feat(brainstorm): divergent style - #290
Open
cukas wants to merge 4 commits into
Open
Conversation
Brainstorm was architecturally the most convergent multi-engine mode: every seat got an identical prompt, qualityScore rewarded file-grounded concreteness, and the synthesis prompt collapsed all drafts into one 'specific and actionable' answer. Over-specified inbound questions from orchestrators then anchored the whole panel on the stated framing. - style option on runBrainstorm: 'divergent' (new default) | 'grounded' (previous behavior); --style flag on the CLI command - divergent seats get distinct stances (anchor/contrarian/first-principles/ outsider/expansionist/wildcard) via the system prompt, shuffled per run so no engine is systematically stuck with a low-scoring stance (Glicko stays unbiased); the protocol draft prompt stays byte-identical so the draft-block output contract is undisturbed - seats are told to treat the question's framing as a hypothesis about the underlying problem, not a hard constraint - structuralScore drops the keyFiles reward in divergent style so reframing drafts are not buried by winner selection - divergent synthesis presents 2-3 distinct directions (at least one challenging the framing) and still closes with a single recommendation so downstream automation keeps one decidable answer - scout path (runScout) is untouched: no style passed, grounded scoring ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Co-Authored-By: agon (KERN) <noreply@kernlang.dev>
Review consensus (minimax) flagged that the divergent default shipped with only the structuralScore math asserted. Adds fake-adapter coverage: assignStances assigns all six distinct stances and cycles on larger panels; divergent seats carry distinct stances in the system prompt while the protocol draft prompt stays byte-identical and stance-free; grounded/absent style keeps the plain instruction; runBrainstorm routes the synthesis prompt by style (divergent keeps distinct directions plus one closing recommendation, grounded restores the convergent prompt). ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Co-Authored-By: agon (KERN) <noreply@kernlang.dev>
Root-caused the long-standing 'claude returns "I'll quickly verify..."
instead of the answer' flakiness. Claude exec dispatches route through
the stream-json companion (--max-turns 1): when the model tries a tool,
the only turn ends at the tool call and the pre-tool preamble comes back
as a successful answer. Two adjacent defects fixed alongside:
- companion stream-json silently DROPPED the caller's systemPrompt (only
ACP/JSONRPC forwarded it) — brainstorm seat instructions ('do NOT use
tools') and the new divergent stances never reached claude. New
CompanionConfig.systemPromptFlag appends it to serverCmd
(claude: --system-prompt).
- stripStreamJson concatenated assistant text blocks AND the final
result event — duplicating one-turn answers and keeping multi-turn
tool preambles. The result event is now authoritative; assistant
texts remain the fallback for streams the result never closed
(timeout kill, is_error/max-tokens truncation).
Fix mechanics:
- DispatchOptions.textOnly (opt-in, set by brainstorm seats + synthesis;
never by agon ask/room work which may legitimately want tools) appends
CompanionConfig.textOnlyArgs (claude: --tools '') so the model gets no
tool definitions and the single turn is always the complete answer.
- Backstop: a stream-json exec turn that still ends on tool_use returns
empty stdout (preamble preserved in stderr for diagnosability), so the
adapter's existing empty-output fall-through retries via the plain CLI
spawn (--max-turns 10), which can finish the tool loop; the fixed
stripStreamJson then extracts the final answer from its result event.
Nero adversarial pass shaped the design: textOnly is per-dispatch opt-in
(not blanket exec), the backstop keeps the preamble in stderr, and the
result-event preference falls back to assistant texts on truncation.
⚔️ Forged by [Agon](https://github.com/KERNlang/agon)
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: agon (KERN) <noreply@kernlang.dev>
Full-roster review (high risk: schema + shared dispatch contract) on f50eca6 found one verified blocker and three consensus gaps, all fixed: - BLOCKER (codex): CompanionConfigSchema omitted systemPromptFlag and textOnlyArgs, so Zod stripped both from registry-loaded claude.json and the new behavior was inactive at runtime. Schema extended; regression test asserts the fields survive validateEngineConfig, and a fresh-dist registry load was verified manually. - tool_use backstop now covers review mode too (codex/zai/kimi/minimax consensus): review turns inherently want tools, so via the --max-turns 1 companion they hit the preamble truncation on nearly every dispatch; empty stdout now falls through to the CLI review path (--max-turns 50). - textOnly extended to every dispatch whose system prompt already says 'do NOT use tools': tribunal debates + verdict, council advisors + chair, nero critic, campfire, synthesis-modus drafts/improve/judge, thinking critique/steps (codex finding on tribunal, swept repo-wide). - stripStreamJson: only a STRING result event is authoritative; non-string payloads keep the old stringify-into-parts behavior so an object result can never mask real assistant text (agy finding). - systemPromptFlag argv append gated to the stream-json protocol — ACP/JSONRPC forward the system prompt in-band and must never get it twice (minimax finding). ⚔️ Forged by [Agon](https://github.com/KERNlang/agon) Co-Authored-By: Claude Fable 5 <noreply@anthropic.com> Co-Authored-By: agon (KERN) <noreply@kernlang.dev>
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
2 participants
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.
No description provided.