fix: orchestrator prompt minimizes worker fan-out for single tasks#429
fix: orchestrator prompt minimizes worker fan-out for single tasks#429
Conversation
🔍 Squad Review — PR #429 Round 1Models: claude-opus-4.6 ×2, claude-sonnet-4.6, gemini-3-pro-preview, gpt-5.3-codex PR SummarySingle-commit, 6-line diff in Findings
Confirmed Clean
Suggested Fix for F1// In BuildDelegationNudgePrompt(), change:
sb.AppendLine("Produce @worker blocks for ALL workers now. Do NOT explain, do NOT summarize previous work, ONLY output @worker blocks.");
// To:
sb.AppendLine("Produce @worker blocks for the workers you need. Do NOT explain, do NOT summarize previous work, ONLY output @worker blocks.");Suggested Fix for F2// Restore unconditional emphasis:
sb.AppendLine("IMPORTANT: Produce all @worker blocks in THIS SINGLE RESPONSE — not one at a time.");Verdict:
|
The orchestrator prompt previously led with 'Assign ALL workers that have relevant work' which caused models to invent micro-tasks to fill all workers even for single-item requests (e.g., one PR review dispatched to 5 workers doing commit checks, builds, file greps separately). Flip the emphasis so the default is ONE worker and fan-out is the exception: - 'Assign the MINIMUM number of workers needed' - 'Do NOT split a single task into micro-tasks across workers' - Only fan out for genuinely independent tasks Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
F1: BuildDelegationNudgePrompt() still said 'for ALL workers' which contradicts the minimum-workers fix on the recovery path. Changed to 'for the workers you need'. F2: Restored unconditional emphasis on single-response requirement. Changed conditional 'If you assign multiple workers, produce all...' back to unconditional 'IMPORTANT: Produce all @worker blocks in THIS SINGLE RESPONSE'. Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
81ff785 to
e03206f
Compare
🔄 Squad Re-Review — PR #429 Round 2Models: claude-opus-4.6 ×2, claude-sonnet-4.6, gemini-3-pro-preview, gpt-5.3-codex Previous Findings Status
New Findings🟢 🟢 CI StatusNo CI checks visible via Test CoveragePrompt-text-only change — existing structural tests ( Verdict: ✅ Approve — Both 🟡 R1 findings cleanly resolved. No new moderate or critical issues. The prompt now reads as a coherent minimize→batch instruction set. Ship it. |
…es (#458) ## Summary Adds a new skill at `.claude/skills/release-notes/SKILL.md` that teaches agents how to generate polished, highlight-driven release notes instead of the default GitHub auto-generated PR lists. ## What the skill covers - **Gathering changes** between tags via `git log` and `gh` CLI - **Categorizing commits** into Highlights, Features, Fixes, Performance, Mobile, Multi-Agent, Infrastructure, Docs - **Writing curated highlights** with user-focused, active-voice language - **Publishing** via `gh release edit` to update existing GitHub Releases - **README update strategy** — when to add features, what sections to touch, what to avoid - **Quick reference commands** for common release tasks ## Before vs After **Before** (current v1.0.15 release notes): > ## What's Changed > * fix: cleared input fields re-fill with stale draft text on re-render by @PureWeen in #435 > * fix: orchestrator prompt minimizes worker fan-out for single tasks by @PureWeen in #429 > *(17 more undifferentiated bullet points...)* **After** (skill-guided format): > ## 🚀 Highlights > - **Fiesta Mode** — Discover and orchestrate across multiple PolyPilot instances on your LAN with push-to-pair (#322) > - **Mixed-Model PR Review Squad** — New preset combining Opus, Sonnet, and Codex for comprehensive code review (#451) > > ## ✨ New Features > - Surface CLI subagent and skill events directly in chat + `/agent` command (#445) > *(categorized, concise, scannable...)* ## Testing This is a documentation-only skill file — no code changes, no tests needed. --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…ureWeen#429) ## Problem The orchestrator dispatch prompt led with **"Assign ALL workers that have relevant work"** which caused models to invent micro-tasks to fill all 5 workers even for single-item requests. For example, a single PR re-review would get split into: - Worker-1: check commits - Worker-2: do 5-model dispatch - Worker-3: grep one file - Worker-4: build and test - Worker-5: grep another file This wastes resources and makes sessions appear stuck (all 5 workers busy for 10+ minutes on what should be a single-worker job). ## Fix Flip the emphasis so the **default is ONE worker** and fan-out is the exception: - `Assign the MINIMUM number of workers needed` - `Do NOT split a single task into micro-tasks across workers` - Only fan out for genuinely independent tasks (e.g., "review PR PureWeen#100 and PR PureWeen#200") ## History - **Mar 10** (`a79f9f40`): Introduced aggressive fan-out with retry loop - **Mar 12** (`9c58ebf0`): Removed retry loop, added "only relevant" caveat - **This PR**: Flips the default from "assign all" to "assign minimum" --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
…es (PureWeen#458) ## Summary Adds a new skill at `.claude/skills/release-notes/SKILL.md` that teaches agents how to generate polished, highlight-driven release notes instead of the default GitHub auto-generated PR lists. ## What the skill covers - **Gathering changes** between tags via `git log` and `gh` CLI - **Categorizing commits** into Highlights, Features, Fixes, Performance, Mobile, Multi-Agent, Infrastructure, Docs - **Writing curated highlights** with user-focused, active-voice language - **Publishing** via `gh release edit` to update existing GitHub Releases - **README update strategy** — when to add features, what sections to touch, what to avoid - **Quick reference commands** for common release tasks ## Before vs After **Before** (current v1.0.15 release notes): > ## What's Changed > * fix: cleared input fields re-fill with stale draft text on re-render by @PureWeen in PureWeen#435 > * fix: orchestrator prompt minimizes worker fan-out for single tasks by @PureWeen in PureWeen#429 > *(17 more undifferentiated bullet points...)* **After** (skill-guided format): > ## 🚀 Highlights > - **Fiesta Mode** — Discover and orchestrate across multiple PolyPilot instances on your LAN with push-to-pair (PureWeen#322) > - **Mixed-Model PR Review Squad** — New preset combining Opus, Sonnet, and Codex for comprehensive code review (PureWeen#451) > > ## ✨ New Features > - Surface CLI subagent and skill events directly in chat + `/agent` command (PureWeen#445) > *(categorized, concise, scannable...)* ## Testing This is a documentation-only skill file — no code changes, no tests needed. --------- Co-authored-by: Copilot <223556219+Copilot@users.noreply.github.com>
Problem
The orchestrator dispatch prompt led with "Assign ALL workers that have relevant work" which caused models to invent micro-tasks to fill all 5 workers even for single-item requests. For example, a single PR re-review would get split into:
This wastes resources and makes sessions appear stuck (all 5 workers busy for 10+ minutes on what should be a single-worker job).
Fix
Flip the emphasis so the default is ONE worker and fan-out is the exception:
Assign the MINIMUM number of workers neededDo NOT split a single task into micro-tasks across workersHistory
a79f9f40): Introduced aggressive fan-out with retry loop9c58ebf0): Removed retry loop, added "only relevant" caveat