[dmccoystephenson]

Summary

Four additions to the generated dev-loop template, grounded in RESEARCH.md §4 (long-horizon decay + context rot):

1. Phase 2 → 3 "Plan summary" — compressed plan block (issues, branch, expected files, invariants) re-read at the start of Phase 3.
2. Phase 3 → 4 "Implementation summary" — compressed implementation block (files modified, commit summaries, validation result, open carryovers) re-read at the start of Phase 4.
3. Phase 8 → 9 "Cycle summary" — compressed cycle block (issues closed, PR merged, surprises, followups) re-read at the start of Phase 9 to seed the self-audit.
4. Edge case: abort budget — if the cycle exceeds ~500 tool calls or ~200k tokens without converging, abort: mark the PR Draft (or push WIP), file a gap issue with a specific template, exit. The half-life model predicts persistence past budget is worse than restart with fresh context.

Closes #19

Test plan

• Phase 2 ends with a "Plan summary (re-read at the start of Phase 3)" block (line 167 area)
• Phase 3 ends with an "Implementation summary (re-read at the start of Phase 4)" block
• Phase 8 ends with a "Cycle summary (re-read at the start of Phase 9)" block
• Edge cases section has a new "Cycle exceeds abort budget" entry with explicit thresholds (~500 calls, ~200k tokens) and a 3-step abort procedure
• The abort-procedure gap-issue template is concrete (title format + 4-bullet body shape)
• Placeholder count unchanged at 20 (no new {{placeholders}} introduced)
• Diff size 29 LOC across 2 files — under the 400/10 soft ceiling (PR Add hard PR scope gates: LOC ceiling, do-not-auto-merge paths #30)
• Localization verified: create-dev-loop.md exists; "Phase 2", "Phase 3 — Implementation", "Phase 8 — Merge", "Edge cases" headings all confirmed present before editing
• Per the rule shipped in PR Promote 'verify each Closes #N' to template Phase 4 #26, ran gh issue view 19 and confirmed title/body match
• Per the rule shipped in PR Track implementation outcomes back into RESEARCH.md #27, added an Implementations entry under RESEARCH.md §4
• Do-not-auto-merge check (PR Add hard PR scope gates: LOC ceiling, do-not-auto-merge paths #30): modified files are create-dev-loop.md and RESEARCH.md; no matches

Open question (#19 noted it)

Whether ScheduleWakeup should be used to fresh-context-restart on abort. Out of scope for this PR — the abort instruction tells the agent to exit, and the user (or harness) initiates the next cycle.

---

drafted by Claude on behalf of Daniel Stephenson

[dmccoystephenson]

Self-review rubric:

• Scope: PASS — 2 files (create-dev-loop.md, RESEARCH.md); 29 lines total, each tied to a #19 AC item or the CLAUDE.md PR-27 Implementations rule.
• Tests-new: PASS — boundary-summary instructions are self-exemplifying (re-reading them at the next phase boundary is the test). The abort-budget edge case is exercised by this very cycle staying under it (well under 500 tool calls / 200k tokens for this PR).
• Tests-fix: N/A — not a bug fix.
• Sibling structure: PASS — three boundary-summary blocks share an identical shape (**X summary (re-read at the start of Phase Y).** heading + bullet list of structured fields); edge-case entry matches the **Label:** prose + numbered steps pattern of the existing edge cases above it.
• Sibling renames: N/A — no identifiers renamed.
• Docs: PASS — Phase 7 sources (create-dev-loop.md, README.md) reviewed; README is high-level and doesn't enumerate inner-phase summary blocks. RESEARCH.md §4 gets an Implementations entry per the CLAUDE.md PR-27 rule.
• Issue resolution: PASS — all five AC items from #19 met: (1) Phase 2 plan summary; (2) Phase 3 implementation summary; (3) Phase 8 cycle summary; (4) abort-budget edge case with explicit thresholds; (5) gap-issue template is concrete (title format + 4-bullet body shape).
• Validation: PASS — placeholder count unchanged at 20; no new fenced blocks introduced (the new prose uses inline backticks only); diff under scope ceiling.
• {{EXTERNAL_SIGNAL_LABEL}}: N/A — no CI; manual validation against the diff is the anchor (which I just did).
• Placeholder-consistency: N/A — no new placeholders.
• Escaping: N/A — no new fenced code blocks added.

Adversarial probe — 'Aren't 500 tool calls / 200k tokens just made up numbers?' Yes — RESEARCH.md §4 cites the half-life model but doesn't give a calibrated threshold. The 500/200k figures match #19's own placeholder values from the issue body, which the author marked as 'numbers to calibrate empirically.' Documented limitation; current values are reasonable starting points that future cycles can refine based on observed abort triggers.