feat: Local-First TensorZero Gateway & AgentGym Integration

[POWERFULMOVES]

TensorZero: Local-First Architecture & Network Hardening

Summary

This PR transitions TensorZero to a production-ready "Local-First" architecture, implementing the features/gateway with strict 5-tier network segmentation, ClickHouse observability, and a routing hierarchy that prioritizes local/free models. It also introduces the AgentGym-RL Coordinator for the simulation-to-reality training loop.

Key Changes

1. Local-First Routing Hierarchy

• Tier 1 (Local): Ollama (qwen2.5:32b, qwen3-embedding) - Always try first.
• Tier 1 (Free): Cloudflare Workers AI (llama-3-8b).
• Tier 2 (Plans): OpenRouter, GLM, Claude Code.
• Tier 3 (Fallback): Google Gemini -> Anthropic -> OpenAI (Emergency).

2. Security Hardening

• Network Segmentation: Services deployed to isolated api_tier, data_tier, app_tier, bus_tier.
• Non-Root Execution: All new containers run as user 65532 (distroless) or 101 (service specific).
• Read-Only Filesystems: Root filesystems are read-only; write access limited to volumes (e.g., ClickHouse data).
• Capability Dropping: cap_drop: [ALL] applied universally.

3. New Services

• Gateway: features/gateway (TensorZero + ClickHouse).
• AgentGym Coordinator: pmoves/services/agentgym-rl-coordinator with NATS & Hugging Face integration (HF_TOKEN auth).
• Hyperdimensions UI: Pmoves-hyperdimensions served via unprivileged Nginx.

4. Integration

• Documentation: Complete rewrite of docs/PMOVES_TensorZero_Implementation.md with service integration matrix.

Verification

• Routing: Verify chat function falls back from Ollama -> Cloudflare -> Gemini.
• Security: Verify containers run as non-root and cannot write to root FS.
• Bus: Verify Coordinator connects to NATS geometry.event.v1.

Summary by CodeRabbit

• New Features

  • Introduced local-first routing with tiered model prioritization (local, free tier, and cloud fallback).
  • Added comprehensive provider configuration supporting multiple AI services.
  • Enabled training initiation and dataset publishing capabilities.

• Documentation

  • Expanded implementation guide with service integration matrix, security hardening guidelines, and deployment workflows.

_{✏️ Tip: You can customize this high-level summary in your review settings.}

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.

[coderabbitai]

Walkthrough

This PR introduces a comprehensive TensorZero gateway integration with tiered model routing, a new AgentGym RL Coordinator service for training/dataset management, and multi-service Docker deployment with ClickHouse observability. Documentation is restructured to reflect local-first architecture with cloud fallbacks and security hardening requirements.

Changes

Cohort / File(s)	Summary
TensorZero Gateway Configuration features/gateway/config/tensorzero.toml, features/gateway/docker-compose.yml	Introduces comprehensive gateway configuration with three-tier provider routing hierarchy (Local/Free: ollama, cloudflare; Specific Plans: openrouter, glm; Cloud Fallback: google_gemini, anthropic, openai). Defines model-to-provider mappings and weighted function routing variants. Docker setup includes ClickHouse observability service with security hardening (non-root 101:101, capability dropping, healthchecks) and gateway service with read-only filesystem, distroless image, and environment-based API key provisioning.
AgentGym RL Coordinator Service pmoves/services/agentgym-rl-coordinator/Dockerfile, pmoves/services/agentgym-rl-coordinator/app.py, pmoves/services/agentgym-rl-coordinator/docker-compose.yml, pmoves/services/agentgym-rl-coordinator/requirements.txt	Adds new FastAPI service with NATS client lifecycle management, subscribing to geometry.event.v1 events. Exposes /health endpoint (NATS/HuggingFace status), /agentgym/train/start route for training initiation, and /agentgym/dataset/publish route for HuggingFace dataset publishing with token validation. Multi-stage distroless Docker build optimizes image size. Requirements include fastapi, uvicorn, nats-py, huggingface_hub. Security hardening: non-root user 65532:65532, read-only filesystem, dropped capabilities.
Documentation docs/PMOVES_TensorZero_Implementation.md	Restructures from concise implementation reference to architecture-focused guide. Introduces local-first routing philosophy, reworks provider configuration (section 2) with tiered layout, replaces tool-centric definitions with model-centric blocks and weighted routing variants. Adds comprehensive "Exhaustive Service Integration Matrix" (section 3) detailing MIGRATE/INTEGRATE/NETWORK/OBSERVE strategies across 8 service categories. Removes prior best-practices sections, introduces Security Hardening Compliance and Hugging Face Training & Publishing Pipeline (section 6) with dataset creation, LoRA/finetune artifacts, and deployment readiness guidance.

Sequence Diagram(s)

sequenceDiagram
    actor User
    participant GW as TensorZero Gateway
    participant Ollama as Ollama (Local)
    participant CF as CloudFlare Workers
    participant Gemini as Google Gemini
    participant Claude as Anthropic Claude
    participant GPT as OpenAI GPT

    User->>GW: chat request (weight distribution)
    rect rgb(200, 220, 255)
        Note over GW,Ollama: Tier 1: Local/Free (weight: 1.0)
        GW->>Ollama: attempt local inference
        alt Success
            Ollama-->>GW: response
            GW-->>User: response (local)
        else Ollama unavailable
            Note over GW: fallback triggered
        end
    end
    
    rect rgb(200, 240, 200)
        Note over GW,CF: Tier 1b: Free Cloud (weight: 0.5)
        GW->>CF: attempt CloudFlare llama
        alt Success
            CF-->>GW: response
            GW-->>User: response (cloudflare)
        else CF unavailable
            Note over GW: fallback triggered
        end
    end
    
    rect rgb(255, 240, 200)
        Note over GW,Gemini: Tier 2: Free Plans (weight: 0.4)
        GW->>Gemini: attempt Gemini
        alt Success
            Gemini-->>GW: response
            GW-->>User: response (gemini)
        else Gemini unavailable
            Note over GW: fallback triggered
        end
    end
    
    rect rgb(255, 220, 200)
        Note over GW,Claude: Tier 3: Cloud Premium (weight: 0.3)
        GW->>Claude: attempt Claude
        alt Success
            Claude-->>GW: response
            GW-->>User: response (claude)
        else Claude unavailable
            Note over GW: final fallback
        end
    end
    
    GW->>GPT: attempt GPT-4o (weight: 0.1)
    GPT-->>GW: response
    GW-->>User: response (gpt-4o)

Loading

sequenceDiagram
    actor User
    participant API as AgentGym API
    participant NATS as NATS Bus
    participant HF as HuggingFace Hub
    participant Cache as HF Cache Volume

    User->>API: POST /agentgym/train/start (run_id)
    API->>NATS: subscribe geometry.event.v1
    Note over NATS: Await training events
    NATS-->>API: training events stream
    API->>API: log training initiation
    API-->>User: {status: training_started}
    
    User->>API: POST /agentgym/dataset/publish (dataset_name)
    rect rgb(200, 220, 255)
        Note over API: Validate HF_TOKEN
        alt HF_TOKEN present
            API->>Cache: check local cache
            Cache-->>API: cached artifacts (if exists)
            API->>HF: publish dataset/model
            HF-->>API: publication status
            API-->>User: {status: published}
        else HF_TOKEN missing
            API-->>User: HTTP 500 (token required)
        end
    end

Loading

Estimated code review effort

🎯 4 (Complex) | ⏱️ ~45 minutes

Possibly related PRs

• TensorZero: Local-First Architecture & Supabase Integration #336: Implements identical TensorZero local-first gateway integration with shared config/tensorzero.toml and service orchestration patterns.
• test(tensorzero): add agent function integration tests #288: Includes integration tests that exercise the gateway health checks, ClickHouse connectivity, and agent function endpoints introduced in this PR.
• feat: Build fixes, comprehensive testing, and TAC integration complete #286: Provides TensorZero model configuration fixes and NATS CLI behavior corrections that directly complement the gateway/model routing configuration added here.

Suggested reviewers

• hunnibear

Pre-merge checks and finishing touches

❌ Failed checks (1 warning, 1 inconclusive)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	You can run @coderabbitai generate docstrings to improve docstring coverage.
Description check	❓ Inconclusive	The PR description covers key changes (routing, security, new services) and verification steps, but is missing required Testing section and some Required Checks documentation.	Add a Testing section with specific commands/outputs and mark off or document the Required Checks (CHIT Contract Check, contract updates, documentation updates) to fully comply with the template.

✅ Passed checks (1 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title accurately summarizes the main change: introducing a local-first TensorZero Gateway with AgentGym integration, matching the core scope of the PR.

✨ Finishing touches

• 📝 Generate docstrings

🧪 Generate unit tests (beta)

• Create PR with unit tests
• Post copyable unit tests in a comment
• Commit unit tests in branch feature/tensorzero-integration

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

[coderabbitai]

Actionable comments posted: 8

🧹 Nitpick comments (4)

pmoves/services/agentgym-rl-coordinator/Dockerfile (1)

15-16: Optional: Optimize COPY to exclude build artifacts.

The COPY . /app statement copies all files including requirements.txt, Dockerfile, and potentially other build artifacts that aren't needed in production. This slightly increases image size.

🔎 Proposed optimization

-COPY . /app
+COPY app.py /app/
 WORKDIR /app

Alternatively, add a .dockerignore file to exclude unnecessary files:

requirements.txt
Dockerfile
docker-compose.yml
*.md
.git
__pycache__
*.pyc

features/gateway/docker-compose.yml (1)

21-24: Hardcoded credentials in ClickHouse configuration.

The credentials tensorzero:tensorzero are hardcoded in the environment variables and healthcheck. While acceptable for local development, these should be parameterized for production deployments.

🔎 Proposed improvement

     environment:
-      - CLICKHOUSE_USER=tensorzero
-      - CLICKHOUSE_PASSWORD=tensorzero
+      - CLICKHOUSE_USER=${CLICKHOUSE_USER:-tensorzero}
+      - CLICKHOUSE_PASSWORD=${CLICKHOUSE_PASSWORD:-tensorzero}
       - CLICKHOUSE_DEFAULT_ACCESS_MANAGEMENT=1

And update the healthcheck to use environment variable interpolation (note: may require docker-compose v2.x):

     healthcheck:
-      test: [ "CMD-SHELL", "wget --spider http://tensorzero:tensorzero@localhost:8123/ping" ]
+      test: [ "CMD-SHELL", "wget --spider http://$${CLICKHOUSE_USER}:$${CLICKHOUSE_PASSWORD}@localhost:8123/ping" ]

Document in README or deployment guide that production deployments must set unique credentials.

features/gateway/config/tensorzero.toml (1)

112-118: Consider adding cloud fallback variants for embedding function.

The embed function only has a local variant (qwen-embedding-local). If the local Ollama service is unavailable, embedding requests will fail. Consider adding cloud fallback options similar to the chat function.

🔎 Proposed enhancement

Add cloud embedding models:

+# Cloud embedding models
+[models.openai-embedding]
+provider = "openai"
+model = "text-embedding-3-small"
+
 # Embedding: Local First
 [functions.embed]
 type = "embedding"
-variants = ["local_embed"]
+variants = ["local_embed", "cloud_embed"]
 
 [functions.embed.variants.local_embed]
 model = "models.qwen-embedding-local"
+weight = 1.0
+
+[functions.embed.variants.cloud_embed]
+model = "models.openai-embedding"
+weight = 0.3

This ensures embedding functionality remains available even when Ollama is down, aligning with the "Local-First" philosophy's resilience goals.

docs/PMOVES_TensorZero_Implementation.md (1)

169-245: Add blank lines around all tables for proper markdown formatting.

Markdown tables should be surrounded by blank lines. This is flagged by the linter at multiple locations (lines 169, 177, 185, 195, 207, 219, 230, 241) and can cause rendering issues in various markdown processors.

🔎 Example fix

 ### 3.1 Core Infrastructure
+
 | Service | Current Status | Integration Strategy | Action Item |
 | :--- | :--- | :--- | :--- |
 | `tensorzero-gateway` | **NEW** | **CORE**: The central hub. | Deploy & Secure |
 ...
 | `nats` | **EXISTING** | **CORE**: The Geometry Bus carrier. | Ensure `bus_tier` isolation |
+
 ### 3.2 Agent Orchestration

Apply to all tables in section 3.

📜 Review details

Configuration used: defaults

Review profile: CHILL

Plan: Pro

📥 Commits

Reviewing files that changed from the base of the PR and between 37048cf and aea2ba6.

📒 Files selected for processing (7)

• docs/PMOVES_TensorZero_Implementation.md
• features/gateway/config/tensorzero.toml
• features/gateway/docker-compose.yml
• pmoves/services/agentgym-rl-coordinator/Dockerfile
• pmoves/services/agentgym-rl-coordinator/app.py
• pmoves/services/agentgym-rl-coordinator/docker-compose.yml
• pmoves/services/agentgym-rl-coordinator/requirements.txt

🧰 Additional context used

📓 Path-based instructions (4)

pmoves/**/docker-compose.yml

📄 CodeRabbit inference engine (pmoves/AGENTS.md)

Use Compose profiles (data, workers) to scope what runs locally in docker-compose.yml

Files:

• pmoves/services/agentgym-rl-coordinator/docker-compose.yml

pmoves/services/**/*.py

📄 CodeRabbit inference engine (pmoves/AGENTS.md)

pmoves/services/**/*.py: Keep modules small and single-purpose; share helpers in services/common/
FastAPI routes: snake_case function names; path names kebab-case only in URLs
Validate payloads against schemas before publishing events using services/common/events.py

Files:

• pmoves/services/agentgym-rl-coordinator/app.py

pmoves/**/*.py

📄 CodeRabbit inference engine (pmoves/AGENTS.md)

Python 3.11+, 4-space indentation, prefer type hints

Files:

• pmoves/services/agentgym-rl-coordinator/app.py

**/*.py

📄 CodeRabbit inference engine (AGENTS.md)

Use Python 3.11+ with 4-space indentation and prefer type hints in all Python code

Files:

• pmoves/services/agentgym-rl-coordinator/app.py

🧠 Learnings (10)

📚 Learning: 2025-12-07T11:03:53.415Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: pmoves/AGENTS.md:0-0
Timestamp: 2025-12-07T11:03:53.415Z
Learning: Applies to pmoves/**/environment.yml : Preferred Python: Conda 3.11+ (env name: `PMOVES.AI` or `pmoves-ai`); use `environment.yml` at repo root for setup

Applied to files:

• pmoves/services/agentgym-rl-coordinator/requirements.txt

📚 Learning: 2025-12-15T12:02:18.878Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: pmoves/docs/ARTSTUFF/realtime/GEMINI.md:0-0
Timestamp: 2025-12-15T12:02:18.878Z
Learning: Applies to pmoves/docs/ARTSTUFF/realtime/**/*{install,start,update,reset}.{js,json} : Install Python packages using uv instead of pip when possible, for example uv pip install -r requirements.txt instead of pip install

Applied to files:

• pmoves/services/agentgym-rl-coordinator/requirements.txt

📚 Learning: 2025-12-07T11:03:53.415Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: pmoves/AGENTS.md:0-0
Timestamp: 2025-12-07T11:03:53.415Z
Learning: Applies to pmoves/services/{agent-zero,archon}/**/*.py : For Agents/Archon full-stack validation, follow the 'All Services Up, Then Tests' section in `pmoves/docs/SMOKETESTS.md` and use `make -C pmoves agents-headless-smoke`, `make -C pmoves smoke-gpu`, and `make -C pmoves verify-all`

Applied to files:

• pmoves/services/agentgym-rl-coordinator/requirements.txt
• pmoves/services/agentgym-rl-coordinator/app.py
• docs/PMOVES_TensorZero_Implementation.md

📚 Learning: 2025-12-07T11:03:07.638Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: GEMINI.md:0-0
Timestamp: 2025-12-07T11:03:07.638Z
Learning: Applies to **/pmoves/**/*{qwen,gemma,audio,summary}*.py : Integrate Qwen2-Audio provider and add Gemma summaries to PMOVES.YT endpoints

Applied to files:

• pmoves/services/agentgym-rl-coordinator/requirements.txt
• docs/PMOVES_TensorZero_Implementation.md

📚 Learning: 2025-12-07T11:03:53.415Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: pmoves/AGENTS.md:0-0
Timestamp: 2025-12-07T11:03:53.415Z
Learning: Applies to pmoves/**/docker-compose.yml : Use Compose profiles (`data`, `workers`) to scope what runs locally in docker-compose.yml

Applied to files:

• pmoves/services/agentgym-rl-coordinator/docker-compose.yml
• features/gateway/docker-compose.yml

📚 Learning: 2025-12-07T11:03:53.415Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: pmoves/AGENTS.md:0-0
Timestamp: 2025-12-07T11:03:53.415Z
Learning: Applies to pmoves/services/**/*.py : Keep modules small and single-purpose; share helpers in `services/common/`

Applied to files:

• pmoves/services/agentgym-rl-coordinator/app.py

📚 Learning: 2025-12-15T12:03:17.291Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: AGENTS.md:0-0
Timestamp: 2025-12-15T12:03:17.291Z
Learning: Mandatory context before changes: read pmoves/docs/PMOVES.AI PLANS/ROADMAP.md and pmoves/docs/NEXT_STEPS.md to align with current sprint focus

Applied to files:

• docs/PMOVES_TensorZero_Implementation.md

📚 Learning: 2025-12-07T11:03:53.415Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: pmoves/AGENTS.md:0-0
Timestamp: 2025-12-07T11:03:53.415Z
Learning: Read `pmoves/docs/PMOVES.AI PLANS/ROADMAP.md` and `pmoves/docs/NEXT_STEPS.md` before making changes to align with current sprint focus

Applied to files:

• docs/PMOVES_TensorZero_Implementation.md

📚 Learning: 2025-12-15T12:03:17.291Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: AGENTS.md:0-0
Timestamp: 2025-12-15T12:03:17.291Z
Learning: Applies to services/**/README.md : Update services/*/README.md and pmoves/docs/PMOVES.AI PLANS/ runbooks when touching service operational code

Applied to files:

• docs/PMOVES_TensorZero_Implementation.md

📚 Learning: 2025-12-15T12:03:17.291Z

Learnt from: CR
Repo: POWERFULMOVES/PMOVES.AI PR: 0
File: AGENTS.md:0-0
Timestamp: 2025-12-15T12:03:17.291Z
Learning: Agents/Archon: for full-stack validation, follow the 'All Services Up, Then Tests' section in pmoves/docs/SMOKETESTS.md and the Archon service guide

Applied to files:

• docs/PMOVES_TensorZero_Implementation.md

🪛 Checkov (3.2.334)

features/gateway/docker-compose.yml

[medium] 41-42: Basic Auth Credentials

(CKV_SECRET_4)

🪛 markdownlint-cli2 (0.18.1)

docs/PMOVES_TensorZero_Implementation.md

169-169: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

177-177: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

179-179: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

180-180: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

181-181: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

182-182: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

185-185: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

187-187: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

188-188: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

189-189: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

190-190: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

191-191: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

192-192: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

195-195: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

197-197: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

198-198: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

199-199: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

200-200: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

201-201: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

202-202: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

203-203: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

204-204: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

207-207: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

209-209: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

210-210: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

211-211: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

212-212: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

213-213: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

214-214: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

215-215: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

216-216: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

219-219: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

221-221: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

222-222: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

223-223: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

224-224: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

225-225: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

226-226: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

227-227: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

230-230: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

232-232: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

233-233: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

234-234: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

235-235: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

236-236: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

237-237: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

238-238: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

241-241: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

243-243: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

244-244: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

245-245: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

🪛 OSV Scanner (2.3.1)

pmoves/services/agentgym-rl-coordinator/requirements.txt

[HIGH] 1-1: fastapi 0.109.0: undefined

(PYSEC-2024-38)

---

[HIGH] 1-1: python-multipart 0.0.6: python-multipart vulnerable to Content-Type Header ReDoS

(GHSA-2jv5-9r88-3w3p)

---

[HIGH] 1-1: python-multipart 0.0.6: Denial of service (DoS) via deformation multipart/form-data boundary

(GHSA-59g5-xgcq-4qw3)

---

[HIGH] 1-1: starlette 0.35.1: Starlette has possible denial-of-service vector when parsing large files in multipart forms

(GHSA-2c2j-9gv5-cj73)

---

[HIGH] 1-1: starlette 0.35.1: Starlette Denial of service (DoS) via multipart/form-data

(GHSA-f96h-pmfr-66vw)

🪛 Ruff (0.14.10)

pmoves/services/agentgym-rl-coordinator/app.py

21-21: Unused function argument: app

(ARG001)

---

38-38: Do not catch blind exception: Exception

(BLE001)

---

39-39: Use logging.exception instead of logging.error

Replace with exception

(TRY400)

⏰ Context from checks skipped due to timeout of 90000ms. You can increase the timeout in your CodeRabbit configuration to a maximum of 15 minutes (900000ms). (2)

• GitHub Check: tests (3.11)
• GitHub Check: Analyze (python)

🔇 Additional comments (6)

pmoves/services/agentgym-rl-coordinator/app.py (1)

49-53: LGTM!

The health check endpoint provides clear visibility into NATS connectivity and HuggingFace configuration status.

pmoves/services/agentgym-rl-coordinator/Dockerfile (1)

1-25: LGTM! Excellent security posture.

The multi-stage build using distroless with a non-root user aligns perfectly with the PR's security hardening objectives. The PATH configuration and HF_HOME setup are appropriate for the service's HuggingFace integration needs.

pmoves/services/agentgym-rl-coordinator/docker-compose.yml (1)

39-43: LGTM! Security hardening properly implemented.

The security configuration follows the PR's hardening requirements:

• Non-root distroless user (65532:65532)
• Read-only filesystem with volume-specific write access
• All capabilities dropped

This aligns with the documented Production Hardening Guidelines.

features/gateway/docker-compose.yml (1)

32-44: LGTM! Strong security hardening implementation.

Both services properly implement:

• Non-root users (101 for ClickHouse, 65532 for Gateway)
• Capability dropping with minimal re-additions for ClickHouse
• Read-only filesystem for gateway
• Health-based startup ordering
• Network segmentation

The ClickHouse capability additions (SYS_NICE, NET_ADMIN, IPC_LOCK) are justified for database performance and networking.

Also applies to: 72-78

features/gateway/config/tensorzero.toml (1)

83-110: LGTM! Well-structured tiered routing hierarchy.

The function routing configuration clearly implements the Local-First philosophy with appropriate weight distribution:

• Local (1.0) → Free (0.5) → Gemini (0.4) → Anthropic (0.3) → OpenAI (0.1)

The decreasing weights properly prioritize cost-effective options while maintaining fallback resilience.

docs/PMOVES_TensorZero_Implementation.md (1)

1-304: Excellent comprehensive documentation.

This documentation provides:

• Clear architectural overview of the Local-First routing philosophy
• Detailed configuration guidance with practical examples
• Exhaustive service integration matrix covering 55+ services
• Security hardening requirements aligned with the PR objectives
• HuggingFace training pipeline integration details

The structured approach will significantly help with the migration and deployment process.

[coderabbitai]

🛠️ Refactor suggestion | 🟠 Major

Fix table column mismatch in Agent Orchestration section.

The table header defines 3 columns, but the data rows contain 4 columns (the separator row also has only 3 columns). This causes rendering issues and is flagged by the markdown linter.

🔎 Proposed fix

Looking at the table structure, it appears the first column in data rows should be part of the header. The table should have either:

Option 1: Keep 3 columns by removing the first data cell:

 ### 3.2 Agent Orchestration
 | Service | Integration Strategy | Action Item |
 | :--- | :--- | :--- |
-| `agent-zero` | **MIGRATE** | Point `OPENAI_BASE_URL` to Gateway. Use `tensorzero::model` names. | Update `.env` |
+| `agent-zero` | **MIGRATE**: Point `OPENAI_BASE_URL` to Gateway. Use `tensorzero::model` names. | Update `.env` |

Option 2: Add a 4th column to the header (recommended based on content):

 ### 3.2 Agent Orchestration
-| Service | Integration Strategy | Action Item |
-| :--- | :--- | :--- |
+| Service | Integration Strategy | Details | Action Item |
+| :--- | :--- | :--- | :--- |

Apply the same fix to all tables in sections 3.2 through 3.8 (lines 177-245).

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	| Service | Integration Strategy | Action Item |
	| :--- | :--- | :--- |
	| `agent-zero` | **MIGRATE** | Point `OPENAI_BASE_URL` to Gateway. Use `tensorzero::model` names. | Update `.env` |
	| `archon` | **MIGRATE** | Same as Agent Zero. Ensure all sub-agents route through Gateway. | Update `.env` |
	| `mesh-agent` | **NETWORK** | Ensure connectivity to Gateway for health checks. | Verify Network |
	| `channel-monitor` | **INTEGRATE** | Publish `content.new.v1` as Geometry Packet. | Add NATS logic |
	| Service | Integration Strategy | Details | Action Item |
	| :--- | :--- | :--- | :--- |
	| `agent-zero` | **MIGRATE** | Point `OPENAI_BASE_URL` to Gateway. Use `tensorzero::model` names. | Update `.env` |
	| `archon` | **MIGRATE** | Same as Agent Zero. Ensure all sub-agents route through Gateway. | Update `.env` |
	| `mesh-agent` | **NETWORK** | Ensure connectivity to Gateway for health checks. | Verify Network |
	| `channel-monitor` | **INTEGRATE** | Publish `content.new.v1` as Geometry Packet. | Add NATS logic |

🧰 Tools

🪛 markdownlint-cli2 (0.18.1)

177-177: Tables should be surrounded by blank lines

(MD058, blanks-around-tables)

---

179-179: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

180-180: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

181-181: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

---

182-182: Table column count
Expected: 3; Actual: 4; Too many cells, extra data will be missing

(MD056, table-column-count)

🤖 Prompt for AI Agents

In docs/PMOVES_TensorZero_Implementation.md around lines 177 to 182 (and apply
across sections 3.2–3.8, lines ~177–245): the markdown tables declare 3 header
columns but the data rows contain 4 cells, causing lint/render errors; fix by
adding a 4th header column and updating the separator row so every table has the
same number of pipes (or alternatively remove the extra data cell), then update
all corresponding tables in sections 3.2–3.8 so header, separator, and data rows
have matching column counts.

[coderabbitai]

⚠️ Potential issue | 🟡 Minor

🧩 Analysis chain

🏁 Script executed:

# First, let's explore the repository structure
git ls-files | head -30

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 1037

---

🏁 Script executed:

# Search for docker-compose files
fd docker-compose

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 2984

---

🏁 Script executed:

# Look for the tensorzero.toml file and its context
fd tensorzero.toml

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 254

---

🏁 Script executed:

# Search for any documentation about deployment or Linux compatibility
fd -e md -e rst | grep -E "(README|DEPLOY|INSTALL|GUIDE|doc)" | head -20

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 989

---

🏁 Script executed:

# Search for references to OLLAMA_URL or environment variables
rg "OLLAMA_URL|host.docker.internal" -i

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 33717

---

🏁 Script executed:

# Check the features/gateway directory structure and docker-compose
ls -la features/gateway/ 2>/dev/null || echo "Directory not found or doesn't exist"

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 301

---

🏁 Script executed:

# Read the features/gateway docker-compose file
cat -n features/gateway/docker-compose.yml 2>/dev/null || echo "File not found"

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 2610

---

🏁 Script executed:

# Read the tensorzero.toml file in full context
cat -n features/gateway/config/tensorzero.toml

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 3431

---

🏁 Script executed:

# Check if there are any environment variable substitutions in tensorzero.toml
rg -A 5 -B 5 "tensorzero.toml" features/gateway/ || echo "No additional context found"

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 842

---

🏁 Script executed:

# Check what documentation exists specifically for features/gateway
find features/gateway -type f \( -name "*.md" -o -name "README*" -o -name "*.txt" \) 2>/dev/null

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 49

---

Add Linux compatibility configuration to the TensorZero gateway Docker Compose setup.

The hardcoded URL http://host.docker.internal:11434 is unavailable on Linux Docker hosts without explicit configuration. Add to features/gateway/docker-compose.yml under the tensorzero-gateway service:

extra_hosts:
  - "host.docker.internal:host-gateway"

Alternatively, modify features/gateway/config/tensorzero.toml line 10 to support environment variable override:

url = "${env.OLLAMA_URL:-http://host.docker.internal:11434}"

and add OLLAMA_URL to the environment section in docker-compose.yml.

Document this Linux requirement in a features/gateway/README.md or update deployment docs to note that Linux deployments must either use the extra_hosts approach or configure a network override.

🤖 Prompt for AI Agents

In features/gateway/config/tensorzero.toml around lines 7 to 10, the hardcoded
Ollama URL uses host.docker.internal which is not available on Linux Docker
hosts; update the configuration to allow an environment override for the URL
(read OLLAMA_URL with a fallback to the existing value) and then update
features/gateway/docker-compose.yml to either add extra_hosts:
["host.docker.internal:host-gateway"] under the tensorzero-gateway service or
add OLLAMA_URL to the service environment so operators can set a
Linux-compatible address; also add a short note to features/gateway/README.md
(or your deployment docs) explaining the Linux requirement and the two supported
workarounds.

[coderabbitai]

🛠️ Refactor suggestion | 🟠 Major

Add Compose profile for gateway services.

Per coding guidelines: "Use Compose profiles (data, workers) to scope what runs locally in docker-compose.yml"

The gateway services should have a profile (e.g., gateway or data) to control when they're started.

🔎 Proposed change

 services:
   tensorzero-clickhouse:
     image: clickhouse/clickhouse-server:24.12-alpine
     restart: unless-stopped
+    profiles:
+      - data
     environment:

And similarly for tensorzero-gateway:

   tensorzero-gateway:
     image: tensorzero/gateway:latest
     restart: unless-stopped
+    profiles:
+      - gateway
     command: [ "--config-file", "/app/config/tensorzero.toml" ]

Based on coding guidelines: "Use Compose profiles (data, workers) to scope what runs locally in docker-compose.yml"

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	services:
	tensorzero-clickhouse:
	image: clickhouse/clickhouse-server:24.12-alpine
	restart: unless-stopped
	environment:
	- CLICKHOUSE_USER=tensorzero
	- CLICKHOUSE_PASSWORD=tensorzero
	- CLICKHOUSE_DEFAULT_ACCESS_MANAGEMENT=1
	ports:
	- "8123:8123"
	volumes:
	- ./clickhouse-data:/var/lib/clickhouse
	networks:
	- data_tier
	- monitoring_tier
	# Security Hardening
	user: "101:101" # ClickHouse default non-root user
	cap_drop:
	- ALL
	cap_add:
	- SYS_NICE
	- NET_ADMIN
	- IPC_LOCK
	healthcheck:
	test: [ "CMD-SHELL", "wget --spider http://tensorzero:tensorzero@localhost:8123/ping" ]
	interval: 5s
	timeout: 2s
	retries: 5
	services:
	tensorzero-clickhouse:
	image: clickhouse/clickhouse-server:24.12-alpine
	restart: unless-stopped
	profiles:
	- data
	environment:
	- CLICKHOUSE_USER=tensorzero
	- CLICKHOUSE_PASSWORD=tensorzero
	- CLICKHOUSE_DEFAULT_ACCESS_MANAGEMENT=1
	ports:
	- "8123:8123"
	volumes:
	- ./clickhouse-data:/var/lib/clickhouse
	networks:
	- data_tier
	- monitoring_tier
	# Security Hardening
	user: "101:101" # ClickHouse default non-root user
	cap_drop:
	- ALL
	cap_add:
	- SYS_NICE
	- NET_ADMIN
	- IPC_LOCK
	healthcheck:
	test: [ "CMD-SHELL", "wget --spider http://tensorzero:tensorzero@localhost:8123/ping" ]
	interval: 5s
	timeout: 2s
	retries: 5

🧰 Tools

🪛 Checkov (3.2.334)

[medium] 41-42: Basic Auth Credentials

(CKV_SECRET_4)

🤖 Prompt for AI Agents

In features/gateway/docker-compose.yml around lines 17-44, the
tensorzero-clickhouse service lacks a Compose profile so it always starts; add a
profiles key (e.g., profiles: ["data"] or ["gateway"]) to scope startup,
matching project guidelines, and do the same for the tensorzero-gateway service
elsewhere in this file so both services only run when the matching Compose
profile is enabled.

[coderabbitai]

⚠️ Potential issue | 🟠 Major

Critical: NATS connection failure doesn't prevent app startup.

If NATS connection fails during startup, the exception is logged but the app continues running. This means health checks will report "disconnected" status, and events won't be processed. Services depending on this coordinator will fail silently.

Additionally:

• No connection timeout specified, which could cause indefinite hangs.
• The message handler (lines 30-35) has no error handling; a malformed message could crash the subscription.
• Line 39: Use logging.exception instead of logging.error to capture the full traceback.

🔎 Proposed fixes

 @asynccontextmanager
 async def lifespan(app: FastAPI):
     # Startup
     global nc
     try:
         logger.info(f"Connecting to NATS at {NATS_URL}...")
-        nc = await nats.connect(NATS_URL)
+        nc = await nats.connect(NATS_URL, connect_timeout=10)
         logger.info("Connected to NATS.")
         
         # Subscribe to Geometry Bus events (Simulation Actions)
         async def message_handler(msg):
-            subject = msg.subject
-            data = msg.data.decode()
-            logger.info(f"Received geometry event on {subject}: {len(data)} bytes")
-            # TODO: Accumulate trajectory for dataset creation
+            try:
+                subject = msg.subject
+                data = msg.data.decode()
+                logger.info(f"Received geometry event on {subject}: {len(data)} bytes")
+                # TODO: Accumulate trajectory for dataset creation
+            except Exception as e:
+                logger.exception(f"Error processing message from {msg.subject}: {e}")
             
         await nc.subscribe("geometry.event.v1", cb=message_handler)
         
-    except Exception as e:
-        logger.error(f"Failed to connect to NATS: {e}")
+    except (ConnectionClosedError, TimeoutError, NoServersError) as e:
+        logger.exception(f"Failed to connect to NATS: {e}")
+        raise  # Prevent app from starting without NATS
 
     yield
     
     # Shutdown
     if nc:
         await nc.close()

🧰 Tools

🪛 Ruff (0.14.10)

21-21: Unused function argument: app

(ARG001)

---

38-38: Do not catch blind exception: Exception

(BLE001)

---

39-39: Use logging.exception instead of logging.error

Replace with exception

(TRY400)

🤖 Prompt for AI Agents

In pmoves/services/agentgym-rl-coordinator/app.py around lines 20 to 45, the
lifespan handler currently logs NATS connection failures but allows the app to
continue, has no connection timeout, and the message_handler lacks error
handling; change it so that the startup fails when NATS cannot be established
(raise the exception or call sys.exit after logging), use a connect call with a
timeout/connection options (or wrap await nats.connect(...) in asyncio.wait_for
with a reasonable timeout), replace logger.error with logger.exception to
capture tracebacks, and wrap the message_handler body in a try/except that logs
exceptions (using logger.exception) and safely skips malformed messages; also
ensure shutdown checks that nc exists and is connected before awaiting
nc.close().

[coderabbitai]

🛠️ Refactor suggestion | 🟠 Major

Add request body validation with Pydantic.

The run_id parameter should be part of a validated request body rather than a query parameter. This aligns with REST best practices and the coding guideline: "Validate payloads against schemas before publishing events."

🔎 Proposed refactor

Add a Pydantic model at the top of the file after imports:

+from pydantic import BaseModel, Field
+
+class TrainingStartRequest(BaseModel):
+    run_id: str = Field(..., min_length=1, description="Unique training run identifier")

Then update the endpoint:

 @app.post("/agentgym/train/start")
-async def start_training(run_id: str):
+async def start_training(request: TrainingStartRequest):
+    run_id = request.run_id
     logger.info(f"Triggering training run: {run_id}")
     # TODO: Implement PPO training loop triggering
     return {"status": "training_started", "run_id": run_id}

Based on coding guidelines: "Validate payloads against schemas before publishing events using services/common/events.py"

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	@app.post("/agentgym/train/start")
	async def start_training(run_id: str):
	logger.info(f"Triggering training run: {run_id}")
	# TODO: Implement PPO training loop triggering
	return {"status": "training_started", "run_id": run_id}
	@app.post("/agentgym/train/start")
	async def start_training(request: TrainingStartRequest):
	run_id = request.run_id
	logger.info(f"Triggering training run: {run_id}")
	# TODO: Implement PPO training loop triggering
	return {"status": "training_started", "run_id": run_id}

🤖 Prompt for AI Agents

In pmoves/services/agentgym-rl-coordinator/app.py around lines 55 to 59, the
endpoint currently takes run_id as a query parameter; change it to accept and
validate a request body by adding a Pydantic model (e.g., TrainStartRequest with
a run_id: str) after the imports, update the POST handler signature to accept
that model (async def start_training(req: TrainStartRequest)), use req.run_id in
the handler, and ensure any published events use the validated payload and the
existing services/common/events.py helpers.

[coderabbitai]

🛠️ Refactor suggestion | 🟠 Major

Improve validation and error handling.

Issues:

1. dataset_name should be part of a validated request body, not a query parameter.
2. HF_TOKEN validation should occur at startup (in the lifespan manager) rather than per-request to fail fast.
3. HTTP 500 is incorrect for missing configuration; use 503 (Service Unavailable) or validate at startup and refuse to start.

🔎 Proposed refactor

Add validation at startup in the lifespan manager:

 @asynccontextmanager
 async def lifespan(app: FastAPI):
     # Startup
     global nc
+    
+    # Validate HF_TOKEN at startup
+    if not HF_TOKEN:
+        logger.error("HF_TOKEN not configured. Dataset publishing will be unavailable.")
+        # Optionally: raise ValueError("HF_TOKEN required") to prevent startup
+    
     try:

Add a Pydantic model:

+class DatasetPublishRequest(BaseModel):
+    dataset_name: str = Field(..., min_length=1, pattern=r'^[\w\-]+$', description="Dataset name")

Update the endpoint:

 @app.post("/agentgym/dataset/publish")
-async def publish_dataset(dataset_name: str):
+async def publish_dataset(request: DatasetPublishRequest):
     if not HF_TOKEN:
-        raise HTTPException(status_code=500, detail="HF_TOKEN not configured")
+        raise HTTPException(status_code=503, detail="HF_TOKEN not configured - dataset publishing unavailable")
     
+    dataset_name = request.dataset_name
     logger.info(f"Publishing dataset {dataset_name} to Hugging Face...")
     # TODO: Implement dataset push logic via huggingface_hub
     return {"status": "publishing_started", "dataset": dataset_name}

Based on coding guidelines: "Validate payloads against schemas before publishing events using services/common/events.py"

Committable suggestion skipped: line range outside the PR's diff.

🤖 Prompt for AI Agents

pmoves/services/agentgym-rl-coordinator/app.py around lines 61-68: change the
endpoint to accept a Pydantic request body (e.g., DatasetPublish model) so
dataset_name is validated as part of the body rather than a query param; remove
the per-request HF_TOKEN check and instead validate HF_TOKEN in the
application's lifespan manager (fail fast on startup and log a clear error or
raise StartupError) so requests never proceed without config; return proper HTTP
status for service unavailability (503) only when appropriate or rely on startup
failure to prevent running; and validate the incoming payload against the
canonical schema in services/common/events.py before publishing, returning
validation errors (422) if it fails.

[coderabbitai]

🛠️ Refactor suggestion | 🟠 Major

Add Compose profile and clarify network requirements.

Per coding guidelines: "Use Compose profiles (data, workers) to scope what runs locally in docker-compose.yml"

Additionally, the service is attached to all four networks. Clarify if api_tier access is required—typically, internal coordinators only need app_tier, bus_tier, and monitoring_tier.

🔎 Proposed changes

 services:
   agentgym-coordinator:
     build: .
     restart: unless-stopped
+    profiles:
+      - workers  # or 'ai' if there's a profile for AI/ML services
     ports:
       - "8114:8114"
     environment:
       - NATS_URL=nats://nats:4222
       - PORT=8114
       - HF_TOKEN=${HF_TOKEN}
       - HF_HOME=/home/nonroot/.cache/huggingface
     volumes:
       - huggingface-cache:/home/nonroot/.cache/huggingface
     networks:
       - app_tier
       - bus_tier
-      - api_tier
       - monitoring_tier

If api_tier access is genuinely required (e.g., the service exposes a public API), document the reason in a comment.

Based on coding guidelines: "Use Compose profiles (data, workers) to scope what runs locally in docker-compose.yml"

📝 Committable suggestion

‼️ IMPORTANT
Carefully review the code before committing. Ensure that it accurately replaces the highlighted code, contains no missing lines, and has no issues with indentation. Thoroughly test & benchmark the code to ensure it meets the requirements.

Suggested change

	services:
	agentgym-coordinator:
	build: .
	restart: unless-stopped
	ports:
	- "8114:8114"
	environment:
	- NATS_URL=nats://nats:4222
	- PORT=8114
	- HF_TOKEN=${HF_TOKEN}
	- HF_HOME=/home/nonroot/.cache/huggingface
	volumes:
	- huggingface-cache:/home/nonroot/.cache/huggingface
	networks:
	- app_tier
	- bus_tier
	- api_tier
	- monitoring_tier
	# Security Hardening
	user: "65532:65532" # Non-root distroless
	read_only: true
	cap_drop:
	- ALL
	services:
	agentgym-coordinator:
	build: .
	restart: unless-stopped
	profiles:
	- workers # or 'ai' if there's a profile for AI/ML services
	ports:
	- "8114:8114"
	environment:
	- NATS_URL=nats://nats:4222
	- PORT=8114
	- HF_TOKEN=${HF_TOKEN}
	- HF_HOME=/home/nonroot/.cache/huggingface
	volumes:
	- huggingface-cache:/home/nonroot/.cache/huggingface
	networks:
	- app_tier
	- bus_tier
	- monitoring_tier
	# Security Hardening
	user: "65532:65532" # Non-root distroless
	read_only: true
	cap_drop:
	- ALL

🤖 Prompt for AI Agents

In pmoves/services/agentgym-rl-coordinator/docker-compose.yml around lines 21 to
43, add a Compose profile (e.g., "workers") to the agentgym-coordinator service
so it only runs under the intended local scope and adjust the networks to only
include the required internal networks (app_tier, bus_tier, monitoring_tier)
unless external API access is needed; if api_tier is genuinely required, add a
clear inline comment explaining why the service must be on api_tier (e.g.,
exposes public API) and otherwise remove api_tier from the networks list. Ensure
the service entry includes the profile name and update any documentation or
compose comments to reflect the chosen profile and network rationale.

[coderabbitai]

⚠️ Potential issue | 🔴 Critical

🧩 Analysis chain

🌐 Web query:

PYSEC-2024-38 fastapi vulnerability

💡 Result:

Summary — what it is

• PYSEC-2024-38 is a Regular Expression Denial-of-Service (ReDoS) affecting form-data parsing via python-multipart used by FastAPI. An attacker can craft a Content-Type option that causes the regex to consume CPU and stall the event loop, resulting in DoS. Published Feb 5, 2024. [1][2]

Identifiers

• PYSEC-2024-38 = CVE-2024-24762; GHSA-2jv5-9r88-3w3p / GHSA-qf9m-vfgh-m389. [1]

Affected software

• FastAPI versions before 0.109.1 (when using form-data parsing via python-multipart).
• python-multipart versions prior to 0.0.7. [1][3]

Impact

• High (CVSS 3.1 7.5): Denial of Service (process becomes unresponsive while parsing a single malicious request). [1][3]

Mitigation / Fix

• Upgrade FastAPI to >= 0.109.1.
• Upgrade python-multipart to >= 0.0.7 (if your app directly depends on it).
• If you cannot upgrade immediately, block or validate suspicious Content-Type headers at your edge (WAF/proxy) to reject unusually long or weird option strings. [1][3]

References

• OSV advisory (PYSEC-2024-38 / FastAPI). [1]
• GitLab / CVE-2024-24762 (python-multipart / Starlette notes). [2]
• PT Security / dbugs summary of CVE-2024-24762 and remediation. [3]

---

🌐 Web query:

GHSA-2jv5-9r88-3w3p python-multipart Content-Type Header ReDoS

💡 Result:

Summary — CVE-2024-24762 / GHSA-2jv5-9r88-3w3p

• What: A ReDoS (regular-expression DoS) in python-multipart when parsing the Content-Type header options for form data. A maliciously crafted Content-Type option can cause catastrophic regex backtracking, consuming CPU and stalling the event loop. [1][2]
• Affected versions: python-multipart <= 0.0.6 (all versions before 0.0.7). [1][3]
• Severity: CVSS 7.5 (High). [1][3]
• Fixed: Upgrade to python-multipart >= 0.0.7. Commit/patch referenced in the advisory. [1][2]

Actionable mitigation

1. Upgrade immediately: pip install --upgrade python-multipart (ensure version >= 0.0.7). [1][2]
2. If you cannot upgrade immediately, apply network mitigations (rate limiting, request throttling, WAF rules to block suspicious Content-Type headers) and avoid parsing untrusted form Content-Type options until patched. [3][4]

References

1. GitLab Advisory / package advisory summary (python-multipart — CVE-2024-24762). [turn0search0]
2. Chainguard / advisory summary (CVE-2024-24762). [turn0search2]
3. Snyk / Debian security tracker summary for CVE-2024-24762. [turn0search4]
4. Additional vendor write-ups / vulnerability databases (Vulert / Seal) summarizing exploit and mitigation. [turn0search5][turn0search1]

---

🌐 Web query:

GHSA-59g5-xgcq-4qw3 python-multipart multipart form-data DoS

💡 Result:

Summary

• GHSA-59g5-xgcq-4qw3 (CVE-2024-53981) affects python-multipart versions < 0.0.18: a malicious multipart/form-data request that places lots of bytes before the first boundary or after the last boundary triggers one-byte-at-a-time skips and excessive logging, allowing high CPU use and an ASGI event-loop stall (DoS). Patch: upgrade to 0.0.18+. [1][2]

Mitigation

• Upgrade python-multipart to >= 0.0.18 and deploy/restart services.
• If you vendor/pin dependencies, update lockfiles (requirements.txt/Poetry/Pipfile) and rebuild images.
• Consider WAF/rate-limiting for multipart endpoints until patched.

Related note

• python-multipart also had an earlier ReDoS issue (CVE-2024-24762) fixed in 0.0.7; ensure you’re on recent non-vulnerable versions for both issues. [3]

Sources
[1] GitHub Advisory GHSA-59g5-xgcq-4qw3 (python-multipart).
[2] CVE-2024-53981 / CVE feed entry.
[3] Snyk / GitLab advisories (python-multipart vulnerability summaries).

---

🌐 Web query:

GHSA-2c2j-9gv5-cj73 starlette DoS vulnerability

💡 Result:

Summary: GHSA-2c2j-9gv5-cj73 is a moderate-severity denial‑of‑service in Starlette (CVE‑2025‑54121). When parsing multipart/form-data with large file parts (larger than the default in‑memory spool size) Starlette can block the event/main thread while rolling files to disk, preventing the app from accepting new connections. Affects Starlette versions < 0.47.2. Fixed in 0.47.2 — upgrade or pin to ≥ 0.47.2. [1][2][3]

References

• GitHub security advisory (GHSA-2c2j-9gv5-cj73). [1]
• OSV / vulnerability record (CVE‑2025‑54121). [2]
• Safety / vulnerability summary. [3]

---

🌐 Web query:

GHSA-f96h-pmfr-66vw starlette DoS vulnerability

💡 Result:

Yes — GHSA-f96h-pmfr-66vw is a published high-severity DoS in Starlette (CVE-2024-47874). Summary, impact, and mitigation:

• Issue: multipart/form-data parts that lack a filename are treated as text fields and buffered without size limits, allowing an attacker to send very large form fields and exhaust server memory (DoS). PoC uses request.form() with a large field. [1][2]
• Affected versions: all starlette < 0.40.0. [1][3]
• Fixed in: starlette 0.40.0 — adds max_part_size to MultiPartParser to enforce limits. Upgrade to >= 0.40.0. [1][3]
• Mitigation alternatives: enforce upstream request-size limits (may not fully prevent this), use a reverse proxy with strict per-part limits, or explicitly set max_part_size via Starlette 0.40.0+ configuration. [1][2]

Sources:

1. GitHub Advisory GHSA-f96h-pmfr-66vw / CVE-2024-47874. [GitHub Advisory]. [1]
2. OSV entry for GHSA-f96h-pmfr-66vw. [2]
3. Safety / package advisories and changelog noting fix in 0.40.0. [3]

---

🏁 Script executed:

# Check what version of starlette is required by fastapi 0.109.0
curl -s "https://pypi.org/pypi/fastapi/0.109.0/json" | jq -r '.info.requires_dist[] | select(contains("starlette"))'

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 92

---

🏁 Script executed:

# Check the current versions of these packages on PyPI
for pkg in fastapi python-multipart starlette; do
  echo "=== $pkg latest version ==="
  curl -s "https://pypi.org/pypi/$pkg/json" | jq -r '.info | "\(.version) (released: \(.last_modified))"'
done

Repository: POWERFULMOVES/PMOVES.AI

Length of output: 243

---

Critical: Multiple HIGH severity security vulnerabilities in dependencies.

python-multipart 0.0.6 is affected by two HIGH severity DoS vulnerabilities:

• GHSA-2jv5-9r88-3w3p (CVE-2024-24762): Content-Type header ReDoS—fixed in 0.0.7
• GHSA-59g5-xgcq-4qw3 (CVE-2024-53981): Multipart boundary parsing DoS—fixed in 0.0.18

fastapi 0.109.0 itself is patched in 0.109.1 for PYSEC-2024-38 (the ReDoS issue above).

starlette 0.35.x (transitive via fastapi) is vulnerable to:

• GHSA-f96h-pmfr-66vw (CVE-2024-47874): HIGH severity—unbounded buffering of form fields leads to memory exhaustion; fixed in 0.40.0
• GHSA-2c2j-9gv5-cj73 (CVE-2025-54121): Moderate severity—file spooling DoS; fixed in 0.47.2

These vulnerabilities directly impact production readiness and require immediate updates to: fastapi ≥ 0.109.1, python-multipart ≥ 0.0.18, and starlette ≥ 0.40.0 (or higher for full coverage).

🧰 Tools

🪛 OSV Scanner (2.3.1)

[HIGH] 1-1: fastapi 0.109.0: undefined

(PYSEC-2024-38)

---

[HIGH] 1-1: python-multipart 0.0.6: python-multipart vulnerable to Content-Type Header ReDoS

(GHSA-2jv5-9r88-3w3p)

---

[HIGH] 1-1: python-multipart 0.0.6: Denial of service (DoS) via deformation multipart/form-data boundary

(GHSA-59g5-xgcq-4qw3)

---

[HIGH] 1-1: starlette 0.35.1: Starlette has possible denial-of-service vector when parsing large files in multipart forms

(GHSA-2c2j-9gv5-cj73)

---

[HIGH] 1-1: starlette 0.35.1: Starlette Denial of service (DoS) via multipart/form-data

(GHSA-f96h-pmfr-66vw)

🤖 Prompt for AI Agents

In pmoves/services/agentgym-rl-coordinator/requirements.txt lines 1-7, update
vulnerable dependency versions: bump fastapi to at least 0.109.1,
python-multipart to at least 0.0.18, and add an explicit starlette>=0.40.0 (or
higher) to ensure transitive fixes are applied; then run a fresh install and
regenerate the lock/constraints file (pip-compile or pip freeze) and run the
test suite/quick smoke tests to ensure compatibility with the new versions.