Apprentice

A local, multi-provider code-delegation pipeline. A master orchestrator (e.g. Claude Code) delegates routine coding to apprentice models — a local model on your own GPU (via Ollama), Gemini (Vertex AI), GPT/Codex (OpenAI), or any OpenAI-compatible endpoint (Groq, OpenRouter, LM Studio, vLLM, …) — then mechanically verifies, corrects, and learns from the results over time via in-context retrieval. The expensive brain is spent on judgment; the cheap brains do the typing.

The economics in one line: local apprentices are free but weaker; cloud apprentices are stronger but metered — so the pipeline verifies everything mechanically (compile → lint → your project's own tests), starts cheap, escalates only on failure, prices every cloud call, and enforces daily budgets.

Two ways to use it:

1. Standalone agent — apprentice chat gives you a coding agent in your terminal, driven by one model of your choice. No orchestrator subscription needed. Every change it makes is verified and auto-reverted if it breaks your tests. → docs/AGENT.md
2. Delegation server — an MCP server exposing delegate, assign, log_correction, so an orchestrator (Claude Code, …) can offload routine coding to cheaper models and stay the judge.

Both share the same engine: providers, mechanical gate, tests, budgets, and the corrections store that makes the workers better over time.

New here? Start with docs/MULTI_AGENT.md — it explains, in beginner terms, what an "agent" is and how the boss + two-worker model fits together.

Note: the project was formerly qwen-pipeline. Its default working directory and the MCP server id are still qwen-pipeline / qwen; only the project brand is Apprentice.

Why this exists

Many orchestrators (Claude Code among them) can only run their own model family — there's no way to point a sub-agent at a local model or at Gemini. So the worker models live behind a local MCP server instead. The orchestrator calls a tool, the server runs the chosen model, the orchestrator reviews the output.

        ┌──────────────────────────────────────────────┐
        │  ORCHESTRATOR  — the "boss" / decision-maker   │
        │  splits tasks, picks provider, REVIEWS output, │
        │  fixes mistakes, logs corrections, commits     │
        └───────────────┬────────────────────────────────┘
                        │ MCP tool call (stdio)
                        ▼
        ┌──────────────────────────────────────────────┐
        │  MCP server  (src/server.py, FastMCP)          │
        │   delegate(task, role, provider?, model?, …)   │
        │   assign(task, done_when, repo, …)             │
        │   log_correction(…)                            │
        └───┬───────────────┬───────────────┬────────────┘
            ▼               ▼               ▼
         qwen            gemini        openai + any
      (Ollama,        (Vertex AI:    openai-compatible
       local GPU,      flash / pro,   endpoint (GPT/Codex,
       FREE)           metered $)     Groq, LM Studio, …)

Providers are config, not code. Any entry in config providers.<name> with a known kind (ollama-local, openai-compatible, vertex-ai) becomes a valid provider= value — adding Groq or a second local model is a 6-line JSON block. See docs/CONFIGURATION.md. Every provider runs through the same gate, retries, retrieval, metering, and budgets.

The "specialized agents" (test writer, C++ implementer, …) are not separate models — they are role values that select a different system prompt for the same worker.

Requirements

• Python 3.11+
• Ollama running locally, with a worker model + an embedder pulled.
• (optional, for the assign file-aware agent) Aider in its own venv.
• (optional, for the Gemini worker) google-genai + Google Cloud Vertex AI credentials.
• An MCP-capable orchestrator (e.g. Claude Code) to drive the tools.

The reference machine is an RTX 5090 (32 GB VRAM) + 64 GB RAM, but the pipeline runs anywhere Ollama can serve a model — scale the worker model to your hardware.

Getting started

Option A — install as a package (quickest)

pipx install git+https://github.com/m-555/Apprentice.git   # or: pip install apprentice-pipeline
apprentice init      # creates the data home (~/.apprentice or $APPRENTICE_HOME),
                     # seeds the config, checks Ollama, prints the MCP registration cmd
apprentice doctor    # environment check any time

# pull the worker + embedder models (scale the worker to your hardware)
ollama pull qwen3-coder-next
ollama pull nomic-embed-text

# register with your orchestrator (Claude Code example; `init` prints this too)
claude mcp add --scope local qwen -- apprentice serve

Config and data live in ~/.apprentice (override with APPRENTICE_HOME). Edit ~/.apprentice/config/qwen.local.json for machine-local values and secrets. The Gemini provider is an extra: pipx install 'apprentice-pipeline[gemini] @ git+https://github.com/m-555/Apprentice.git'.

Option B — clone the repo (for hacking on the pipeline itself)

git clone https://github.com/m-555/Apprentice.git qwen-pipeline
cd qwen-pipeline

# 1. core deps (pinned)
python -m venv .venv
.venv/Scripts/pip install -r requirements.txt          # Windows
# .venv/bin/pip install -r requirements.txt             # Linux/macOS

# 2. pull the worker + embedder models (sizes approximate)
ollama pull qwen3-coder-next        # ~51 GB, Q4_K_M — scale down for smaller GPUs
ollama pull nomic-embed-text        # ~274 MB — for retrieval

# 3. (optional) the file-aware `assign` agent, in an ISOLATED venv
python -m venv .aider-venv
.aider-venv/Scripts/pip install -r requirements-aider.txt

# 4. register the MCP server with your orchestrator (Claude Code example)
claude mcp add --scope local qwen -- ".venv/Scripts/python.exe" "src/server.py"
claude mcp list      # -> qwen ... ✓ Connected

In a checkout, config and data live in the repo (config/, corrections/, outputs/, metrics/) — see docs/CONFIGURATION.md. To add the Gemini worker, see Enabling Gemini below.

The worker model & expert-offload

Expert-offload: the ~48 GB model does not fit in 32 GB VRAM. Ollama keeps the attention/shared layers on the GPU and streams the MoE experts from system RAM. GPU utilisation looks low because it is memory-bandwidth-bound — normal for an MoE with few active params. Warm throughput ≈ 50–58 tok/s; cold load ≈ 55 s.

Warm model: Ollama keeps the model resident for keep_alive after the last request (default 30m here). Requests within that window skip the load (~0.1 s). It's deliberately not infinite — a warm model holds the whole GPU, so a moderate timeout frees it for other work when idle.

Model storage gotcha (Ollama desktop app): the desktop app stores its model location in db.sqlite and, when it spawns the server, sets OLLAMA_MODELS to that value — overriding your env var. If a large ollama pull fills the wrong disk, point both the env var and the app's DB at your intended path, then confirm with ollama list.

The standalone agent

cd /path/to/your-repo
apprentice chat                                    # local free model, verification on
apprentice chat --provider gemini --model pro      # a stronger cloud model
apprentice run "add mul(a,b)" --done-when "pytest -q"   # unattended

It reads, searches, edits, and runs commands in your repo — and after every turn its changes are checked. If they fail your project's tests, they are reverted byte-for-byte and the verbatim failure goes back to the model to fix:

  -> read_file(calc.py)
  -> edit_file(calc.py)
  -> run_tests()
  [OK] verified (tests)

--verify off | gate | tests picks how strict that is; /undo, /provider, /cost and friends work mid-session; a model that keeps failing escalates to a stronger tier automatically. Shell commands are allowlist/denylist-checked and prompt before running. Add --json to stream JSON-lines events instead of text — the integration surface for a VS Code extension, a web UI, or CI. Full guide: docs/AGENT.md.

The MCP tools

delegate(task, role, provider="", context="", model="", repo="", context_files="", apply_to="", apply_mode="append", test_cmd="", return_mode="") -> str

Sends {system: ROLES[role] (+ repo conventions), user: task (+context)} to the chosen provider and returns the generated text, plus a status footer with the gate verdict and an output_id. The pipeline mechanically verifies the output and auto-retries the worker on failure before returning.

• roles: ts_implementer, cpp_implementer, py_implementer, test_writer, refactorer
• providers: qwen (local, default), gemini (Vertex AI), openai (GPT/Codex), or any config-defined provider
• model: optional per-call override — for gemini, "flash" (routine) or "pro" (hard).

Token-cheap mode — the orchestrator's expensive output tokens should never carry code:

• repo= + context_files="src/a.ts src/b.ts:20-80" — send paths, not code; the server reads the content locally (size-capped, path-traversal-guarded) and builds the context block.
• apply_to="src/a.ts" (+ apply_mode: append|create|overwrite) — the server writes the gate-passed code into the real file itself.
• test_cmd="npx vitest run …" — after applying, the server runs your project's own acceptance command in repo; on red it reverts the file, bounces the verbatim test output back to the worker, and retries — a full TDD loop with zero orchestrator tokens. The tree is never left broken. If the tier keeps failing, the task escalates through the cascade (e.g. local → gemini) carrying the failing code + test output, budget-guarded, before it ever reaches the orchestrator (footer shows test_tier= when that happened).
• return_mode="summary" — receive only the status footer + a one-line preview instead of the full code (it's already in the file and the output store).

A routine function then costs the orchestrator roughly: task spec in, two-line footer out.

delegate(task="Add mul(a,b)…", role="py_implementer",
         repo="/path/to/proj", context_files="mathx.py",
         apply_to="mathx.py", test_cmd="python test_mathx.py",
         return_mode="summary")
→ [summary] code_lines=3 first_line='def mul(a: float, b: float) -> float:'
  [qwen-pipeline] machine_verified=true check=py_compile attempts=1 tier=qwen
  output_id=… applied=true apply_to=mathx.py test=pass test_attempts=1

assign(task, done_when, repo, provider="", files="", max_iters=0, apply=True, model="") -> dict

A file-aware worker agent (Aider) that reads repo itself and grinds a whole task to an objective "done" with no orchestrator in the loop. The boss's role = define task + define done + commit.

When to use which: for a known target file, prefer delegate in token-cheap mode (context_files + apply_to + test_cmd) — it's simpler, faster, and needs no Aider install. Reach for assign when the task is genuinely exploratory or multi-file ("find where X is handled and fix it") — that's what the repo-map agent is for.

• Runs Aider (isolated venv, pinned) in a disposable git worktree off repo's HEAD — the real tree is untouched. Loops: worker edits → run done_when (a shell cmd that must exit 0) → on failure feed the verbatim output back to the worker (up to max_iters).
• On green: extracts a clean diff (build/worker junk filtered) and, if apply, mechanically applies it to the real tree. You then just commit.
• Returns a cheap summary: {done_passed, applied, iterations, files_changed, patch_path, done_log_tail, worker_log_tail, output_id} — the full diff is in patch_path.

log_correction(role, task, error_category, explanation, output_id="", correction_patch="", …) -> {"ok": true}

Appends one record to corrections/corrections.jsonl (and indexes it for retrieval). Call it after every delegation, even when the worker was correct (error_category="none", empty patch).

• error_category: logic | compile | style | edge_case | security | api_misuse | none.
• Prefer the diff-only form: pass output_id (from the delegate footer) + a unified-diff correction_patch instead of re-sending the code — the pipeline reconstructs both sides.

The delegate → review → fix → log loop

1. Split the task; delegate only the well-specified, self-contained part.
2. delegate(...) (snippet) or assign(...) (whole file-aware task) with the right role/provider — see config/routing.md.
3. Review for: correctness, compiles/runs, project conventions, edge cases, security, and any language-specific concerns (e.g. version-guarding for C++).
4. Fix if needed (else corrected == worker output).
5. log_correction(...) — always.

The mechanical gate + worker→worker auto-retry handle most fixes with zero orchestrator tokens; the boss only steps in for judgment. Full routing rules: config/routing.md.

Cost model — free-but-weaker vs. strong-but-metered

The routing philosophy the pipeline is built around:

What keeps this honest:

• The mechanical gate levels the field — a weak model whose output compiles and passes your tests is worth the same as a strong one, and it cost nothing. Failures bounce back to the worker, not to the orchestrator.
• Auto-escalation (cascade) retries a persistently failing task one tier up, carrying the failed attempt + checker error so the stronger model doesn't start cold.
• Pricing — set providers.<name>.cost (USD per Mtok, flat or per tier) and every cloud call is priced into metrics/metrics.jsonl; python src/metering.py shows per-tier and total estimated spend.
• Budgets are enforced — metering.budgets.<name>_tokens_per_day / <name>_usd_per_day: over-budget providers are refused with a clear error and the cascade won't escalate to them. A runaway retry loop cannot drain your credits.

In-context retrieval — learning without training

The pipeline gets better over time via retrieval, not weight training (an 80B can't be fine-tuned on one 32 GB GPU). The mechanism (src/retrieval.py):

• On log_correction: the task is embedded with nomic-embed-text and a compact entry (vector + role/provider/category + few-shot fields) is appended to corrections/index.jsonl.
• On delegate: the incoming task is embedded, the top-k most similar past corrections for the same provider+role are selected (favoring real mistakes per mistake_vs_correct_mix) and injected into the system prompt as few-shot examples.
• Fail-safe: if the embedder is unreachable, delegation still runs (just without examples) and corrections are still saved — re-embed later with python src/retrieval.py reindex.

Tunables in config/qwen.json → retrieval: enabled, top_k, role_filter, prefer_error_categories, mistake_vs_correct_mix.

Enabling Gemini (Vertex AI)

Secrets and machine-local values go in config/qwen.local.json (gitignored), which is deep-merged over config/qwen.json at load time — so the committed config never holds a secret.

1. .venv/Scripts/pip install -r requirements-gemini.txt
2. cp config/qwen.local.example.json config/qwen.local.json, then fill in your GCP project, the service-account JSON path (credentials_file), the model ids for flash/pro, and set enabled: true.
3. Delegate to a tier: delegate(..., provider="gemini", model="pro") or assign(..., provider="gemini", model="flash").

⚠️ The assign (Aider) model ids must use litellm's vertex_ai/ prefix for a service account (e.g. vertex_ai/gemini-2.5-pro), NOT gemini/ (the AI-Studio API-key path). Full walkthrough + the two-model-id-forms gotcha: docs/CONFIGURATION.md.

Use on another project (it's project-agnostic)

Nothing in the gate/agent layer is tied to a particular codebase. To use it on another repo:

1. Point delegate(repo="/path/to/your-repo", …) / assign(repo=…, done_when=…) at it.
2. (Optional) drop <your-repo>/.qwen-pipeline.json for per-project settings:
   hljs language-json

   Copy

   {
     "conventions": "TypeScript strict; no `any`. Zod for validation. snake_case file names.",
     "agent": { "max_iters": 4, "diff_excludes": [".aider*", "node_modules", "dist", "*.pyc"] }
   }
   

   conventions is injected into the worker prompt on every delegate(repo=…) — your style rules are enforced up front instead of corrected after the fact. The agent block merges over config/qwen.json → agent (repo wins) for assign.
3. Gate languages (gate.languages.*) and any batched build step are config-driven — enable/point them per project. The MCP surface (delegate, assign, log_correction) is unchanged.

Repository layout

qwen-pipeline/
├── README.md
├── CHANGELOG.md
├── CONTRIBUTING.md
├── LICENSE                       # MIT
├── requirements.txt              # core, PINNED (mcp, numpy)
├── requirements-gemini.txt       # optional: Gemini/Vertex provider
├── requirements-aider.txt        # optional: the `assign` agent (install in .aider-venv)
├── config/
│   ├── qwen.json                 # canonical config (committed, NO secrets)
│   ├── qwen.local.example.json   # template for the gitignored local overlay
│   ├── qwen.local.json           # GITIGNORED: project id, creds path, model ids, enabled flags
│   └── routing.md                # what to delegate, to which provider/role/tier
├── docs/
│   ├── CONFIGURATION.md          # config reference + enabling Gemini
│   └── MULTI_AGENT.md            # how the boss + two-worker model works (beginner-friendly)
├── src/
│   ├── cli.py                    # the `apprentice` command (init/chat/run/serve/doctor/…)
│   ├── server.py                 # FastMCP stdio server: delegate / assign / log_correction
│   ├── providers.py              # provider registry: ollama-local / openai-compatible / vertex-ai
│   ├── chat_providers.py         # multi-turn chat + tool calls, normalized across providers
│   ├── loop.py / session.py      # the agent loop; history, repo map, compaction, transcripts
│   ├── tools.py / verify.py      # agent tools (repo-scoped) ; snapshot → check → revert
│   ├── chat_ui.py                # the REPL for `apprentice chat` / `apprentice run`
│   ├── deliver.py                # server-side context fetch + apply/test/revert (token-cheap mode)
│   ├── budgets.py / corrections.py  # shared daily caps; the corrections writer
│   ├── agent.py                  # the `assign` file-aware agent (Aider + disposable worktree)
│   ├── gate.py / gate_cli.py     # mechanical gate (compile/lint) + worker-retry
│   ├── store.py                  # output-id store + unified-diff apply
│   ├── retrieval.py              # embed + cosine retrieval of past corrections
│   ├── metering.py               # per-delegation cost/outcome log
│   ├── host_verify.py            # optional batched build/test runner (project-specific)
│   └── roles.py                  # role -> system-prompt map
├── tests/test_pipeline.py        # deterministic, offline (stubs providers/embeddings)
└── corrections/                  # GITIGNORED contents: corrections + retrieval index (local only)

Gitignored (never pushed): config/qwen.local.json, secrets/, corrections/*.jsonl, outputs/, metrics/, models/, .venv/, .aider-venv/, node_modules/.

Troubleshooting

Conventions & safety

• Pin dependencies. Never float the MCP SDK (a 2026 stdio command-injection advisory makes pinning the documented mitigation). The stdio server runs with full user privileges — keep scope tight.
• Keep the tool surface small (three tools) — schemas load into the orchestrator's context every turn.
• corrections/ may contain private code — it stays local (gitignored) and is not committed.
• Token generation on the worker; judgment on the boss. That's the only place the cost win comes from — if a task type keeps coming back wrong, stop delegating it (config/routing.md).

Documentation

• docs/MULTI_AGENT.md — how the boss + two-worker model works, in beginner terms (what an agent is; Claude Code vs. Aider vs. Codex vs. OpenClaw; who does what).
• docs/CONFIGURATION.md — full config reference, the committed vs. local overlay, and enabling the Gemini/Vertex worker.
• CONTRIBUTING.md — dev setup, tests, and ground rules.
• CHANGELOG.md — notable changes.

License

MIT © 2026 Mohsen Mirzaei.