mirror of https://fastgit.cc/github.com/Yeachan-Heo/oh-my-claudecode synced 2026-04-21 13:20:34 +08:00

Files

Bellman 8641e541c6 refactor(agents): consolidate 4 overlapping agent pairs (22→18 agents) (#1426 )

* refactor(skills): eliminate 5 thin wrapper skills + CLAUDE.md diet

Part 1: Delete 5 thin wrapper skills that duplicated agent prompts:
- analyze -> debugger/architect (merged unique content)
- build-fix -> build-fixer (no unique content)
- tdd -> test-engineer (merged TDD enforcement)
- code-review -> code-reviewer (merged review checklist)
- security-review -> security-reviewer (merged OWASP/checklists)

Keyword routing preserved: tdd/analyze keywords now route through
bridge.ts message constants (TDD_MESSAGE, ANALYZE_MESSAGE) instead
of skill invocations in keyword-detector.mjs.

Part 2: Trim docs/CLAUDE.md from 177 to 83 lines (53% reduction).
Agent catalog compressed to bare name+model list. Skills section
compacted. All section headers preserved.

Updated: tests, REFERENCE.md (en/ko), MIGRATION.md (en/ko),
skills/AGENTS.md, skill-state protection map, keyword-detector
priority array.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(agents): consolidate 4 overlapping agent pairs (22→18 agents)

Merge 4 agent pairs using always-thorough strategy (no flags, no modes):
- code-reviewer absorbs quality-reviewer (adds style, performance, quality strategy modes)
- critic absorbs harsh-critic (adds 5-phase investigation, adaptive harshness, gap analysis)
- debugger absorbs build-fixer (adds build/compile error handling, lsp_diagnostics_directory)
- executor absorbs deep-executor (adds task classification, explore-first protocol, circuit breaker)

Registry changes:
- Remove quality-reviewer, deep-executor, build-fixer from agent registry
- Remove harsh-critic conditional registration (enableHarshCritic feature toggle deleted)
- Agent count: 22→18 (deterministic, no conditional agents)

Backward compatibility:
- DEPRECATED_ROLE_ALIASES in delegation routing for all 4 absorbed agents
- HUD badge fallbacks map old names to survivor abbreviations

Updates across: delegation routing, permission handler, autopilot prompts,
HUD elements, team skill, docs, tests, and config schema.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* chore(review): sync generated artifacts for agent consolidation

* fix(review): restore compatibility contracts after agent consolidation

---------

Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

2026-03-08 12:00:54 +09:00

3.7 KiB

Raw Permalink Blame History

Mode Selection Guide

Quick Decision

If you want...	Use this	Keyword
Clarify vague requirements first	`deep-interview`	"deep interview", "ouroboros", "don't assume"
Full autonomous build from idea	`autopilot`	"autopilot", "build me", "I want a"
Parallel autonomous (3-5x faster)	`team` (replaces `ultrapilot`)	`/team N:executor "task"`
Persistence until verified done	`ralph`	"ralph", "don't stop"
Parallel execution, manual oversight	`ultrawork`	"ulw", "ultrawork"
Cost-efficient execution	`` (modifier)	"eco", "budget"
Many similar independent tasks	`team` (replaces `swarm`)	`/team N:executor "task"`

Note: ultrapilot and swarm are deprecated — they now route to team mode.

If You're Confused or Uncertain

Don't know what you don't know? Start with /deep-interview - it uses Socratic questioning to clarify vague ideas, expose hidden assumptions, and measure clarity before any code is written.

Already have a clear idea? Start with autopilot - it handles most scenarios and transitions to other modes automatically.

Detailed Decision Flowchart

Uncertain about requirements or have a vague idea?
├── YES: Use deep-interview to clarify before execution
└── NO: Continue below

Want autonomous execution?
├── YES: Is task parallelizable into 3+ independent components?
│   ├── YES: team N:executor (parallel autonomous with file ownership)
│   └── NO: autopilot (sequential with ralph phases)
└── NO: Want parallel execution with manual oversight?
    ├── YES: Do you want cost optimization?
    │   ├── YES: eco + ultrawork
    │   └── NO: ultrawork alone
    └── NO: Want persistence until verified done?
        ├── YES: ralph (persistence + ultrawork + verification)
        └── NO: Standard orchestration (delegate to agents directly)

Have many similar independent tasks (e.g., "fix 47 errors")?
└── YES: team N:executor (N agents claiming from task pool)

Examples

User Request	Best Mode	Why
"Build me a REST API"	autopilot	Single coherent deliverable
"Build frontend, backend, and database"	team 3:executor	Clear component boundaries
"Fix all 47 TypeScript errors"	team 5:executor	Many independent similar tasks
"Refactor auth module thoroughly"	ralph	Need persistence + verification
"Quick parallel execution"	ultrawork	Manual oversight preferred
"Save tokens while fixing errors"	+ ultrawork	Cost-conscious parallel
"Don't stop until done"	ralph	Persistence keyword detected

Mode Types

Standalone Modes

These run independently:

autopilot: Autonomous end-to-end execution
team: Canonical orchestration with coordinated agents (replaces ultrapilot and swarm)

Deprecated: ultrapilot and swarm now route to team mode.

Wrapper Modes

These wrap other modes:

ralph: Adds persistence + verification around ultrawork

Component Modes

These are used by other modes:

ultrawork: Parallel execution engine (used by ralph, autopilot)

Modifier Modes

These modify how other modes work:

****: Changes model routing to prefer cheaper tiers

Valid Combinations

Combination	Effect
`eco ralph`	Ralph persistence with cheaper agents
`eco ultrawork`	Parallel execution with cheaper agents
`eco autopilot`	Autonomous execution with cost savings

Invalid Combinations

Combination	Why Invalid
`autopilot team`	Both are standalone - use one
`` alone	Needs an execution mode to modify

3.7 KiB Raw Permalink Blame History