feat(agents): OpenAI-compatible endpoints for Pi via vault custom connections#5345
Merged
Merged
Conversation
dosubot
Bot
added
the
size:XXL
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
Member
Author
|
@coderabbitai review |
✅ Action performedReview finished.
|
📝 WalkthroughWalkthroughThis change adds OpenAI-compatible custom connection support for Pi. It normalizes and validates custom connections, generates isolated ChangesPi OpenAI-compatible custom models
Estimated code review effort: 4 (Complex) | ~60 minutes Sequence Diagram(s)sequenceDiagram
participant Client
participant SDKResolver
participant SandboxAgent
participant PiConfig
participant Daytona
Client->>SDKResolver: submit named custom connection
SDKResolver->>SDKResolver: resolve provider, deployment, endpoint, and secret name
SDKResolver-->>SandboxAgent: resolved run request
SandboxAgent->>PiConfig: build and serialize models.json
PiConfig-->>SandboxAgent: Pi model configuration
SandboxAgent->>Daytona: upload models.json before session creation
Daytona-->>SandboxAgent: prepared Pi environment
SandboxAgent->>Daytona: apply qualified slug/model identifier
Possibly related issues
Possibly related PRs
🚥 Pre-merge checks | ✅ 4 | ❌ 1❌ Failed checks (1 warning)
✅ Passed checks (4 passed)
✨ Finishing Touches📝 Generate docstrings
🧪 Generate unit tests (beta)
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}
There was a problem hiding this comment.
🧹 Nitpick comments (1)
sdks/python/agenta/sdk/agents/handler.py (1)
150-154: 📐 Maintainability & Code Quality | 🔵 Trivial | 💤 Low valueConsider refining the error message for valid providers on unsupported deployments.
If a provider is generally supported by the harness (e.g., Anthropic for Pi) but is rejected for the resolved deployment (e.g., the
customdeployment) viaharness_allows_pair, this logic falls through to raise anUnsupportedProviderError. The resulting error message ("provider 'X' is not supported") might confuse users who know the provider is inherently supported directly.Consider raising a distinct
UnsupportedPairErroror refining the error message in the future to clarify that the combination of the provider and deployment is not supported.
ℹ️ Review info
⚙️ Run configuration
Configuration used: Organization UI
Review profile: CHILL
Plan: Pro Plus
Run ID: ff208924-7dd7-4965-badb-96f7ff89e1f7
📒 Files selected for processing (35)
docs/design/agent-workflows/documentation/adapters/pi.mddocs/design/agent-workflows/interfaces/cross-service/service-to-vault-and-tool-providers.mddocs/design/agent-workflows/interfaces/in-service/model-connection-resolution.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/README.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/context.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/design.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/open-issues.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/plan.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/qa.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/research.mddocs/design/agent-workflows/projects/pi-openai-compatible-models/status.mdsdks/python/agenta/sdk/agents/__init__.pysdks/python/agenta/sdk/agents/adapters/harnesses.pysdks/python/agenta/sdk/agents/capabilities.pysdks/python/agenta/sdk/agents/connections/__init__.pysdks/python/agenta/sdk/agents/connections/errors.pysdks/python/agenta/sdk/agents/handler.pysdks/python/agenta/sdk/agents/platform/connections.pysdks/python/oss/tests/pytest/integration/agents/recordings/e2-pi-custom-openai-compatible.jsonsdks/python/oss/tests/pytest/integration/agents/test_custom_connection_replay.pysdks/python/oss/tests/pytest/unit/agents/connections/test_capabilities.pysdks/python/oss/tests/pytest/unit/agents/platform/test_connections_http.pysdks/python/oss/tests/pytest/unit/agents/test_agent_composition_seam.pysdks/python/oss/tests/pytest/unit/agents/test_harness_adapters.pysdks/python/oss/tests/pytest/unit/agents/test_invoke_failure_status.pyservices/runner/src/engines/sandbox_agent.tsservices/runner/src/engines/sandbox_agent/daytona.tsservices/runner/src/engines/sandbox_agent/pi-assets.tsservices/runner/src/engines/sandbox_agent/pi-model-config.tsservices/runner/tests/unit/sandbox-agent-daytona.test.tsservices/runner/tests/unit/sandbox-agent-model.test.tsservices/runner/tests/unit/sandbox-agent-orchestration.test.tsservices/runner/tests/unit/sandbox-agent-pi-assets.test.tsservices/runner/tests/unit/sandbox-agent-pi-model-config.test.tsservices/runner/tests/unit/session-pool.test.ts
…, and replay test Claude-Session: https://claude.ai/code/session_01AYBi6cYuQCR87dAD3ZdPWc
… test The default resolve stub in test_invoke_handler.py returned a successful runtime_provided ResolvedConnection pinned to provider 'openai'. With the new post-resolve harness capability gate this is rejected on the claude harness (openai is not a claude-consumable provider), breaking test_invoke_cross_harness_same_body_divergent_configs. Make the stub raise ConnectionResolutionError instead, exercising the real no-connection degraded path that every harness tolerates, which is what these response-body/lifecycle/cross-harness tests are meant to check.
mmabrouk
force-pushed
the
feat/pi-openai-compatible-models
branch
from
July 17, 2026 20:29
81636d6 to
4ef236b
Compare
Contributor
Railway Preview Environment
Updated at 2026-07-17T22:42:43.275Z |
mmabrouk
added a commit
that referenced
this pull request
Jul 18, 2026
…position v0.105.4 (#5345) added the OpenAI-compatible custom-connection model-config plan inside acquireEnvironment in the old monolith. This decomposition moved acquireEnvironment into environment.ts + environment-setup.ts, so on rebase that release logic is threaded through the new split: the plan is built and localModelConfigUnwritable computed in prepareEnvironmentSetup and returned in its bundle; the fail-loud/fail-closed throws, the Daytona asset piModelConfig arg, and the fully-qualified wantedModel selection live in acquireEnvironment. Behavior is identical to #5345.
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
1 participant
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.
feat(agents): OpenAI-compatible endpoints for Pi via vault custom connections
Context
You could store a custom connection (a base URL, a key, and a model id) in the vault, pick it for
a Pi agent, and the run would fail before it reached the model. A connection whose kind was not a
known provider family never passed the harness capability gate. Even when it did, the base URL was
dropped: the runner never wrote Pi a
models.json, so Pi had no idea the endpoint existed and fellback to its default provider. OpenAI-compatible endpoints (an Ollama gateway, an in-house proxy,
any self-named Chat Completions endpoint) could not run on Pi at all.
This change makes that path work end to end, with no vault schema change and no
/runschemachange.
Changes
The service formalizes the existing
customkind as an OpenAI-compatible Chat Completionsendpoint. A provider-less named connection with
kind=customresolves to provideropenai,deployment
custom, and its key inOPENAI_API_KEY. An explicit provider family is preserved, soexisting Anthropic gateway connections are unchanged. A new
harness_allows_pairtable decideswhich resolved
(provider, deployment)pairs a harness accepts. Pi plusopenaipluscustomisnewly allowed; arbitrary provider plus
customon Pi is still rejected. Named Agenta connectionsnow defer provider acceptance until after the vault record resolves, so the pre-resolve check
validates only the connection mode.
pi_coreandpi_agentapublish deployments["direct", "custom"].The connection
{mode, slug}reaches the runner throughmodel_refthreading inadapters/harnesses.pyfor the Pi and Agenta harnesses. It was not threading for a named customconnection, which live end-to-end testing surfaced. The slug is connection identity now, not a
provider family.
The runner writes Pi's
models.jsonfrom a new pure builder(
services/runner/src/engines/sandbox_agent/pi-model-config.ts). It applies only for Pi plusprovider
openaiplus deploymentcustomplus connection modeagenta.Before, a custom base URL never reached Pi. After, the runner writes:
{ "providers": { "my-ollama-gateway": { "baseUrl": "https://gateway.internal/v1", "api": "openai-completions", "apiKey": "$OPENAI_API_KEY", "models": [{ "id": "llama-3.1-70b" }] } } }The
apiKeyis an environment reference, so no secret lands on disk. Locally the runner writes thefile
0600and atomically into an isolated per-run Pi dir that carries no operatorauth.json, soa plan run never authenticates with the operator's login. On Daytona it uploads the file to
DAYTONA_PI_DIRbefore session creation. A no-plan run removes any stale file best-effort. When aplan exists, the runner requests the fully qualified
<slug>/<model>id, which fixes a suffixcollision that could silently route a request to
api.openai.com.Fail-loud throughout. An unusable base URL raises
EndpointResolutionError, a new subclass thatreturns HTTP 422 (as do
UnsupportedProviderErrorandUnsupportedDeploymentError; without anexplicit
status_codethese fell through to 500). The connection never resolves withendpoint=None, so the harness cannot fall back to a default endpoint. Builder, write, and uploadfailures are terminal.
Scope / risk
/runwire shape are untouched.No migration.
customdeployment with an unusable base URLnow fails loudly, including a Claude Anthropic gateway. Existing gateways with a valid base URL
are unaffected.
model_refthreading now puts the connection{mode, slug}on the wire for Pi and Agenta, whichchanges the config fingerprint used for warm-session reuse. Expect a one-time cold restart of
parked
self_managedsessions on first deploy. Steady state is unchanged.services/runner, so a runner image rebuild is needed on deploy.models.json.How to QA
and one model id. Pick it for a Pi agent and run. Expected: the run reaches the configured host,
not
api.openai.com, and the reply comes from the configured model.models.jsonis uploaded toDAYTONA_PI_DIRand the sandbox carries no operatorauth.json.the request routes to its own host, proving the slug keys the provider.
kind=customconnection with no base URL and run it.Expected: the run fails with HTTP 422 and an error naming the connection slug, not a 500 and not
a silent fallback.
model. Expected: the request uses the
<slug>/<model>id and reaches the custom host, notapi.openai.com.connection. Expected: both pass unchanged.
Test commands:
Evidence: SDK agents suite 669 passed, runner suite 1190 passed plus wire golden fixtures. Live
end-to-end on the EE dev stack (2026-07-14 and 2026-07-15) passed the happy path local and on
Daytona, the 422 fail-loud paths, and the Pi and Claude regression runs.
https://claude.ai/code/session_01AYBi6cYuQCR87dAD3ZdPWc