Skip to content

feat(check): retire the thinking-disabled workaround — effort buys extended thinking#153

Open
RobbieMcKinstry wants to merge 1 commit into
fix-check-agent-timeoutsfrom
enable-check-agent-thinking
Open

feat(check): retire the thinking-disabled workaround — effort buys extended thinking#153
RobbieMcKinstry wants to merge 1 commit into
fix-check-agent-timeoutsfrom
enable-check-agent-thinking

Conversation

@RobbieMcKinstry

Copy link
Copy Markdown
Contributor

The cersei pin predated the provider's ability to round-trip Anthropic
thinking blocks: its SSE parser dropped signature_delta, so the thinking
block echoed back on the second turn carried an empty signature and the API
rejected it. Effort was therefore applied as a sampling temperature, with a
comment promising to revisit once fixed upstream.

wack/cersei (now the authoritative source; we are not tracking pacifio
upstream) carries the fixes: 94f18b2 accumulates the signature from
signature_delta and 5bd06db preserves redacted_thinking across turns.
Bump the pin (d485d50 -> 4dfac7f — which also brings the bounded,
gitignore-aware GlobTool that closes the unbounded-walk timeout vector)
and map effort onto its natural vehicle:

  • medium/high enable extended thinking (4096/8192-token budgets, cersei's
    own EffortLevel scale, safely under the 16k default max_tokens);
  • low — the default — keeps thinking off for speed and cost, steering with
    temperature exactly as before (0.0, +0.5 per retry attempt).

Exactly one control is applied per run: the API rejects a temperature when
thinking is enabled (thinking samples at 1.0, which also gives thinking
retries the trajectory diversity that attempt_temperature provides on
thinking-free runs).

Not yet verified against the live API (no credentials in the build
environment): a multi check --effort medium run on a real project should
confirm turn-2 thinking blocks are accepted before this merges.

Co-Authored-By: Claude Fable 5 noreply@anthropic.com
Claude-Session: https://claude.ai/code/session_01VxKv1hhPZ4GocfmwHUk1G8

Copy link
Copy Markdown
Contributor Author

…tended thinking

The cersei pin predated the provider's ability to round-trip Anthropic
thinking blocks: its SSE parser dropped `signature_delta`, so the thinking
block echoed back on the second turn carried an empty signature and the API
rejected it. Effort was therefore applied as a sampling temperature, with a
comment promising to revisit once fixed upstream.

wack/cersei (now the authoritative source; we are not tracking pacifio
upstream) carries the fixes: 94f18b2 accumulates the signature from
`signature_delta` and 5bd06db preserves `redacted_thinking` across turns.
Bump the pin (d485d50 -> 4dfac7f — which also brings the bounded,
gitignore-aware GlobTool that closes the unbounded-walk timeout vector)
and map effort onto its natural vehicle:

- medium/high enable extended thinking (4096/8192-token budgets, cersei's
  own EffortLevel scale, safely under the 16k default max_tokens);
- low — the default — keeps thinking off for speed and cost, steering with
  temperature exactly as before (0.0, +0.5 per retry attempt).

Exactly one control is applied per run: the API rejects a temperature when
thinking is enabled (thinking samples at 1.0, which also gives thinking
retries the trajectory diversity that attempt_temperature provides on
thinking-free runs).

Not yet verified against the live API (no credentials in the build
environment): a `multi check --effort medium` run on a real project should
confirm turn-2 thinking blocks are accepted before this merges.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01VxKv1hhPZ4GocfmwHUk1G8
@RobbieMcKinstry RobbieMcKinstry force-pushed the enable-check-agent-thinking branch from bd475d2 to 749b5d8 Compare July 2, 2026 04:53
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