Summary

When the deepen-plan command was folded into ce:plan, the ability to review sub-agent findings individually before they modified the plan was lost. Users working on complex multi-phase projects want to accept or reject each reviewer's suggestions rather than having them auto-integrated.

This adds an interactive deepening mode that activates when a user explicitly asks to deepen an existing plan (e.g., "deepen the plan", /ce:plan deepen). Auto-deepening during plan generation is unchanged.

Fixes #429

What changed

• Skill description and argument-hint updated so ce:plan auto-triggers for deepening requests
• Phase 0.1 — deepen intent now targets plan docs (not requirements docs), with light-touch discovery that confirms with the user when the match isn't obvious
• Phase 5.3 — two deepening modes: auto (default during generation, unchanged) and interactive (on-demand, per-agent accept/reject/discuss review)
• New Phase 5.3.6b — interactive finding review gate between agent dispatch and synthesis. Findings presented one agent at a time; only accepted findings are integrated
• Phase 5.3.7 — interactive mode only synthesizes accepted findings
• Pipeline mode — always auto, no behavioral change for automated workflows
• Signal word tightening — only "deepen"/"deepening" auto-enter the fast path. Generic words like "strengthen", "confidence", "gaps" require user confirmation when targeting the whole plan, and route to normal editing when they name a specific section. Prevents false-positive deepening triggers from requests like "strengthen the risk section"

Design decisions

• One skill, one entry point — behavior adapts to invocation intent rather than adding a separate /deepen-plan command back
• "Discuss" leads to one exchange then a binary accept/reject re-ask — no automatic fallback that modifies the plan without explicit consent
• When all findings are rejected, document-review is skipped (plan unchanged) and flow proceeds to post-generation options
• When findings are accepted, document-review runs as a quality gate on the modifications
• Three-tier signal word hierarchy: "deepen" (auto-trigger) > "strengthen"/"confidence"/"gaps" on whole plan (confirm first) > same words on a specific section (normal edit). This came from eval testing that showed generic words could false-positive into a full deepening pass when the user just wanted a targeted edit

Eval results

Ran 8 test cases across 2 iterations using skill-creator evals:

Test plan

• /ce:plan on a new feature — auto-deepening runs as before, no interactive prompts
• /ce:plan deepen with a plan in docs/plans/ — enters interactive mode
• "deepen the plan" without slash command — skill auto-triggers
• "strengthen the plan" — confirms deepening intent before entering interactive mode
• "strengthen the risk section" — routes to normal editing, not deepening
• Interactive mode: accept some findings, reject others — only accepted ones integrated
• Interactive mode: reject all findings — plan unchanged, skips to 5.4
• Pipeline/headless mode — always auto, no interaction

🤖 Generated with Claude Opus 4.6 (1M context, extended thinking) via Claude Code