feat(feature-planner): major refactor — Task-tool spawning, progressive disclosure, bug fixes
SKILL.md shrunk from 1444 → 245 lines via progressive disclosure. Detailed
workflow content extracted into 8 reference files loaded only when needed.
Critical fixes
- Replace `claude -p` bash invocation with Task-tool spawn patterns (broken in
VS Code extension, no hook/billing/telemetry inheritance, fails on PowerShell)
- Fix malformed Phase Progress Log table (header had 6 cols, row had 5 —
every PLAN.md generated from the template was broken)
- Drop `/mnt/skills/public/` frontend-design path (Claude.ai sandbox only,
never existed in CLI / VS Code / JetBrains installs)
High-impact additions
- Resume / abort triggers added to skill description ("continue feature",
"resume planning", "pick up where we left off")
- Strategy behavior centralized in references/modes-and-strategies.md (was
duplicated in 5+ places — drift was inevitable)
- references/resumability-and-abort.md adds a clean abort flow
- references/agent-spawning.md covers Task-tool patterns, parallel phases
with git worktrees, MODEL_PROFILE cost overrides, local-only repo handling,
and optional code-review-graph MCP integration
- Phase implementer agents now MUST scan CONTEXT.md `## Failed Approaches`
before writing code (#1 source of duplicated mistakes across phases)
- Test writer / implementer division of labor made explicit (implementer
writes happy-path tests, test writer adds adversarial coverage)
- New `## New Tooling Introduced` section in CONTEXT.md prevents tech-stack
staleness across long features
- Local-only repo detection (probe `git remote -v` in Phase 1)
- Cost transparency: documented expected $5-25/feature and added
`MODEL_PROFILE=fast|balanced|thorough` overrides
Structure
- SKILL.md: thin orchestrator (workflow outline + 18 key rules + agent
table + Task-tool intro + cost note)
- references/: 8 files for on-demand detail
- modes-and-strategies.md
- resumability-and-abort.md
- discovery-and-planning.md (Phase 0 + plan templates with FIXED table)
- implementation-loop.md (Phase N + amendments)
- context-management.md (4-tier system)
- final-review-and-merge.md (final phase + rollback)
- agent-spawning.md (Task tool / worktrees / cost / local-only repos)
- tech-stack-analysis.md (slimmed — detection only, references rules
system instead of duplicating best practices)
- agents/: phase-implementer.md and phase-implementer-frontend.md updated
with mandatory Failed Approaches scan; test-writer.md gains division-of-
labor section; other 5 agents unchanged
alexei.dolgolyov
04fe8124fc