Address lightweight follow-up issues #266, #269, #271, #274, #275

[wwind123]

Summary

This PR addresses five lightweight follow-up issues from earlier planning and PR reviews.

#266 — stderr routing uniform across all backends

Verified that run_with_log already uses stderr=subprocess.STDOUT for all three agent backends (Claude, Codex, Gemini). Added a comment in runner.py documenting this intentional, uniform behavior.

#269 — Non-compact plan revision prompt missing workdir guidance

Added _coder_workdir_guidance(config, implementation=False) to the non-compact build_plan_revision_prompt return value (line 1187 in prompts.py), matching the compact path that already included it via _compact_plan_stable_prefix.

#271 — coder_followup path through attempt_envelope_normalization untested

Added three unit tests covering the coder_followup path through attempt_envelope_normalization: duplicate AGENT_STATE footer (fixed), trailing prose after signature (fixed), and prose before footer (returns None).

#274 — Combined envelope+disposition strip

When attempt_envelope_normalization fixes a structural defect but validate(normalized) raises UnknownPriorItemDispositionError, the orchestrator now attempts a combined deterministic strip from the normalized candidate. This resolves both defects in one pass without falling through to the full repair pass. Added an integration test that exercises this combined path.

#275 — json.dumps tight-packing fix

strip_unknown_prior_item_dispositions now inserts a \n separator between the re-encoded JSON and the trailing footer/signature when the original had no whitespace between them. Added tests verifying tight-packed inputs produce valid output and survive parse_structured_pr_review validation.

Test plan

• All 707 tests pass: python3 -m pytest tests/test_agent_loop.py -q
• 8 new tests added covering issues Follow up future review note: The non-compact build_plan_revision_prompt path (prompts.py:1187) does not include workdir #269, Follow up future review note: No dedicated unit tests cover the coder_followup path through `attempt_envelope_normaliz #271, Follow up future plan-review note: Consider a later cleanup to let deterministic disposition stripping operate on an env #274, Follow up future plan-review note: Verify that re-encoding the JSON with json.dumps in strip_unknown_prior_item_disposit #275

Closes #266, #269, #271, #274, #275

🤖 Generated with Claude Code

[wwind123]

Opened PR #277 addressing all five issues.

#266 — stderr routing verified uniform: All three agent backends (Claude, Codex, Gemini) call run_with_log, which already uses stderr=subprocess.STDOUT at the Popen layer. Added a comment in runner.py documenting this intentional, uniform behavior.

#269 — Non-compact plan revision prompt gains workdir guidance: Added _coder_workdir_guidance(config, implementation=False) to the non-compact build_plan_revision_prompt return value (prompts.py), matching the compact path that already included it via _compact_plan_stable_prefix.

#271 — coder_followup path through attempt_envelope_normalization now tested: Added three unit tests: duplicate AGENT_STATE footer (normalized), trailing prose after signature (stripped), prose before footer (returns None).

#274 — Combined envelope+disposition strip: When attempt_envelope_normalization fixes a structural defect but validate(normalized) raises UnknownPriorItemDispositionError, the orchestrator now tries strip_unknown_prior_item_dispositions on the normalized candidate in the same step, resolving both defects without falling through to the full repair pass. Only activates when the original error was structural (not itself an UnknownPriorItemDispositionError), preserving the existing block-2 flow for the simpler case.

#275 — Tight-packing newline fix: strip_unknown_prior_item_dispositions now inserts a \n separator between the re-encoded JSON and the trailing footer when absent. Added two tests: one verifying the separator is present, one confirming the result passes parse_structured_pr_review validation.

Tests: python3 -m pytest tests/test_agent_loop.py -q — 707 passed (8 new tests added).

-- Anthropic Claude

[wwind123]

Review verdict: Approved

No blocking issues found in the focused fixes for prompt workdir guidance, deterministic response repair, stderr capture documentation, or the added regression tests.

-- OpenAI Codex

[wwind123]

Review verdict: Approved

Review complete. The PR successfully addresses the follow-up issues, including adding necessary workdir guidance to the non-compact plan revision prompt, correctly handling combined repair paths, and accurately applying tighter spacing in stripped payloads.

-- Google Gemini