[jadenfix]

No description provided.

[greptile-apps]

Greptile Summary

This PR is a major version bump (0.2.10 → 1.0.0) that replaces the hand-written HTTP layer and Pydantic response models with a thin facade over an OpenAPI-generated client (roe._generated). The custom RoeHTTPClient is deleted, retry logic moves into a new RoeRetryTransport, and all API classes now delegate to generated endpoint modules while preserving the same public method signatures.

• RoeHTTPClient removed: replaced by a shared httpx.Client wrapped in RoeRetryTransport; retry policy broadened to cover 408/429/5xx and transport errors across all methods, with multipart run calls opting out via x-roe-skip-retry.
• Response models deleted: Agent, Policy, UserInfo, and related Pydantic classes are gone; callers should use roe._generated.models; only Job, JobBatch, JobStatus, and FileUpload are kept in roe.models.
• New UsersAPI: client.users.me() exposes the current-user endpoint; covered by new unit tests.

Confidence Score: 3/5

Not safe to merge without verifying the generated AgentJobStatus model shape — batch job polling may silently hang, and failed-job result fetches may raise unexpectedly.

The transport, exceptions, and API facade layers are well-structured. However, Job.wait() now only catches RoeAPIException when fetching a failed job's result — transport errors after retry exhaustion will propagate as uncaught exceptions instead of returning the documented placeholder. More critically, JobBatch.wait() relies on _extract_id finding a job-ID attribute on the generated AgentJobStatus model, but that model may not carry that field (the old hand-written single-job status also lacked it), which would cause every JobBatch.wait() call to spin until timeout.

src/roe/models/job.py warrants the most attention — both the single-job and batch wait loops have potential failure modes that need verification against the actual generated AgentJobStatus model shape.

Important Files Changed

Filename	Overview
src/roe/models/job.py	Core Job/JobBatch helpers migrated to generated models; two logic issues: exception handler narrowed too far for failed-job result fetches, and batch job-ID extraction may always return None (hanging wait loops).
src/roe/api/agents.py	Agents API refactored as a thin facade over generated client; multipart run calls use call_dynamic correctly; retrieve() uses manual from_dict while other methods use resp.parsed — inconsistent but not a bug.
src/roe/api/policies.py	Policies API migrated cleanly; PolicyVersionsAPI now owns its own config/raw reference rather than delegating to a parent; _normalize_policy_version_wire zero-UUID patch is a plausible workaround for nullable field.
src/roe/utils/transport.py	New RoeRetryTransport correctly gates on x-roe-skip-retry header, retries 5xx/408/429 and TransportErrors with exponential backoff capped at 10 s.
src/roe/utils/_dynamic_call.py	Multipart-bypass helper strips generated body and injects correct data=/files= kwargs; sets x-roe-skip-retry to prevent transport-layer replay of non-idempotent uploads.
src/roe/utils/inputs.py	Extracted multipart-builder from http_client.py; adds MIME-type detection for file paths; silently omits None values instead of sending empty strings (behaviour change from old code).
src/roe/exceptions.py	Added translate_response() that handles both httpx.Response and generated Response objects; correctly extracts DRF error shapes including list-style validation errors.
src/roe/client.py	Client wiring simplified: shared httpx.Client with RoeRetryTransport, raw AuthenticatedClient reuses it; new UsersAPI exposed; http_client property removed (breaking change documented in CHANGELOG).
src/roe/api/users.py	New UsersAPI.me() works correctly; manually parses JSON body as a fallback since resp.parsed may be None when the generated client doesn't populate it; covered by unit tests.
src/roe/utils/generated_request.py	request_raw/request_json helpers force JSON content-type to fix openapi-python-client's multipart-first code-generation bug; straightforward and correct.

Sequence Diagram

sequenceDiagram
    participant Caller
    participant RoeClient
    participant AgentsAPI
    participant call_dynamic
    participant RoeRetryTransport
    participant Server

    Caller->>RoeClient: client.agents.run(agent_id, **inputs)
    RoeClient->>AgentsAPI: run(agent_id, inputs, metadata)
    AgentsAPI->>call_dynamic: ep._get_kwargs() + inject multipart
    call_dynamic->>RoeRetryTransport: httpx.Client.request(**kwargs)
    Note over call_dynamic,RoeRetryTransport: x-roe-skip-retry set — no retry
    RoeRetryTransport->>Server: POST /v1/agents/run/{id}/async/
    Server-->>RoeRetryTransport: 200 job_id
    RoeRetryTransport-->>call_dynamic: httpx.Response
    call_dynamic-->>AgentsAPI: httpx.Response
    AgentsAPI-->>Caller: Job(job_id)

    Caller->>Caller: job.wait()
    loop until terminal status
        Caller->>AgentsAPI: jobs.retrieve_status(job_id)
        AgentsAPI->>RoeRetryTransport: GET /v1/agents/jobs/{id}/status/
        Note over AgentsAPI,RoeRetryTransport: 5xx/408/429 retried with backoff
        RoeRetryTransport-->>AgentsAPI: AgentJobStatus
    end
    AgentsAPI->>AgentsAPI: jobs.retrieve_result(job_id)
    AgentsAPI-->>Caller: AgentJobResultResponse

Loading

Comments Outside Diff (1)

1. src/roe/models/job.py, line 375-379 (link)

   Narrowed exception catch may swallow failures for terminal jobs

   The old code used except Exception to ensure that any error fetching the result of a failed/cancelled job (network issues, parse failures, unexpected status after all retries, etc.) was silently swallowed and replaced with a placeholder _empty_result. The new code catches only RoeAPIException, so if retrieve_result() raises an httpx.TransportError or another non-RoeAPIException for a terminal failed job (e.g., after the retry budget is exhausted at the transport layer), the exception propagates to the caller instead of returning the synthesized result. The CHANGELOG documents the placeholder behaviour as an intentional guarantee, so this regression breaks that contract.

   

_{Reviews (1): Last reviewed commit: "Merge pull request #27 from roe-ai/jaden..." | Re-trigger Greptile}

[greptile-apps]

JobBatch.wait() may hang if the generated AgentJobStatus has no job-ID field

_extract_id probes id, agent_job_id, and job_id on each item returned by retrieve_status_many. The old AgentJobStatusBatch model explicitly had an id: str field carrying the job UUID. The new implementation casts the batch-endpoint response to list[AgentJobStatus] (roe._generated.models.agent_job_status), the same type used for single-job status — which, like the old hand-written AgentJobStatus, likely has only status, timestamp, and error_message.

If none of the three probed attribute names resolves to a non-None value, _extract_id returns None and the continue discards every status item. completed_in_this_batch stays empty every cycle, and the batch only terminates when it times out.

[greptile-apps]

Silent None-value omission is a behaviour change from the previous implementation

The old _process_inputs in RoeHTTPClient sent an empty string ("") when a non-file input was None. The new code omits the field entirely. Agents that declare optional inputs and rely on receiving an empty string for unset fields will no longer get that signal; instead, the field is absent from the multipart form, which the server may treat as a validation error or as a distinct "field not provided" semantic.

Suggested change

	else:
	if value is not None:
	form_data[key] = str(value)
	else:
	form_data[key] = str(value) if value is not None else ""