fix(skill/review): enforce parallel agent dispatch for weaker models

[wenshao]

Summary

• Strengthens the /review skill's Step 4 instructions so that qwen3.6-plus (and other models that were previously serializing the review agents) now reliably launches all of them in a single assistant turn.
• Replaces the one-sentence dispatch instruction with a prominent callout containing a rationale, ✅ CORRECT / ❌ WRONG ASCII examples, an explicit self-check before ending the turn, and a pattern-break ("STOP") for the common failure mode.
• Surfaces the rule in the top-level "Critical rules" list at the top of the skill, so the model sees it before it reaches Step 4.

Why

On qwen3.6-plus, Step 4 of /review was launching the 5 review agents sequentially (one per assistant turn) instead of dispatching them all in a single turn. The original phrasing ("invoking all tools in a single response") was too abstract for the model to consistently follow. Sequential dispatch multiplies review latency ~5× for no benefit, since none of the agents depend on another's output.

The new phrasing uses the concrete-example + anti-pattern + self-check style, which weaker models respond to more reliably, without changing any actual review logic.

Scope

Prompt-only change. This PR edits a single Markdown file (packages/core/src/skills/bundled/review/SKILL.md) that is loaded as a runtime prompt for the /review skill. No TypeScript, no build changes, no tests to run — the change takes effect the next time the skill is loaded.

Test plan

• Manually run /review <pr-number> on qwen3.6-plus after this lands and confirm all 5 agents launch in the same assistant turn
• Sanity-check that models which already parallelized correctly (e.g. Claude, GPT-4 class) are unaffected

[github-actions]

📋 Review Summary

This PR strengthens the /review skill's Step 4 instructions to enforce parallel agent dispatch across all models, particularly weaker ones that previously serialized agent launches. The changes are well-targeted, using concrete examples and anti-patterns to guide model behavior. Overall assessment: solid documentation improvement with clear rationale and actionable guidance.

🔍 General Feedback

• Positive approach: The use of ASCII diagrams (✅ CORRECT / ❌ WRONG) is an excellent pattern for guiding model behavior — concrete examples are far more effective than abstract instructions.
• Good structure: Moving the parallel dispatch rule to the "Critical rules" section at the top ensures models encounter it before reaching Step 4, which is a smart intervention point.
• Clear rationale: The "Why it matters" callout explaining the 5× latency multiplication provides necessary context for models to understand the importance.
• Self-check mechanism: The "verify exactly 5 task tool calls" self-check is a practical guard against the sequential dispatch bug.
• Single-file change: Focused modification to only the skill documentation — no code changes, minimal risk.

🎯 Specific Feedback

🔵 Low

• Line 21 (new rule Where is the config saved? #2): The rule numbering shift (old rule 2 → rule 3) is correct, but consider whether the Step 4 parallel dispatch rule should also be referenced in Step 9's "Create Review API" rule for consistency. Currently Step 9 focuses on the API format but doesn't reinforce the parallel dispatch requirement.

• Line 141 (CRITICAL callout): The phrase "extremely rare" regarding runtime rejection of parallel tool calls may be unnecessary hedging. Consider simplifying to: "If the runtime genuinely does not support parallel tool calls, fall back to sequential dispatch..." — this removes the speculative frequency assessment.

• Line 167 (agent prompt guidance): The revised text mentions "under 200 words" but doesn't explain why this matters (to fit all tool calls in one response without hitting output-length limits). The original diff shows this was added, but consider whether weaker models would benefit from an explicit connection: "Keep each agent's prompt short (under 200 words) so all tool calls fit in one assistant turn without hitting output-length limits."

✅ Highlights

• Excellent use of visual patterns: The ✅/❌ ASCII examples with explicit "STOP" callouts directly address the common failure mode of sequential dispatch.
• Strong justification: The PR summary clearly explains why the original phrasing failed ("too abstract for weaker models") and how the new approach addresses it ("concrete-example + anti-pattern + self-check style").
• Comprehensive coverage: The callout addresses multiple failure modes (single-agent sequential, multi-batch sequential) and provides a clear self-check mechanism.
• Good test plan: The test plan covers all three review scenarios (local diff, PR number, cross-repo URL) and explicitly calls out verification of parallel dispatch behavior.