- Create dedicated synthesizer agent to handle research synthesis
- Move SUMMARY.md creation from main orchestrator to subagent
- Reduces main context usage (~2000+ lines of research output)
- Synthesizer reads 4 research files and produces unified summary
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Consolidates 4 separate commands into one unified flow:
- /gsd:new-project now handles: questioning → research → requirements → roadmap
- Creates gsd-roadmapper agent for heavy lifting (goal-backward, coverage validation)
- Adds atomic commits after each stage for crash recovery
- Deprecates standalone research-project, define-requirements, create-roadmap
(kept for mid-project use)
Fixes from audit:
- Add requirements quality criteria (specific, user-centric, atomic, independent)
- Add milestone context to research prompts (greenfield vs subsequent)
- Add quality gates per research dimension
- Add template references for consistent output format
Removes deprecated gsd-researcher.md (replaced by project/phase researchers)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Research decisions made at roadmap creation time are:
- Premature: Can't know if research is needed before discuss-phase
- Redundant: --skip-research flag exists for explicit control
- Counterproductive: Gives Claude an excuse to skip research entirely
Removed:
- Research: Likely/Unlikely fields from roadmap template
- detect_research_needs step from create-roadmap workflow
- "If roadmap flagged Research: Likely" checks from gsd-planner
Research now always runs during plan-phase unless:
- RESEARCH.md already exists (silent reuse)
- --skip-research flag passed
- --gaps flag passed (gap closure mode)
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Split gsd-researcher into two specialized agents:
- gsd-phase-researcher: tailored for phase research before planning
- gsd-project-researcher: tailored for ecosystem research before roadmap
Updated /gsd:plan-phase to auto-research before planning:
- Research runs if no RESEARCH.md exists (silent use if exists)
- --research flag forces re-research
- --skip-research flag bypasses research entirely
- Researchers commit their output
This enables single-command workflow: /gsd:plan-phase does
research → plan → verify in one orchestrated flow.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Adds back guidance that was lost during consolidation:
- Task sizing heuristic (15-60 min target)
- Signals for too large / too small tasks
- Specificity side-by-side examples table
- "The test" for checking task clarity
This agent was correctly deleted in f3e0e69 (subagents can't spawn
subagents), then accidentally re-added in d07ef33 during install fix.
The command audit-milestone.md is the orchestrator. No agent needed.
research-phase.md:
- Added <key_insight> framing ("what don't I know that I don't know")
- Added <downstream_consumer> explaining plan-phase integration
- Added <quality_gate> checklist
research-project.md:
- Added milestone context (greenfield vs v1.1+)
- Added <downstream_consumer> per dimension
- Added <quality_gate> per agent
- Enhanced roadmap implications guidance
gsd-researcher.md:
- Simplified <gsd_integration> to point to orchestrator-provided context
- Agent is now pure research capability, context is injected
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Reduces context exhaustion during /gsd:map-codebase by having agents
write documents directly instead of returning findings to orchestrator.
- Create gsd-codebase-mapper agent with embedded templates
- Agent understands downstream consumers (plan-phase, execute-plan)
- Parameterized by focus: tech, arch, quality, concerns
- Each focus writes specific documents directly to .planning/codebase/
- Orchestrator receives only confirmation + line counts
Context savings: ~40 lines returned vs potentially thousands of lines
of exploration findings previously transferred back to orchestrator.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Explains why research matters: output directly consumed by plan-phase.
Documents the contract between RESEARCH.md sections and planning workflow.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
The agents folder (gsd-executor, gsd-verifier, gsd-integration-checker,
gsd-milestone-auditor) was not being published to npm because it was
missing from the package.json files array.
Also adds the missing gsd-milestone-auditor.md agent.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Step 0 now explicitly states to skip to Step 3 when in re-verification
mode. Steps 1-2 marked as "Initial Mode Only" to make flow clear.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
When VERIFICATION.md already exists with gaps, verifier now:
- Loads previous must-haves instead of re-deriving
- Focuses deep verification on failed items
- Quick regression check on passed items
- Tracks re-verification metadata (gaps_closed, regressions)
Saves tokens by not starting from scratch after gap closure.
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- Add verify_phase_goal step to execute-phase workflow
- Verifier outputs structured gaps: YAML for planner consumption
- Add --gaps flag to plan-phase for gap closure mode
- Route by verification status: passed, gaps_found, human_needed
- Gap closure creates sequential plans (04, 05...) from verification gaps
- User stays in control at each decision point
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
- gsd-verifier.md: dedicated verification subagent with all logic baked in
- Goal-backward verification: checks codebase, not SUMMARY claims
- Creates VERIFICATION.md, returns status to orchestrator
- execute-phase spawns gsd-verifier instead of general-purpose
Co-Authored-By: Claude <noreply@anthropic.com>
Moves all plan execution logic into a dedicated subagent:
- Deviation rules, checkpoint protocols, commit formatting
- Summary creation, state updates, TDD execution
- Authentication gate handling, continuation support
Installer now copies agents/ to ~/.claude/agents/
Co-Authored-By: Claude <noreply@anthropic.com>
