumption viable	Proceed with full implementation (3-4 weeks)
Context injection viable	Proceed with fallback implementation (1 week)
Neither viable	Document limitations, defer to v0.5.0+

Acceptance Criteria

• Feasibility document produced with clear findings
• At least one PoC demonstrates working approach
• Go/no-go decision documented with justification
• Implementation estimate provided for chosen approach
• Risks and mitigations documented

Research Scope

In Scope

• Claude API/CLI documentation review
• Prototype code for session continuation
• Prototype code for context injection
• Performance benchmarking of checkpoints

Out of Scope

• Full implementation (blocked by this spike)
• Production-ready code
• Database schema implementation
• MCP tool implementation

Success Criteria

• Clear recommendation on implementation approach
• Working prototype demonstrating feasibility
• Documented decision with technical justification

---

Created via /specify command

[dean0x]

Research spike complete. Decision: context injection fallback approach chosen.

Findings:
The spike validated feasibility of task resumption through context injection. Rather than producing a formal feasibility document, the research was conducted inline and the decision is documented in ROADMAP.md.

Approach:

• Claude API does not natively support session restoration from conversation ID
• Context injection fallback validated as viable: inject summary of previous attempt and relevant context into new session
• Implementation details documented in task resumption spike findings

The spike successfully served its purpose of validating the implementation approach for task resumption. Full implementation completed in PR #48 (v0.4.0).