feat(agentgym): EvoSwarm → AgentGym-RL → HuggingFace publisher pipeline - #935
Merged
Conversation
|
You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard. |
Contributor
|
Warning Rate limit exceeded
⌛ How to resolve this issue?After the wait time has elapsed, a review can be triggered using the We recommend that you space out your commits to avoid hitting the rate limit. 🚦 How do rate limits work?CodeRabbit enforces hourly rate limits for each developer per organization. Our paid plans have higher rate limits than the trial, open-source and free plans. In all cases, we re-allow further reviews after a brief timeout. Please see our FAQ for further information. ℹ️ Review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Pro Run ID: 📒 Files selected for processing (4)
📝 WalkthroughWalkthroughAdds AgentGym RL training lifecycle NATS subjects and handlers, HF auto-publishing from the RL coordinator, a training-completed publisher in Evo-Controller, new /healthz and /metrics endpoints, and environment variables/default base model updates for HuggingFace integration. Changes
Sequence Diagram(s)sequenceDiagram
participant EC as Evo-Controller
participant NATS as NATS
participant RC as RL Coordinator
participant HF as HuggingFace
participant PL as Skills Pipeline
participant AZ as Agent Zero
EC->>NATS: publish agentgym.train.completed.v1 (training_run_id, trajectory_ids, model_id, fitness_metrics, timestamp)
NATS->>RC: deliver training completion event
RC->>RC: validate JSON payload, parse/validate UUID trajectory_ids, timestamp
RC-->>HF: request dataset/model publish (if HF configured)
HF-->>RC: respond with repo_url / dataset_id
RC->>NATS: publish agentgym.model.published.v1 (training_run_id, model_id, dataset_id, repo_url, trajectory_count, source, timestamp)
RC->>PL: emit skills.pipeline.model-benchmark-viz.v1
RC->>AZ: optionally notify Agent Zero / observability dashboards
RC->>Storage: record training_completed event (if storage configured)
Estimated code review effort🎯 4 (Complex) | ⏱️ ~45 minutes Possibly related PRs
Poem
🚥 Pre-merge checks | ✅ 2 | ❌ 1❌ Failed checks (1 inconclusive)
✅ Passed checks (2 passed)
✏️ Tip: You can configure your own custom pre-merge checks in the settings. ✨ Finishing Touches🧪 Generate unit tests (beta)
📝 Coding Plan
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. Comment |
![coderabbitai[bot]](https://avatars.githubusercontent.com/in/347564?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Actionable comments posted: 4
Caution
Some comments are outside the diff and can’t be posted inline due to platform limitations.
⚠️ Outside diff range comments (2)
pmoves/env.agentgym.example (1)
185-199:⚠️ Potential issue | 🟡 MinorMissing HuggingFace environment variables for the new publisher pipeline.
The PR adds HuggingFace auto-publishing functionality (in
app.pylines 29-30, it readsHF_TOKENandHF_ORG), but this example file is missing the required variables. Without these, users following this example won't be able to use the new HF publisher feature.📝 Proposed fix: Add HuggingFace configuration section
# A/B test traffic split (new_model:old_model) AGENTGYM_AB_TEST_SPLIT=0.1:0.9 +# ============================================================================ +# HuggingFace Publishing (New in PR `#935`) +# ============================================================================ + +# HuggingFace authentication token for auto-publishing datasets +# Required for agentgym.train.completed.v1 → HF publishing flow +HF_TOKEN= # Add your HuggingFace token here + +# HuggingFace organization for dataset publishing (default: pmoves) +HF_ORG=pmoves + # ============================================================================ # Storage Configuration # ============================================================================🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/env.agentgym.example` around lines 185 - 199, The example env file is missing the HuggingFace variables that app.py expects (HF_TOKEN and HF_ORG are read in app.py around the publisher logic), so add a new "HuggingFace Publishing" section to pmoves/env.agentgym.example declaring HF_TOKEN and HF_ORG (and optionally HF_REPO or HF_PRIVATE if your publisher supports them) with brief comments and default/example values; ensure variable names exactly match what app.py reads so the HF publisher code finds them at runtime.pmoves/services/agentgym-rl-coordinator/app.py (1)
253-265:⚠️ Potential issue | 🟡 MinorReplace
/healthendpoint with/healthzand add/metricsendpoint per coding guidelines.The service is missing the required
/healthzand/metricsendpoints. The current/healthendpoint does not match the standardized pattern. Referencepmoves/services/pdf-ingest/app.pyfor the correct implementation:📋 Example from pdf-ingest service
`@app.get`("/healthz") def healthz() -> Dict[str, bool]: return {"ok": True} `@app.get`("/metrics") def metrics(): return Response(generate_latest(), media_type=CONTENT_TYPE_LATEST)Implement
/healthzto return health status (can reuse the current health_check logic) and/metricsfor Prometheus-compatible metrics. As per coding guidelines: "Use FastAPI + uvicorn for API services and include thepmoves_healthrouter for/healthzand/metricsendpoints".🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/services/agentgym-rl-coordinator/app.py` around lines 253 - 265, Replace the existing /health endpoint (function health_check) with the standardized /healthz endpoint and add a /metrics endpoint: rename or move the logic from the current health_check to a new healthz handler that returns the same status fields (nats, supabase, huggingface) and register it as `@app.get`("/healthz"); add a /metrics handler that returns Prometheus metrics using Response(generate_latest(), media_type=CONTENT_TYPE_LATEST) (import generate_latest and CONTENT_TYPE_LATEST from prometheus_client); alternatively register the provided pmoves_health router which already exposes /healthz and /metrics instead of the old /health route to comply with coding guidelines.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@pmoves/env.agentgym.example`:
- Around line 14-15: Update the AGENTGYM_BASE_MODEL value to the correct
HuggingFace model identifier: replace the current placeholder string used for
the base model (AGENTGYM_BASE_MODEL) so it uses the organization prefix and
canonical name (Qwen/Qwen3-8B) instead of the incorrect "Qwen3-8B-Instruct";
locate the AGENTGYM_BASE_MODEL entry in the env.agentgym.example file and set
its value to the canonical HuggingFace ID.
In `@pmoves/services/agentgym-rl-coordinator/app.py`:
- Around line 151-158: Constructed dataset_name from training_run_id is not
validated before calling publish_to_huggingface, which can produce unclear HF
errors; before invoking hf_publisher.publish_to_huggingface, validate or
sanitize training_run_id to conform to DATASET_NAME_PATTERN (reuse the same
regex used by the /agentgym/dataset/publish endpoint), and if it fails either
(a) sanitize by replacing/stripping invalid chars to produce a valid
dataset_name or (b) log and skip publishing with a clear error; update the code
paths around training_run_id and dataset_name and ensure publish_to_huggingface
only receives a name matching DATASET_NAME_PATTERN.
In `@pmoves/services/evo-controller/agentgym_integration.py`:
- Line 385: The timestamp value is using local time while labeling it with "Z";
update the assignments that set "timestamp": time.strftime("%Y-%m-%dT%H:%M:%SZ")
to generate UTC time instead (either by passing time.gmtime() to time.strftime
or by using datetime.utcnow().strftime) so the produced ISO8601 string truly
reflects UTC; apply this change in the place where the "timestamp" key is set
and also in the publish_training_event function where the same pattern occurs.
- Around line 374-387: Refactor on_training_completed and publish_training_event
to use the standardized event publishing helpers from services.common.events
(import envelope, publish) instead of directly POSTing to agent-zero's
/events/publish; add new schema files for agentgym.train.completed.v1 and
agentgym.train.started.v1 under pmoves/contracts/schemas and register both
topics in pmoves/contracts/topics.json so envelope() can validate payloads, then
construct the event with envelope(topic, source="evo-controller", payload=...)
and call publish(envelope) to send to NATS (or the existing publish wrapper)
rather than manual HTTP calls.
---
Outside diff comments:
In `@pmoves/env.agentgym.example`:
- Around line 185-199: The example env file is missing the HuggingFace variables
that app.py expects (HF_TOKEN and HF_ORG are read in app.py around the publisher
logic), so add a new "HuggingFace Publishing" section to
pmoves/env.agentgym.example declaring HF_TOKEN and HF_ORG (and optionally
HF_REPO or HF_PRIVATE if your publisher supports them) with brief comments and
default/example values; ensure variable names exactly match what app.py reads so
the HF publisher code finds them at runtime.
In `@pmoves/services/agentgym-rl-coordinator/app.py`:
- Around line 253-265: Replace the existing /health endpoint (function
health_check) with the standardized /healthz endpoint and add a /metrics
endpoint: rename or move the logic from the current health_check to a new
healthz handler that returns the same status fields (nats, supabase,
huggingface) and register it as `@app.get`("/healthz"); add a /metrics handler
that returns Prometheus metrics using Response(generate_latest(),
media_type=CONTENT_TYPE_LATEST) (import generate_latest and CONTENT_TYPE_LATEST
from prometheus_client); alternatively register the provided pmoves_health
router which already exposes /healthz and /metrics instead of the old /health
route to comply with coding guidelines.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: ffcdbf99-d390-4fb7-a5c2-c9dd7ad63a98
📒 Files selected for processing (4)
.claude/context/nats-subjects.mdpmoves/env.agentgym.examplepmoves/services/agentgym-rl-coordinator/app.pypmoves/services/evo-controller/agentgym_integration.py
POWERFULMOVES
added a commit
that referenced
this pull request
Mar 15, 2026
- Rename /health → /healthz to follow PMOVES convention - Add /metrics endpoint with Prometheus-format gauges - Validate trajectory_ids as UUIDs before HF publish - Fix timestamp: use datetime.now(timezone.utc) instead of local time.strftime - Add HF_TOKEN and HF_ORG to env.agentgym.example Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Owner
Author
|
@coderabbitai generate unit tests |
Owner
Author
|
@coderabbitai generate docstrings |
Owner
Author
|
@coderabbitai generate coding plan |
Contributor
✅ Actions performedInitiated docstring generation; will generate only if new commits exist. |
Contributor
|
Note Docstrings generation - SUCCESS |
coderabbitai Bot
added a commit
that referenced
this pull request
Mar 15, 2026
Docstrings generation was requested by @POWERFULMOVES. * #935 (comment) The following files were modified: * `pmoves/services/agentgym-rl-coordinator/app.py` * `pmoves/services/evo-controller/agentgym_integration.py`
Contributor
|
Tip For best results, initiate chat on the files or code changes. 🧩 Analysis chain🏁 Script executed: Length of output: 57116
Coding Plan: EvoSwarm → AgentGym-RL → HuggingFace PipelineArchitecture OverviewsequenceDiagram
participant EvoSwarm as EvoSwarm Controller
participant AgentZero as Agent Zero (HTTP relay)
participant NATS
participant Coordinator as AgentGym-RL Coordinator
participant HF as HuggingFace Hub
participant BenchViz as Benchmark-Viz Pipeline
EvoSwarm->>AgentZero: POST /events/publish (agentgym.train.completed.v1)
AgentZero->>NATS: publish agentgym.train.completed.v1
NATS->>Coordinator: training_completed_handler
Coordinator->>HF: hf_publisher.publish_to_huggingface()
Coordinator->>NATS: agentgym.model.published.v1
Coordinator->>NATS: skills.pipeline.model-benchmark-viz.v1
NATS->>BenchViz: trigger benchmark visualization
Tasks
Key Issues Found During Analysis
🏁 Script executed: Length of output: 42 🏁 Script executed: Length of output: 268 🏁 Script executed: Length of output: 2346 🏁 Script executed: Length of output: 5811 🏁 Script executed: Length of output: 0 Good findings — let me read the coordinator package and existing test files to complete the plan. 🏁 Script executed: Length of output: 7487 🏁 Script executed: Length of output: 25549
Coding Plan: EvoSwarm → AgentGym-RL → HuggingFace PipelineSystem Flow (as-designed)flowchart TD
A[EvoSwarmController._tick] -->|evaluate_training_trigger| B{Should Train?}
B -->|No| A
B -->|Yes| C[launch_agentgym_training\n POST /agentgym/train/start]
C --> D[publish_training_event\n agentgym.train.started.v1]
C --> E[poll_training_completion\n GET /agentgym/train/status/run_id]
E -->|completed| F[on_training_completed\n POST agent-zero /events/publish]
F --> G[NATS: agentgym.train.completed.v1]
G --> H[training_completed_handler\n in app.py]
H --> I[HuggingFacePublisher\n .publish_to_huggingface]
I --> J[NATS: agentgym.model.published.v1]
I --> K[NATS: skills.pipeline.model-benchmark-viz.v1]
J --> L[Agent Zero / Discord Publisher]
K --> M[Benchmark Viz Pipeline]
Tasks (7 requirements)
Summary of Critical Issues Found
🏁 Script executed: Length of output: 42 🏁 Script executed: Length of output: 2450
The JetStream stream definition ( Final Plan SummaryAll 7 requirements are ready to implement, in priority order:
The test files should follow the Failed to handle agent chat message. Please try again. |
Contributor
|
Request timed out after 900000ms (requestId=e663f1c2-ca1f-422d-a91e-794e711dc7f6) |
Contributor
✅ Actions performedUnit Test Generation started... |
Contributor
There was a problem hiding this comment.
Actionable comments posted: 4
🧹 Nitpick comments (2)
pmoves/services/agentgym-rl-coordinator/app.py (2)
259-298: LGTM! Health and metrics endpoints properly implemented.The
/healthzendpoint matches the Docker healthcheck configuration, and/metricsprovides useful Prometheus gauges for NATS, Supabase, and HuggingFace status.Consider moving the
PlainTextResponseimport to the top of the file for consistency.♻️ Move import to top of file
from fastapi import FastAPI, HTTPException +from fastapi.responses import PlainTextResponse import nats # ... later in metrics() ... - from fastapi.responses import PlainTextResponse return PlainTextResponse("\n".join(lines) + "\n", media_type="text/plain")🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/services/agentgym-rl-coordinator/app.py` around lines 259 - 298, Move the inline import of PlainTextResponse out of the metrics() function and add it to the module-level imports at the top of the file; update the metrics() function to reference PlainTextResponse directly (function name: metrics) so all imports are centralized and consistent with other endpoints.
133-223: Consider adding idempotency guard for training completion handler.If NATS redelivers messages (e.g., with JetStream durable subscriptions), duplicate HuggingFace publishes could occur. The PR objectives mention this as a task. A simple in-memory set or Supabase-backed check could prevent duplicate processing.
♻️ Sketch of idempotency guard
# At module level _processed_training_runs: set[str] = set() async def training_completed_handler(msg): """Handle training completion: auto-publish to HuggingFace.""" try: data = json.loads(msg.data) training_run_id = data.get("training_run_id") # Idempotency check if training_run_id in _processed_training_runs: logger.debug("Already processed training_run_id=%s, skipping", training_run_id) return _processed_training_runs.add(training_run_id) # ... rest of handlerFor persistence across restarts, check Supabase before processing.
🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/services/agentgym-rl-coordinator/app.py` around lines 133 - 223, Add an idempotency guard to training_completed_handler to avoid duplicate processing: introduce a module-level set like _processed_training_runs and check membership of training_run_id at the start of training_completed_handler, return early if already seen and otherwise add it before proceeding; for durable/persistent guarantees also consult storage (e.g., call storage.record_event or a Supabase check) to confirm the training_run_id wasn't processed in prior runs before publishing via hf_publisher.publish_to_huggingface, and ensure any error paths do not leave false positives in the in-memory set (remove on failure or only add after successful publish/record).
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@pmoves/services/agentgym-rl-coordinator/app.py`:
- Around line 491-500: The trajectory_id validation currently catches ValueError
but raises HTTPException without chaining; change the except block in the
trajectory_ids validation (the loop using UUID(tid)) to catch ValueError as e
and re-raise the HTTPException using "from e" so the original traceback is
preserved (i.e., in the except ValueError as e: branch raise HTTPException(... )
from e, referencing UUID, trajectory_ids, and HTTPException).
- Around line 139-162: The handler is passing trajectory_ids directly to
hf_publisher.publish_to_huggingface without UUID validation; update the block
around trajectory_ids/model_id (the same flow that logs "Training completed..."
and calls hf_publisher.publish_to_huggingface) to validate each entry in
trajectory_ids as a proper UUID (use the same validation approach as the
/agentgym/dataset/publish endpoint), and if any ID is invalid either filter them
out and log a warning or reject the publish early with a warning and return;
ensure you reference trajectory_ids and hf_publisher.publish_to_huggingface so
only validated UUIDs reach the publisher.
In `@pmoves/services/evo-controller/agentgym_integration.py`:
- Line 298: The default base_model value for AGENTGYM_BASE_MODEL is missing the
HuggingFace org prefix; update the default in agentgym_integration.py where
base_model is set (variable base_model, env var AGENTGYM_BASE_MODEL) to include
the HF namespace (e.g., "Qwen/Qwen3-8B-Instruct") so the identifier is a valid
HuggingFace model name and matches env.agentgym.example.
---
Nitpick comments:
In `@pmoves/services/agentgym-rl-coordinator/app.py`:
- Around line 259-298: Move the inline import of PlainTextResponse out of the
metrics() function and add it to the module-level imports at the top of the
file; update the metrics() function to reference PlainTextResponse directly
(function name: metrics) so all imports are centralized and consistent with
other endpoints.
- Around line 133-223: Add an idempotency guard to training_completed_handler to
avoid duplicate processing: introduce a module-level set like
_processed_training_runs and check membership of training_run_id at the start of
training_completed_handler, return early if already seen and otherwise add it
before proceeding; for durable/persistent guarantees also consult storage (e.g.,
call storage.record_event or a Supabase check) to confirm the training_run_id
wasn't processed in prior runs before publishing via
hf_publisher.publish_to_huggingface, and ensure any error paths do not leave
false positives in the in-memory set (remove on failure or only add after
successful publish/record).
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: 71c5b045-4fc6-43d8-a3e4-6e896f409a00
📒 Files selected for processing (3)
pmoves/env.agentgym.examplepmoves/services/agentgym-rl-coordinator/app.pypmoves/services/evo-controller/agentgym_integration.py
Contributor
There was a problem hiding this comment.
Actionable comments posted: 2
🧹 Nitpick comments (3)
pmoves/services/agentgym-rl-coordinator/app.py (3)
306-307: Move import to module level.Importing
PlainTextResponseinside the function adds minor overhead on each request. Move it to the top of the file with other FastAPI imports.♻️ Proposed fix
At the top of the file (around line 10):
-from fastapi import FastAPI, HTTPException +from fastapi import FastAPI, HTTPException +from fastapi.responses import PlainTextResponseIn the metrics function:
- from fastapi.responses import PlainTextResponse return PlainTextResponse("\n".join(lines) + "\n", media_type="text/plain")🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/services/agentgym-rl-coordinator/app.py` around lines 306 - 307, The handler currently imports PlainTextResponse inside the function (in app.py near the metrics response), which is inefficient; move the import of PlainTextResponse to the module-level imports alongside other FastAPI imports at the top of app.py, then update the metrics function to return PlainTextResponse(...) using the now-module-level symbol instead of importing it inside the function.
133-233: Well-structured training completion handler with proper validation.The handler correctly:
- Validates
training_run_idformat againstDATASET_NAME_PATTERN- Validates
trajectory_idsas UUIDs before publishing (filtering invalid ones)- Uses UTC timestamps for events
- Publishes downstream events (
agentgym.model.published.v1,skills.pipeline.model-benchmark-viz.v1)- Logs when HF publisher is unavailable
One minor note: there's no idempotency guard to prevent duplicate HF publishes on NATS redelivery. If JetStream is used with at-least-once delivery, the same
training_run_idcould be published multiple times. Consider tracking published run IDs to skip duplicates.🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/services/agentgym-rl-coordinator/app.py` around lines 133 - 233, Add an idempotency guard in training_completed_handler to avoid double publishing on NATS redelivery: before calling hf_publisher.publish_to_huggingface check a persistent marker (e.g. storage.has_published(training_run_id) or storage.get_event/flag) and skip publish if already marked; after a successful publish (and after publishing agentgym.model.published.v1 and pipeline events) record the marker (e.g. storage.record_published(training_run_id) or include it in storage.record_event) so subsequent deliveries detect and skip the work; apply this around hf_publisher.publish_to_huggingface and the nc.publish calls and use the existing storage abstraction (storage.record_event or add storage.has_published/storage.record_published) to ensure idempotency across restarts.
53-61: HF_TOKEN is validated late at publish time, not at startup.Per context snippet 1,
HuggingFacePublisher(line 57) is initialized even ifHF_TOKENis not set. The validation happens atpublish_to_huggingface()time (raisesValueErrorif missing). This means you won't discover a missing token until a training completion event triggers publishing.Consider logging a warning at startup if
HF_TOKENis not configured:💡 Optional improvement
hf_publisher = HuggingFacePublisher(SUPABASE_URL, SUPABASE_KEY, HF_TOKEN, HF_ORG) + if not HF_TOKEN: + logger.warning("HF_TOKEN not configured - HuggingFace auto-publishing will fail") logger.info("Storage and coordinators initialized")🤖 Prompt for AI Agents
Verify each finding against the current code and only fix it if needed. In `@pmoves/services/agentgym-rl-coordinator/app.py` around lines 53 - 61, HuggingFace publishing token (HF_TOKEN) is only validated later in publish_to_huggingface(), so initialize-time will silently succeed and only fail at publish; update the startup block that constructs HuggingFacePublisher to validate HF_TOKEN up-front: if HF_TOKEN is present instantiate HuggingFacePublisher(HF_TOKEN, HF_ORG, SUPABASE_URL, SUPABASE_KEY) as before, otherwise set hf_publisher to None (or a no-op placeholder) and logger.warning that HuggingFace publishing is disabled until HF_TOKEN is configured so missing tokens are detected at startup.
🤖 Prompt for all review comments with AI agents
Verify each finding against the current code and only fix it if needed.
Inline comments:
In `@pmoves/env.agentgym.example`:
- Around line 14-19: The docker-compose default for the AGENTGYM_BASE_MODEL
environment var is inconsistent with the example; update the fallback value used
in the docker-compose agentgym service (the AGENTGYM_BASE_MODEL env entry in
docker-compose.agentgym.yml) from "Qwen2.5-7B-Instruct" to "Qwen/Qwen3-8B" so it
matches the AGENTGYM_BASE_MODEL in pmoves/env.agentgym.example; ensure only the
string value is changed and keep surrounding formatting intact.
In `@pmoves/services/evo-controller/agentgym_integration.py`:
- Line 370: The fallback base model string is inconsistent: update the default
used when computing base_model (the assignment to base_model) in
on_training_completed to match the value used in launch_agentgym_training
("Qwen/Qwen3-8B") so the same model ID is produced across flows; locate the
base_model variable assignment in on_training_completed and replace the current
"Qwen3-8B-Instruct" fallback with "Qwen/Qwen3-8B".
---
Nitpick comments:
In `@pmoves/services/agentgym-rl-coordinator/app.py`:
- Around line 306-307: The handler currently imports PlainTextResponse inside
the function (in app.py near the metrics response), which is inefficient; move
the import of PlainTextResponse to the module-level imports alongside other
FastAPI imports at the top of app.py, then update the metrics function to return
PlainTextResponse(...) using the now-module-level symbol instead of importing it
inside the function.
- Around line 133-233: Add an idempotency guard in training_completed_handler to
avoid double publishing on NATS redelivery: before calling
hf_publisher.publish_to_huggingface check a persistent marker (e.g.
storage.has_published(training_run_id) or storage.get_event/flag) and skip
publish if already marked; after a successful publish (and after publishing
agentgym.model.published.v1 and pipeline events) record the marker (e.g.
storage.record_published(training_run_id) or include it in storage.record_event)
so subsequent deliveries detect and skip the work; apply this around
hf_publisher.publish_to_huggingface and the nc.publish calls and use the
existing storage abstraction (storage.record_event or add
storage.has_published/storage.record_published) to ensure idempotency across
restarts.
- Around line 53-61: HuggingFace publishing token (HF_TOKEN) is only validated
later in publish_to_huggingface(), so initialize-time will silently succeed and
only fail at publish; update the startup block that constructs
HuggingFacePublisher to validate HF_TOKEN up-front: if HF_TOKEN is present
instantiate HuggingFacePublisher(HF_TOKEN, HF_ORG, SUPABASE_URL, SUPABASE_KEY)
as before, otherwise set hf_publisher to None (or a no-op placeholder) and
logger.warning that HuggingFace publishing is disabled until HF_TOKEN is
configured so missing tokens are detected at startup.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Path: .coderabbit.yaml
Review profile: CHILL
Plan: Pro
Run ID: 5406c533-2513-427f-b76f-34e4907fcaab
📒 Files selected for processing (3)
pmoves/env.agentgym.examplepmoves/services/agentgym-rl-coordinator/app.pypmoves/services/evo-controller/agentgym_integration.py
POWERFULMOVES
pushed a commit
that referenced
this pull request
Mar 15, 2026
…w threads New /pr-trim skill with 4-way classification (actionable / design-decision / false-positive / nitpick), GraphQL resolveReviewThread mutation, Make targets, NATS event schema, FlOO$ chain integration, and agent registry entry. Tested: 36 threads resolved across PRs #935-938 in batch trim session. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
5 tasks
…-publish New NATS event chain: agentgym.train.completed.v1 → auto-publish to HF → agentgym.model.published.v1. Updated base model default to Qwen3-8B-Instruct. Triggers benchmark-viz pipeline on completion. - AgentGym-RL coordinator: HF publisher with repo creation and model upload - Evo controller: training completion callback with NATS event emission - Updated env.agentgym.example with HF_TOKEN and HF_ORGANIZATION - Added agentgym.* NATS subjects to context docs Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
- Rename /health → /healthz to follow PMOVES convention - Add /metrics endpoint with Prometheus-format gauges - Validate trajectory_ids as UUIDs before HF publish - Fix timestamp: use datetime.now(timezone.utc) instead of local time.strftime - Add HF_TOKEN and HF_ORG to env.agentgym.example Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
…p to published event The NATS training_completed_handler was bypassing DATASET_NAME_PATTERN validation that the HTTP endpoint enforces. Also adds missing timestamp field to agentgym.model.published.v1 for cross-service event correlation. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Addresses 4 CodeRabbit review threads: - Thread 1+5: Fix HF model ID from 'Qwen3-8B-Instruct' to 'Qwen/Qwen3-8B' (env.agentgym.example + agentgym_integration.py default) - Thread 3: Add UUID validation for trajectory_ids in NATS handler before HuggingFace publishing (skip invalid, log warning) - Thread 4: Chain ValueError with 'from e' in trajectory_id validation Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Second occurrence of old 'Qwen3-8B-Instruct' default missed in previous commit. Now all defaults use 'Qwen/Qwen3-8B'. Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
POWERFULMOVES
force-pushed
the
feat/evoswarm-hf-publisher
branch
from
March 15, 2026 09:30
41df1b7 to
adef556
Compare
7 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
agentgym.train.completed.v1→ auto-publish to HF →agentgym.model.published.v1Files Changed
pmoves/services/agentgym-rl-coordinator/app.py— HF publisher endpointpmoves/services/evo-controller/agentgym_integration.py— training completion callbackpmoves/env.agentgym.example— HF_TOKEN and HF_ORGANIZATION vars.claude/context/nats-subjects.md— agentgym.* NATS subjectsTest plan
🤖 Generated with Claude Code
Summary by CodeRabbit
New Features
Updates
Bug Fixes