Add context-hygiene at phase boundaries: compressed plan restatement and abort budget #19

Problem

Empirical evidence: autonomous agent success decays exponentially with task length, and the dominant practical failure in long runs is now context rot — quality degradation as input grows, even when the context window isn't full. The current template has 10 phases with no compression at boundaries; accumulated tool output and prior reasoning travel intact through the whole loop.

Motivation

Per RESEARCH.md §4 — Long-horizon agent runs decay exponentially; context rot is the practical limit:

• Ord 2025 (arXiv:2505.05115): constant per-minute failure hazard → exponential decay with task length.
• Chroma "Context Rot" 2025 (research.trychroma.com/context-rot): 18 frontier models tested. 20–50% accuracy drop from 10k → 100k+ tokens, even when window isn't full. Models perform better on shuffled haystacks than coherent ones.
• Anthropic, Effective context engineering for AI agents (first-party).

Confidence: high.

Proposal

Two changes:

1. Compressed plan restatement at phase boundaries

At each major boundary (triage → implementation → PR → merge → self-audit), the agent restates in a tight, structured form:

• Which issues are in scope
• Which files will be / were touched
• What invariants must hold

Subsequent phases work from this summary, not the full accumulated transcript. This is consistent with Anthropic's effective-context-engineering guidance.

2. Abort-and-file-gap-issue budget

If the loop hasn't converged after N tool calls or M tokens (numbers to calibrate empirically — start with 500 tool calls / 200k tokens as a placeholder), the agent aborts the current PR, files a gap issue describing what was attempted and why it didn't converge, and exits. The half-life model predicts persistence past a budget is strictly worse than restart with fresh context.

Acceptance criteria

• Phase 2 ends with a "Plan summary" block to be re-read at the start of Phase 3
• Phase 3 ends with an "Implementation summary" block to be re-read at Phase 4
• Phase 8 ends with a "Cycle summary" block to be re-read at Phase 9
• An abort-budget edge case added to Edge cases section, with explicit thresholds
• Aborted cycles produce a gap issue with a specific template

Open question

Whether to use ScheduleWakeup with a fresh-context restart when an abort fires. The harness supports this; the dev-loop skill could be authored to handle it.

Retrofit

Per CLAUDE.md's promotion rule, retrofit PRs needed for all four child skills.

Related

• RESEARCH.md §4
• Top-5 ranking in RESEARCH.md identifies this as the Create a private GitHub repo for each new dev loop skill #4 highest-impact change
• Issue Scope create-dev-loop to also update existing dev loops #16 (update-mode) is adjacent: a working update mode would let us iterate on this without per-skill manual retrofits

drafted by Claude on behalf of Daniel Stephenson