Repository navigation
Conversation
Match the versioned design-doc convention in docs/designs/ (PLAN_TUNING_V0.md, PLAN_TUNING_V1.md, SELF_LEARNING_V0.md, PACING_UPDATES_V0.md). The original PLAN_ROLLOUT_DOGFOOD.md filename introduced a new "_DOGFOOD" suffix that didn't match any existing pattern and read like an evidence appendix rather than a design doc. Restructure: - New "Design" section at the top describing what /plan-rollout is, what v0 ships, and what's deferred to v1.1+ - "Dogfood: PR garrytan#1241" section retains the worked example (file breakdown, reader-time estimate, verdict, findings) - New "v1.1 roadmap" section consolidates the four follow-up todos All original dogfood content preserved verbatim under its new section heading. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Long decision-brief bodies packed into AskUserQuestion's `question` field and multi-tab batches make the question card unreadable in panel hosts (VSCode). The brief belongs in markdown before the call; the tool payload should stay compact. generate-ask-user-format.ts now documents, in the Format section: - the two-part contract (markdown brief, then a compact tool_use payload); - "Tool payload rules": `question` is one sentence / no newlines / <=80 chars, brief content stays in markdown, batch at most two related questions/tabs (sequence the rest), and no duplicated trade-off text between `question` and `options[].description`; - four matching self-check items. This grafts only the compact-payload guidance onto the current resolver — it preserves the Conductor / runtime-failure-fallback content that landed on main since this PR was first opened. Regenerated all host SKILL.md variants (--host all) and refreshed the claude/codex/factory ship golden fixtures so the generated docs carry the new rules. Tests: new resolver assertions (compact-question, tab-batching, no-duplicate-tradeoff, self-check items) and a gen-skill-docs assertion that the rules reach generated tier 2+ SKILL.md. Affected suites (resolver-ask-user-format, gen-skill-docs, host-config) green. Fixes garrytan#1208 Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
360f345 to
8d79caf
Compare
|
Rebased onto current Reconciliation note: I grafted only the compact-payload guidance (two-part brief + tool payload, Regenerated all host SKILL.md variants ( |
|
@16francej Merge this version because it passed the Mac checks and is the chosen implementation for compact question payloads. |
garrytan#452 Three reviewed bug-fix overlays (MCP-packaging review, office-hours inline approval doc, host-neutral ## Review calibration), pinned set 22 -> 25 with executable regression fixtures. garrytan#1241 (compact AUQ preamble) held back: it drifts 43 baseline-render oracle hashes and needs a separate provenance call. Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
…k outputs Question-format class (garrytan#1208/garrytan#1241/garrytan#1066/garrytan#2373/garrytan#2035): - generate-skill-tree emits ONE references/QUESTION-FORMAT.md per dispatcher tree: decision brief as direct assistant text, short question field (~120 chars), 12-char tab labels, max 4 options with split protocol, and the unavailable-tool prose fallback that NEVER silently defaults. - Dangling module pointers to "the preamble's AskUserQuestion Format section" (excluded from canonical execution) repoint to the packaged contract: plan-ceo-review, office-hours, benchmark-models, sync-gbrain, plan-eng-review sections. - SHARED-JUDGMENT gains rule 14: unavailable question tool renders identical options as numbered prose and stops; never silently default. - question-preference-hook emits only documented PreToolUse outputs: the invalid permissionDecision 'defer' (garrytan#2035/garrytan#1924) is gone; no opinion is empty stdout, additionalContext rides without a permissionDecision. Tests that pinned 'defer' as correct now pin the documented shape. - Parity pins: +15 checks (3 per tree), 4402 -> 4417. - Tripwires: QUESTION-FORMAT exists + dispatcher-loaded in all trees, no preamble-format pointers survive, every hook permissionDecision literal is in the documented allow/deny/ask set. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
|
Thanks for this, @jbetala7. The fix is now on main: the AskUserQuestion format was reworked in the Opus 5.5 prompt cleanup (see v1.91.15.0). Closing as superseded; if you still see the problem on the latest gstack, please open an issue with ./setup --status output. |
Fixes #1208.
Summary
questionpayload and into the preceding markdownquestionto be one sentence, no newlines, <=80 charsTesting