CLike Documentation
CLike is an AI-native development pipeline built around three runtime surfaces:
- VS Code extension for chat, Harper commands, code actions, local agent execution, RAG attachment handling, and Git-assisted promotion.
- Orchestrator for workflow control, repository-aware preparation, RAG APIs, eval/gate orchestration, MCP exposure, and gateway delegation.
- Gateway for model catalog resolution, provider abstraction, Harper prompt execution, embeddings, and telemetry APIs.
This documentation reflects the current source tree in the inspected package and treats code as the source of truth.
Documentation map
INSTALL.mdβ requirements, configuration, services, extension, local agents, MCP for Claude Code / Codexchat-ui.mdβ the chat in VS Code: what it is and what it does (cockpit, run card, Stop, history, autocompletion), with its pictureGET_STARTED.mdβ first project, agent chat, auto-eval, acceptance-first, CLike from Claude Code / Codex, troubleshootingarchitecture.mdβ system structure and runtime boundariessetup-and-runtime.mdβ local services, ports, configs, and startup modelcommands.mdβ current slash commands, code actions, and extension commandsharper-workflow.mdβ IDEA β SPEC β PLAN β KIT β EVAL β GATE β FINALIZE, then/idea --extendfor the next releaseauto-eval.mdβ/eval REQ --fix: governed KIT β EVAL repair loop, regression of promoted REQs, gate warnings policy, extra cycles while improvingartifacts.mdβ canonical docs, run artifacts, candidate artifacts, and promotion fileslocal-agents.mdβ current local execution path, GPT Codex / Claude Code integration, and restrictionscontext-sources.mdβ.clike/context.yaml: project files (requirements, UI references, data, β¦) read by the phases, release folders (docs/inputs/vN/,docs/reference/vN/), formats and rulesrag.mdβ indexing, search, fetch, attachment strategy, repository grounding, and the Memory panel (project memory and CLike guide)git-and-promotion.mdβ current Git behavior, stacked REQ branches, gate auto-merge promotion (3-way, conflicts, report), merge flowmcp.mdβ MCP servers (orchestrator, extension), tools and authenticationhowto-mcp-project.mdβ run a whole project from Claude Code or Codex through MCP (setup, the phase loop, next release, troubleshooting)clike-at-a-glance.mdβ CLike at a glance: the picture of the whole, area by areahowto-mcp-bmad-claude.mdβ from/initto/finalizewith BMAD, driven by Claude Code through MCPapi-reference.mdβ endpoint census for orchestrator, gateway, and agent surfacesagent-operating-model.mdβ developerβagent and agentβCLike operating models, methodology profilesarchitecture/ADOPT_AND_IDEA_EXTEND.mdβ/idea --extend(implemented: release cascade of a delivered project with ADDED/MODIFIED/REMOVED requirements) and the/adoptproposal for brownfield repositories (not implemented)development.mdβ developer guide: local runs, tests, golden snapshots, lint baseline, hygiene, CI, rebuild, eval sandbox and eval environment (docker/eval.env, provisioned variables, test databases, services of other REQs)../SECURITY.mdβ security model, known limitations, vulnerability reporting../CHANGELOG.mdβ release notes
Current product shape
Extension
The VS Code extension currently provides:
- Chat modes: free, coding, harper
- Harper slash commands and webview-based execution
- AI code actions for docstrings, refactoring, tests, and fix flows
- File attachments with inline vs RAG paths
- Local agent execution for compatible Harper flows
- Git helpers for branch, commit, PR, and promotion
- Eval/Gate integration from the chat UI and commands
Orchestrator
The orchestrator currently provides:
- Legacy
/v1/chatand/v1/generate(/v1/applywas removed in the 2026-10 hardening) - Harper workflow APIs under
/v1/harper/* - RAG APIs under
/v1/rag/* - Eval/Gate endpoints
- Agent endpoint for code-action style tasks
- Router resolution endpoint
- Optional read-only MCP server mounted at
/mcp - Eval sandbox (isolated container), eval environment provisioning, eval tools catalog (
/v1/eval/tools)
Gateway
The gateway currently provides:
- Model catalog and validation
- Provider-routed chat completions
- Embeddings
- Harper phase execution via
/v1/harper/run - Telemetry APIs and telemetry UI
What is current vs experimental
Current
The following are implemented in the inspected sources:
- Harper phases and slash-command entrypoints
- Candidate artifact generation under
runs/kit/<REQ-ID>/... - RAG index/search/fetch APIs
- Local agent integration with Claude Code and GPT Codex
AGENT_EXECUTION_CONTEXT.json- Eval pre-pass support for local agents
/eval --fix//gate --fixrepair loops, regression of promoted REQs, gate auto-merge promotion, merge-on-gate/finalizesolution verification (ci/SOLUTION_LTC.json,FINALIZE_REPORT.md, local run scripts)/idea --extendrelease cascade of a delivered project- MCP server mounted by orchestrator when enabled; extension MCP server for Claude Code / Codex
- Gateway model routing and provider abstraction
- Telemetry endpoints for Harper metrics
Present but restricted
The following are implemented with explicit guardrails:
- Local agents run every Harper phase (IDEA β¦ FINALIZE,
/extend,/idea --extend) under the phase write policy; the verdict of/evaland/gateis always the canonical sandbox eval, the local/evalpath is a pre-pass - The orchestrator MCP server is read-only; the extension MCP server runs phases,
eval_runandgate_checkthrough the governed chat flow, without overrides, promotion or Git
Not documented as current product guarantees
The inspected codebase does not justify treating these as stable user-facing guarantees:
- MCP tools that write files, promote or run Git
- Remote external agent orchestration as a production contract
Naming and compatibility notes
The sources currently show both legacy and normalized naming in the local-agent area.
Execution preference
The extension normalizes legacy values:
prefer_claude_codeβprefer_local_agentclaude_code_onlyβlocal_agent_onlyautoβcloud_onlyhybridβprefer_local_agent
Current normalized execution preference values are:
cloud_onlyprefer_local_agentlocal_agent_only(default ofclike.execution.defaultPreference)
Local executors
Current supported local executors are:
claude_codegpt_codex
Both are exposed as local agent executors behind the same extension workflow.
Recommended reading order
For onboarding, read in this order:
architecture.mdagent-operating-model.mdsetup-and-runtime.mdcommands.mdharper-workflow.mdartifacts.mdapi-reference.mddevelopment.md(contributors)