Skip to content

code-walkthrough: stub generated files in prompts, recover from safeguard flags, bump to 1.4.1 - #65

Merged
crisperity merged 2 commits into
mainfrom
fix/walkthrough-safeguard-flag
Nov 9, 2026
Merged

crisperity merged 2 commits into
mainfrom
fix/walkthrough-safeguard-flag

Conversation

@crisperity

Copy link
Copy Markdown
Owner

Batch 3 of a private PR failed on a Sonnet 5.5 safeguards false positive (AUP). Likely trigger is a golden fixture with single lines up to 43 KB of escaped OOXML. Not confirmed by bisecting.

  • Prompt copy stubs binary, golden, snapshot, lock and huge-line/huge files, keeping hunk headers. Validation and merge still use the real diff.
  • Batches are also capped by characters (batch_max_chars, default 120000).
  • A flagged error is its own kind, skips escalate, and retries with the largest file stubbed, up to 5 rounds. Withheld files get a fixed note the validator accepts and are listed in meta and walkthrough_get.
  • Small route and routed retries get the same handling.
  • A batch that is still flagged with everything stubbed fails with the batch number and files named.

Not done: skipping a still-flagged batch and finishing the rest, because every later step needs every file covered by a fragment. Known gap: the model can still read a stubbed file through its read tools.

Tests: 190 passed across test_cw_run, test_cw_llm, test_fanout, test_validate_analysis, test_prompts; 140 across server, mcp, store, pipeline, e2e, sibling. To confirm: re-run the original PR.

…uard flags, bump to 1.4.1

A live-mode batch with a golden fixture (lines up to 43 KB of escaped XML) was rejected by Sonnet 5.5 safeguards. The failure was deterministic. The escalate retry re-sent the same prompt.
- Prompt copy stubs binary, golden, snapshot, lock and huge-line/huge files, keeping hunk headers. Validation and merge still use the real diff.
- Batches are also capped by characters (batch_max_chars, default 120000).
- A flagged error is its own kind, skips escalate, and retries with the largest file stubbed, up to 5 rounds. Withheld files get a fixed note the validator accepts and are listed in meta and walkthrough_get.
- Small route and routed retries get the same handling.
- A batch that is still flagged with everything stubbed fails with the batch number and files named.
…uard-flag

# Conflicts:
#	.claude-plugin/marketplace.json
@crisperity
crisperity merged commit 3f147d3 into main Nov 9, 2026
1 check passed
@crisperity
crisperity deleted the fix/walkthrough-safeguard-flag branch November 9, 2026 21:28
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant