Repository navigation
fix(prompt-engineer): don't ask for <thinking> tags on models with native extended thinking - #1023
Open
vidaunited wants to merge 1 commit into
Open
vidaunited wants to merge 1 commit into
vidaunited wants to merge 1 commit into
Conversation
…native extended thinking
TheRealVitja
pushed a commit
to TheRealVitja/agency-agents
that referenced
this pull request
Oct 6, 2026
…nts) Curated sync of the open upstream pull requests as of 2026-10-06: - Script and CI fixes: msitarzewski#1055, msitarzewski#1030, msitarzewski#865, msitarzewski#860, msitarzewski#967, msitarzewski#870, msitarzewski#869, msitarzewski#889, msitarzewski#1056, msitarzewski#755, msitarzewski#868, msitarzewski#771, msitarzewski#867; ported msitarzewski#523, msitarzewski#512 and the permissions part of msitarzewski#790. - Existing-agent fixes: msitarzewski#1033-msitarzewski#1052, msitarzewski#1053, msitarzewski#1054, msitarzewski#1023, msitarzewski#756-msitarzewski#759, msitarzewski#799, msitarzewski#805, msitarzewski#715, msitarzewski#752, msitarzewski#784, msitarzewski#793, msitarzewski#858, msitarzewski#812, msitarzewski#789, msitarzewski#1007. - New agents: msitarzewski#702, msitarzewski#707, msitarzewski#731, msitarzewski#732, msitarzewski#764, msitarzewski#848, msitarzewski#859, msitarzewski#862, msitarzewski#863, msitarzewski#886, msitarzewski#908, msitarzewski#982-msitarzewski#985, msitarzewski#1031, msitarzewski#1032. - Docs: msitarzewski#577, msitarzewski#743, msitarzewski#762, msitarzewski#785, msitarzewski#786, msitarzewski#815, msitarzewski#816. Fixes found while integrating: Bash 3.2 guard for the msitarzewski#755 worker argv, Windsurf re-conversion over a stale .windsurfrules, locale-independent check-divisions.sh, agency- prefix handling in the outputs eval, and the India Business Navigator's YAML and headings. Resolves upstream issues msitarzewski#229, msitarzewski#763, msitarzewski#821 and msitarzewski#1027. Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01NwgfpJ9tGbUh5g84u5VSgv
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
engineering/engineering-prompt-engineer.mdtells the agent (line 45, in the example system prompt) to "think step-by-step inside<thinking>tags", and lists<thinking>→<answer>scaffolds as its reasoning-chain technique (line 163).On models with native extended thinking (Claude 4 and later), that instruction is counter-productive: the model already reasons before answering, and a
<thinking>scratchpad asked for in the prompt duplicates that reasoning and lands in the visible output — longer responses, and the "thinking" leaks to end users unless every caller strips the tags.Change
Two lines, no structural change:
<thinking>tags; use<answer>only when a downstream parser needs the delimiter.<thinking>→<answer>scaffold, scoped to models that do not have native thinking.Why it matters for this repo
The agent is the one users reach for to write other agents' prompts, so the pattern propagates. Anthropic's extended-thinking docs recommend against prompting for a scratchpad when thinking is enabled for exactly this reason.