최근 완료된 8시간 회차 사이 GitHub 스타 증가량 순
Qwen3.5‑4B 전용 조향형 뉴런 탐색기, 모든 MLP 뉴런을 밀어 결과를 보관·탐색
Rust 코드를 실시간으로 편집하고, 실행 중인 프로그램에서 밀리초 단위로 변화를 확인
TraceRoot는 오픈소스 에이전트‑네이티브 개선 레이어로, 생산 추적·탐지 신호·평가를 코딩 워크플로에 통합합니다
대만 제작, 단일 Go 바이너리 자체 호스팅 AI 에이전트, 도구 작성·테스트·MCP로 Claude Code·Codex 공유
대형 AI 모델을 내 PC에서 실행: Strata 엔진에서 Qwen3.8‑Flash‑Next, GLM 5.3 Flash 등 다양한 모델 지원
Terminal-native AI agent orchestrator with safety gates, repo intelligence, and controlled execution workflows.
저장소 README에서 가져옴
README 전체를 불러오는 중입니다. 아래는 저장된 발췌입니다.
Terminal-native AI agent orchestrator for the mq ecosystem.
Screenshots/gallery: the GitHub Pages demo page shows the current command
surface and release proof.
mq-agent orchestrates; the whole-stack architecture and memory loop are
documented in mqobsidian (source of truth):
docs/architecture/mq-stack.md.
This README is the human-facing canonical project document and is intentionally
complete. Agents should not use it as their first-read surface — read the thin,
instructional entrypoints instead:README 전문 보기 ↗Terminal-native AI agent orchestrator for the mq ecosystem.
mq-agent doctor # check environment
mq-agent score . # README + publish score (no API key)
mq-agent audit . --dry-run # repo audit plan
mq-agent release-check --dry-run # release readiness plan
flowchart TD
A[mqlaunch] --> B[mq-agent]
B --> C[mq-hal\nreasoning]
B --> D[mq-mcp\ntool layer]
B --> E[repo-signal\nrepo intel]
B --> F[Planner]
B --> G[Executor]
B --> H[Verifier]
B --> I[Memory]
B --> J[Safety]
Screenshots/gallery: the GitHub Pages demo page shows the current command surface and release proof.
mq-agent orchestrates; the whole-stack architecture and memory loop are
documented in mqobsidian (source of truth):
docs/architecture/mq-stack.md.
v1.30 adds an evidence-first active-versus-shadow loop for context selection. The shadow path is zero-effect: it measures candidate context without changing the production prompt, routing, tools, repository, durable memory, approvals or task result.
mq-agent feedback status
mq-agent feedback run --task-class repo-review --repo . --task "review release boundaries"
mq-agent feedback compare <feedback-run-id> --fixture path/to/relevance-fixture.json
mq-agent feedback candidates
mq-agent feedback report --task-class repo-review --since 30d --json
flowchart LR
T[repo-review task] --> A[active context]
T --> S[shadow context]
A --> E[immutable evidence]
S --> E
E --> C[deterministic comparison]
C --> P[reviewable candidate]
P --> H[human review]
C --> X[JSON / MCP / HAL / mqlaunch]
Operational metrics alone do not prove a candidate is better. A
CANDIDATE_BETTER verdict requires explicit deterministic relevance
evidence, and candidates never activate production policy automatically.
See Feedback Engine and the client contract.
This README is the human-facing canonical project document and is intentionally complete. Agents should not use it as their first-read surface — read the thin, instructional entrypoints instead:
AGENTS.md — Codex entrypoint (MQ memory read-order)CLAUDE.md — Claude Code entrypoint (MQ memory read-order).mq/context/repo-card.md — repo identity, role, owns / does-not-own.mq/context/integration-map.md — reads-from / writes-to / use-when.mq/context/task-pack.md — task-specific context, generated per task (not committed)How agents should read MQ memory — read order, truth boundary, and token
budgets — is defined by the mqobsidian context contract: see mqobsidian
docs/CONTEXT_CONTRACT.md and docs/TOKEN_BUDGET.md. mqobsidian owns the
contract; mq-agent is a consumer.
Select existing skills for a task with local, deterministic routing:
mq-agent skills route "fix CI and open PR" --repo mq-agent --target codex --explain
Supports --json, observed Codex/Claude discovery, and required safety skills.
Requires mqobsidian's skill-selection contracts. See Skill Selection.
Most AI coding tools either wrap a model around shell commands or hide execution behind a chat UI.
mq-agent is different: it treats agent work as a controlled terminal workflow with explicit planning, tool routing, verification, memory and safety gates. Every action is traceable, every operation is gated, and the model never runs unsupervised.
mq-agent is useful when you want to:
Not an AI wrapper script. An actual orchestrator with:
| Layer | Responsibility |
|---|---|
| Planner | Decomposes goals into steps |
| Executor | Runs steps through tools |
| Verifier | Checks each result with AI |
| Memory | Session + persistent state |
| Safety | Enforces operation modes |
# Install
uv pip install -e ".[dev,signal]"
# Check environment
mq-agent doctor
# Score a repo (no API key needed)
mq-agent score .
# AI-backed repo assessment (local Qwen via Ollama by default)
mq-agent signal .
# Read-only audit
mq-agent audit .
main is protected. Use a branch and pull request for all development and
release-prep changes:
git switch -c chore/release-vX.Y.Z
mq-agent release-check --dry-run
git push -u origin chore/release-vX.Y.Z
gh pr create --base main --head chore/release-vX.Y.Z
Merge the PR only after CI is green. Create release tags from the merged
main commit.
uv pip install -e ".[dev,signal]"
pipx install git+https://github.com/MCamner/mq-agent.git
pipx install mq-agent # coming soon
Planning and verification default to local qwen3.5:9b via Ollama and need no
OpenAI credentials. Select MQ_AGENT_PROVIDER=openai explicitly for cloud
planning; OpenAI-backed commands, including mq-mcp reviews, require OpenAI
credentials. score, doctor, repo-summary and tools work without them.
$ mq-agent score .
╭──────────────────────── README Score ─────────────────────────╮
│ README score: 100/100 [██████████] │
│ ✓ title ✓ install ✓ usage ✓ examples ✓ badges │
│ ✓ license ✓ roadmap ✓ contributing (none missing) │
╰───────────────────────────────────────────────────────────────╯
╭──────────────────── Publish Checklist ────────────────────────╮
│ Publish checklist: 16/16 [PASS] │
│ Repo looks publish-ready from the static checklist. │
╰───────────────────────────────────────────────────────────────╯
See docs/DEMO.md for the full end-to-end walkthrough.
mq-agent audit . # Read-only repo audit
mq-agent release-plan # Show release plan
mq-agent release-check # Validate release readiness (suggest mode)
mq-agent release-check --approve # Execute checks
mq-agent ship status --target 1.25.0 # Read-only release decision
mq-agent ship proof --target 1.25.0 # Release evidence
mq-agent ship audit --target 1.25.0 # Post-release audit
mq-agent repo-summary . # Quick repo overview
mq-agent run "pytest" --approve # Safe shell execution
mq-agent fix-ci # Diagnose CI failures
mq-agent doctor # Check environment
mq-agent tools # List registered tools
mq-agent tools --mcp # Include discovered MCP tools
mq-agent tools --describe <name> # Show tool metadata and safety class
mq-agent mcp status # Check mq-mcp reachability and tool counts
mq-agent mcp tools # List all MCP tools with safety classes
mq-agent run-tool <tool> # Run an MCP tool through safety gates
mq-agent review file <path> # Review one file through mq-mcp
mq-agent review diff # Review current diff through mq-mcp
mq-agent review repo [path] # Review repo through mq-mcp
mq-agent review file <path> --fast # Prefer Class A tools (mq-mcp routes)
mq-agent review diff --fast
mq-agent review file <path> --architecture-image docs/arch.png
mq-agent learn status # Check mq-mcp learn system
mq-agent learn search <query> # Search learned review patterns
mq-agent learn explain <pattern> # Fetch pattern explanation
mq-agent dashboard # Operator snapshot: stack, brain, Ollama, contracts
mq-agent stack loop # Controlled autonomous stack loop
mq-agent stack loop --execute --approve
mqlaunch agent → 19 # Manual stack loop plan from the menu
mq-agent tui # Launch Textual dashboard
# Examples of commands with explicit --dry-run support
mq-agent audit . --dry-run
mq-agent run-tool read_repo_file --arg path=README.md --dry-run
Review commands are pass-through orchestration for mq-mcp review tools. They
route to review_file, review_diff, review_repo, or supported risk review
tools through MCPBridge; mq-agent does not implement local review logic,
severity scoring, architecture reasoning, or risk classification.
read-only → only read tools allowed
suggest → plan shown, nothing executed (default for most commands)
execute → runs after safety check (requires --approve for destructive ops)
dangerous → no restrictions (requires explicit flag)
mq_agent/
├── core/
│ ├── state.py # AgentState, PlanStep, SafetyMode, StepStatus
│ ├── planner.py # Plan generation via selected provider
│ ├── executor.py # Tool routing + safety enforcement
│ ├── verification.py # Result verification via selected provider
│ ├── memory.py # Session + persistent memory
│ └── safety.py # Safety gate (mode enforcement)
├── tools/
│ ├── git_tools.py # git_status, git_log, git_diff, git_branch, git_remote
│ ├── shell_tools.py # run_command (blocked pattern list)
│ ├── repo_tools.py # repo_summary, list_files, read_file, find_files
│ ├── signal_tools.py # repo_scan, repo_readme_score, repo_publish_checklist, repo_analyze
│ └── mcp_bridge.py # HTTP bridge to mq-mcp
├── agents/
│ ├── audit_agent.py # Repo audit (read-only)
│ ├── release_agent.py # Release validation
│ ├── ci_agent.py # CI failure diagnosis
│ ├── docs_agent.py # Documentation audit
│ └── signal_agent.py # repo-signal static scan + AI improvement plan
├── tui/
│ └── app.py # Textual dashboard
├── prompts/
│ ├── planner.md # System prompt for Planner
│ └── verifier.md # System prompt for Verifier
└── tasks/
├── release.yaml # Declarative release task
├── audit.yaml # Declarative audit task
└── fix_ci.yaml # Declarative CI fix task
uv run pytest -v — no OpenAI calls requiredmq-agent score . — 100/100 README, 16/16 publish checklist [PASS]mq-agent doctor — all required checks passmq-agent audit . --dry-run — safe, read-only plan generationmq-agent signal . --dry-run — repo-signal assessment without execution--dry-run and --json are command-specific; check mq-agent <command> --help.
For example, doctor supports neither flag; tools supports --json but not --dry-run.uv run pytest tests/ -v
ship status, ship proof, and ship auditmq_release_cockpit.v1 JSON contractstack release remains the release engineReal-run collection is now a separate content-addressed layer over the v2 measurement engine.
mq.hybrid-retrieval-suite.v1 describes operator-selected task casesmq.hybrid-retrieval-evidence-set.v1 binds exact live run fingerprintsmemory hybrid-evidence collect executes fixture-backed live channelsmemory hybrid-evidence status re-verifies every stored run referenceHybrid Retrieval v2 broadens the existing zero-effect retrieval shadow into a measured multi-channel surface while keeping mq-mcp semantic memory authoritative.
mq.hybrid-retrieval.v2 contract; v1 remains readable and callablepromotion_eligible=falsecontext pack --codegraph auto behavior is unchangedFeedback activation readiness no longer treats two comparisons on two snapshots as sufficient promotion evidence.
mq.execution-outcome.v1 recordsWorkflow continuity now uses bounded resume references over the existing workflow run state instead of capturing conversations or cloning runtime state.
mq.workflow-checkpoint.v1workflow checkpoint resume --dry-run performs zero-write verification--approvePortable runtime recovery now covers deterministic Notebook metadata/checkpoint
state without turning mq-agent state into a general backup mechanism.
mq.state-snapshot.v2 manifests are content-addressedsha256: fingerprint evidencemq.state-snapshot.v1 snapshots remain readable/restorablestate restore --dry-run verifies and resolves targets without writes--approveThe visual review path is now contract-bound end to end rather than stopping at perception extraction.
perception.v1mq.perception-review.v1mq.perception-review-receipt.v1 binds perception, review,
reviewer runtime and an exact repo commit when requestedmq-agent review perception orchestrates the chain and fails closed on
malformed/mismatched evidence or a refused requested receiptreview_perception toolv1.36.0 adds changed-file verification with receipts and a machine-readable memory review queue.
verify-changes runs the repo-owned checks in .mq/verification.json for
files changed against a base, and binds the result to the working tree in a
receipt under .git/verify-changes --check-receipt validates the last receipt without
running tests; repo-signal doctor --verification reads itmemory review-status --json forwards mqobsidian's
mqobsidian.memory-status.v1; macos-scripts Pulse reads the held queues from itskipped when the local provider omits
the cloud probetools/model_runtime.py send
think: false to Ollama for every modelv1.35.0 checks every MCP call mq-agent makes against what the producers publish, and hardens the gates around evidence.
stack compatibility checks MCP tool names, classes and arguments against
each producer's committed tool surface (MQC019–MQC023), without a server--live compares the running servers with their committed surface
(MQC024, WARN only); --repo-path NAME=PATH points the gate at another checkoutlearn store <candidate-id> --validation … stores one explicitly chosen
inbox candidate; learn extract and learn list feed and show the inboxreview repo --risk uses mq-mcp's risk_review_repo; a failed review
exits 1route divergence holds the new era to ten complete successful runs per
route across fourteen days; automatic routing stays disabledv1.34.0 consolidates the work landed since Policy Registry v2: the controlled activation gate proven end to end, evidence-gated Hybrid Retrieval, Session Checkpoints, State Recovery v2, Perception Review v2 and measured execution completeness.
repo-review evidence: approval,
Canary v2, promotion, activation, post-activation check and exact rollbackfeedback activate needs both receiptsreview repo records repo-review execution outcomes, fails closed on
unmeasured coveragecontext prepare rebuilds views, exports context and verifies in one stepPolicy Registry v2 makes task-class policy state explicitly recoverable and addressable.
mq.feedback-policy-snapshot.v1feedback policy surfaces current registry snapshot--codegraph auto semantics remain unchangedCanary v2 replaces one-shot activation evidence with a bounded append-only PLAN/RESULT contract.
mq.feedback-canary.v1 is packaged and declared by mq-agentfeedback canary-plan binds candidate, approval and policy snapshotfeedback canary-run executes a bounded zero-effect experiment budgetfeedback canary-status reads authoritative PLAN/RESULT evidence--codegraph auto remains unchangedEvidence-Gated Activation and Recovery. v1.31 keeps v1.30's evidence-first Feedback Engine and adds an explicit human-controlled path from a ready candidate to a bounded canary, one task-class activation, post-activation regression checking and deterministic rollback.
MQ_FEEDBACK_ACTIVATION=off kill switchcontext pack --codegraph auto behavior remains unchanged;
feedback policy is consumed only with --codegraph policyEvidence-Grounded Feedback and Governance Hardening. mq-agent can measure zero-effect active-versus-shadow context experiments, derive bounded improvement candidates, evaluate activation readiness without changing policy, persist exact-code review receipts, and verify that protected-main workflows do not depend on direct branch writes.
NotebookLM Knowledge Pipeline. The NotebookLM archive is now a governed MQ knowledge source with deterministic corpus accounting, local retrieval, incremental semantic sync, evidence provenance, and an Atlas-safe handoff.
Runtime Provenance. Every identity layer — checkout, integration, remote, installed, running, release — is observed separately and compared one edge at a time, so an execution record can say which build produced it.
mq-agent stack release --all — multi-repo release plan over the single-repo primitive; every repo categorized ready / blocked / up-to-dateorigin — the shape that otherwise aborts mid-release after the commitstack release --all --preflight — read-only refusal surface: dirty, off-main, unpushed, tag exists, version mismatch, and each repo's own release-check.shrepo_release_check.v1 contract adopted by all 8 stack repos behind one canonical read-only entrypointstack release --all --execute --approve — fail-fast gate, dependency-ordered execute, stop-on-first-failure; repos after a failure are SKIPPED and an already-released repo is never rolled backrelease.sh syncs .mq/repo-contract.json or aborts, so the contract cannot drift regardless of which path runsmq-agent obsidian inbox list|read|rank — the promotion inbox surface mqlaunch delegates to, read only through mqobsidian's canonical truth-export-index.json; no raw-vault fallbackmq-agent obsidian promote|reject|defer|rollback|deprecate — five explicit verbs, preview by default, --confirm to apply; writes delegated to mqobsidian's write-ahead journalinbox_promotion_orchestration.v1 in the repo contract, validated against the payload the CLI actually emitsrepo-signal / mq-mcp evidence adapters, pending producer contracts (mqobsidian DEC-004)mq-hal reads: stack cockpit, stack brain-gate, run --stack, stack release-check, dashboardmq-hal stack, brain-status, release-status and next-action depend onmq_hal_operator_contract.v1 in the repo contract, plus the documented mq-agent → mq-hal read contractmq_stack_loop_audit.v1 — approved stack-loop executions append one audit record for read-only display by mq-halmq-agent stack loop — controlled planning and approved one-step executionmq-agent stack loop planningmq-agent memory ingest — local mqobsidian Markdown indexmq-agent memory query / memory search-vault — read-only vault searchmq-agent memory summarize — section summary across truth, reviews, learn, releases, architecture and decisionsmq-agent memory link — read-only link candidates between notesdocs/MEMORY_ENGINE.mdmq-agent dashboard — read-only operator snapshot for stack health, contracts, mqobsidian truth freshness and Ollama profile statusmq-agent tui — starts with the same operator snapshot before command executionmq-agent models — first-class Ollama model runtime command groupmq-agent models list/current/doctor/switch/bench~/.mq-agent/models.jsonmq-agent stack run surfaces the active model profile in the Ollama checkmq-agent stack run — one runtime gate for repo-signal, mq-mcp, Ollama, brain export rendering and release readinessmq-agent run --stack — canonical root alias for the stack runtime pipeline--json and --markdowndocs/MQ_CONTROL_PLANE.md — system map for signal, review, learn, memory and releasemq-agent stack cockpit — one merged view of the whole stack: version, branch, contract, release gate, brain-export freshness, next action per repo--dry-run never writes, --json machine-readable, --brain respects --dry-run, --approve required for write flows (locked by tests/test_flag_contract.py)mq-agent brain structure — standard mqobsidian export structure: check, --init --approve, legacy detectionmq-agent stack brain-gate — pre-release checklist: contract-check + release-check + truth-export dry-run + vault structure + review→brain write pathsignal --brain --dry-run no longer writes to the braindocs/STACK_COCKPIT.md, docs/VAULT_STRUCTURE.md, docs/BRAIN_GATE.md — reference docsmq-agent stack release --repo <name> — orchestrated single-repo release pipeline--execute applies; --bump patch|minor|major or explicit --versiontruth-export to mqobsidiandocs/STACK_RELEASE.md — reference docmq-agent stack truth-export — durable stack truth note (contract + release gates) to mqobsidianmq-agent stack export kept as backwards-compatible alias~/mqobsidian/memory/stack-truth/YYYY-MM-DD-mq-stack-truth.mdstack contract-check --ci / stack release-check --ci — CI mode with SKIPPED for missing reposmq-stack-gate.yml split: fast --ci gate on PRs, full multi-repo gate on main/nightlydocs/STACK_TRUTH_EXPORT.md — reference doc.github/workflows/mq-stack-gate.yml — CI workflow for MQ stack gates~/... pathsmq-agent stack contract-check --json runs on pull requests and pushes to mainmq-agent stack release-check --json runs on pull requests and pushes to mainmq-agent stack contract-check — validates .mq/repo-contract.json across all stack reposmq-agent stack contract-check --json — machine-readable output_contract_entry() helper with READY / REVIEW / DRIFT / BLOCKED status modelschemas/mq_stack_repo_contract.schema.json — JSON Schema for contract manifests.mq/repo-contract.json deployed to all 8 MQ stack reposdocs/STACK_CONTRACT_GATE.md — reference docmq-agent stack release-notes — draft notes from git commits since last tag, per repomq-agent stack release-notes --repo <name> — single repo filtermq-agent stack release-notes --json — machine-readable outputdocs/STACK_RELEASE_NOTES.md — reference docmq-agent stack report — score, trend, alert, readiness per repo in one tablemq-agent stack report --json — machine-readablemq-agent stack release-check — local release gate across all stack repos (exit 1 on blocker)mq-agent stack release-check --json — GO/NO-GO with per-repo detaildocs/STACK_REPORT.md — reference with example output and workflowmq-agent stack alert — exits 1 when a repo drops ≥ threshold or falls below min-scoremq-agent stack sweep --alert — inline alert check at the end of a sweep--threshold N / --min-score N — configurable thresholds--json — machine-readable alert list, CI-friendly exit codesdocs/STACK_ALERT.md — reference with CI integration examplesmq-agent stack sweep appends every run to ~/.mq-agent/sweep-history.jsonlmq-agent stack history — tabular trend view across last N sweepsmq-agent stack history --diff — delta table between last two sweepsmq-agent stack history --json / --limit N — scripting and filteringdocs/STACK_HISTORY.md — reference with example output and JSON scriptingmq-agent stack sweep — loop repo-signal over all mq-stack repos in one passmq-agent stack sweep --brain — write a brain note per repo to mqobsidianmq-agent stack sweep --decide — consolidated health ADR via brain_record_decisionmq-agent stack sweep --dry-run — preview without writesmq-agent stack sweep --json — machine-readable summary tabledocs/STACK_HEALTH.md — multi-repo sweep reference with example outputmq-agent signal . --brain → mq-agent review repo . --brain → mq-agent release-check --dry-runmqlaunch/commands/demo-flow.sh — standalone chain scriptdocs/DEMO.md rewritten as canonical v1.5.0 referencemq-image-analyze MCP endpoint registered for visual perception toolsmq-agent run-tool observe_architecture routes through mq-image-analyzemq-agent run-tool image_ocr routes through mq-image-analyzemq-agent review file|diff|repo --architecture-image <path> adds visual_architecture_observation.v1 contextlist_architecture_decisions)--fast flag on all review commands — mq-mcp routes to Class A toolsmq-agent learn status/search/explain — read-only access to mq-mcp learned patternsMultiMCPBridge._call_optional_tool() — silent None when tool not availableuv run pytest -v — no OpenAI calls requiredmq-agent review file/diff/repo — pass-through review orchestration via mq-mcp--security, --architecture, --risk flags forwarded to mq-mcp review contracts--dry-run on all review commands — shows planned mq-mcp call without executingvalidate_orchestration_contract in mq-agent doctoruv run pytest -v — no OpenAI calls required for test suitecli/render.py + core/diagnostics.py — orchestration logic extracted from main.pyPlanner wired to MqAgentConfig.effective_model() — config-driven model selectionuv run pytest -v — no OpenAI calls required for test suitemq_agent/config.py — MqAgentConfig dataclass with load/save
and MQ_AGENT_MODEL env overrideuv run pytest -v — no OpenAI calls requiredmq-agent task list/rundocs/ARCHITECTURE.mduv run pytest -v — no OpenAI calls requiredmq-agent doctor — full environment checkmq-agent mcp status / mq-agent mcp tools — MCP inspectionmq-agent tools --describe <name> — tool metadata and safety classmq-agent run-tool <tool> — MCP tool through safety gatesscripts/smoke-mqlaunch.sh — verifies mqlaunch agent ... reaches mq-agentdocs/MQLAUNCH_INTEGRATION.md — bridge architecture and usagedocs/COMMAND_SURFACE.md — single source of truth for command countsuv run pytest -v — no OpenAI calls requiredmq-agent v0.5.0 adds semantic repository memory via repo-signal.
mq-agent memory status # check vector store and repo-signal
mq-agent memory doctor # diagnose environment with actionable fixes
mq-agent memory build . # preview upload (dry-run default)
mq-agent memory refresh . --approve # upload when ready
mq-agent memory status --json # machine-readable output
Memory upload is explicit and gated. mq-agent never uploads silently.
$ mq-agent memory status
╭────────────────────────────── Semantic Memory ───────────────────────────────╮
│ status: ready │
│ vector store: vs_69ffa9a4ef5c81919d7d237c3ecdc260 (canonical) │
│ repo-signal: available │
│ repo: /path/to/mq-agent │
╰──────────────────────────────────────────────────────────────────────────────╯
$ mq-agent memory build .
[dry-run] Would run: repo-signal semantic-upload
Add --no-dry-run to execute, or use memory refresh --approve.
$ mq-agent memory doctor
╭──────────────────────── Memory Doctor ───────────────────────────╮
│ ✓ vector store: vs_69ffa9a4ef5c81919d7d237c3ecdc260 (canonical) │
│ ✓ repo-signal: available │
│ ✓ repo path: /path/to/mq-agent │
╰──────────────────────────────────────────────────────────────────╯
mq-agent is fully wired into mqlaunch as both a direct command surface and
an interactive menu.
mqlaunch agent
mqlaunch agent score .
mqlaunch agent audit .
mqlaunch agent doctor
mqlaunch agent release-check --dry-run
mqlaunch agent mcp-status
mqlaunch agent mcp-tools
See docs/COMMAND_SURFACE.md for the canonical command surface and docs/MQLAUNCH_INTEGRATION.md for bridge details.
See docs/ROADMAP.md for the current roadmap.
Current release is v1.36.0: changed-file verification receipts and a machine-readable memory review queue.
The next focus is collecting current-era docs-review routing evidence:
ten complete successful runs per route across at least fourteen days.
Automatic routing remains disabled pending explicit operator promotion review.
The current development tree also provides verify-changes: repo-owned checks
and source-bound receipts, with stale evidence detected before reuse.
See verification rules and limits.
Across these workflows, mq-agent owns orchestration, mqobsidian owns durable memory, mq-mcp owns review and runtime reasoning, and mqlaunch remains a thin delegation surface.
Do not commit API keys, local secrets, or private environment files.
When multiple Git worktrees exist, pass a repeatable --repo-path NAME=PATH
option to mq-agent stack compatibility to select the checkout used for both
committed tool-contract validation and optional --live comparison:
mq-agent stack compatibility --repo-path "mq-mcp=$HOME/mq-mcp-main" --live
Unknown repository names, duplicate overrides and missing directories are rejected. Without this option, repository discovery is unchanged. The option only reads the selected worktrees and does not switch branches or modify them.
Routing, planning, verification and all local profiles default to qwen3.5:9b.
Local failures do not fall back to OpenAI. Nomic embeddings are unchanged.
Use MQ_AGENT_PROVIDER=openai MQ_AGENT_MODEL=<cloud-model> mq-agent plan <goal>
only for an explicitly selected cloud task; the existing OpenAI credentials are
required for that selection. Set MQ_AGENT_VERIFIER_MODEL separately when a
cloud task needs a different verification model. Credentials are not required
for local inference. Routing allowlists and execution safety gates are unchanged.
OpenAI vector-store status/refresh also requires explicit
MQ_AGENT_PROVIDER=openai. Existing remote stores are preserved; local Nomic
indexes are not rebuilt or changed by this consolidation.
Run repo-owned checks selected from changed files and validate local receipts:
mq-agent verify-changes . --plan
mq-agent verify-changes .
mq-agent verify-changes . --check-receipt
Verification rules, receipts and limits.
mqlaunch verify-changes delegates to the same command with the caller's repo.
mq-agent memory status / build / refreshmq-agent memory doctor — diagnose memory environment with actionable fixesmq-agent memory status --json / mq-agent memory doctor --json — machine-readable output