Reach a host's Ollama from a containerized coder

Reach a host's Ollama from a containerized coder

The local lane hardcoded 127.0.0.1:11434 in three places — the reply
source, model resolution, and default discovery — and inside a
container that address is the container. All three now honor the
standard OLLAMA_HOST variable, resolving against the server the
session will actually talk to; the first containerized attempt failed
in resolution ("none are installed") against a host that had the
model. The Gym adapter maps Harbor's ollama/<name> spelling onto the
CLI's --model ollama:<name> local lane, passes the host, and drops the
token requirement there — a local model spends no grant.

Proven by the third leg of the first model comparison: qwen3.8 27B on
the host's own Ollama passed the same graded task as the two cloud
lanes.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GoYpb8FEmdxVErsv7ABCYi

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GoYpb8FEmdxVErsv7ABCYi

Deploy story

What this commit did to the running system — joined from the forge receipt chain, the part a commit page elsewhere cannot show.

Not deployed through the forge lane

No push, promotion, build, or deploy receipt references this commit (receipts are scanned over a bounded recent window). Changes shipped by full node replacement carry their proof in the release gate receipt instead.

Changed files

modified bench/adapters/openagents_coder.py
modified packages/openagents-cli/src/cli.ts

Diff

2 files changed, +29 -5

    @property
    def _token(self) -> str:
        token = os.environ.get("OPENAGENTS_TOKEN", "")
        if not token:
        if not token and not self._local_lane:
            raise ValueError(
                "OPENAGENTS_TOKEN is not set. The coder's thread lane needs a "
                "chat:account token for the server at OPENAGENTS_CODER_API_URL."

        if not self.model_name:
            return None
        # Harbor spells models provider/name; the catalog id is the name.
        return self.model_name.split("/", 1)[-1]
        # An `ollama/<name>` model selects the coder's local lane instead:
        # the CLI's `--model ollama:<name>` shape, answered by an Ollama
        # server the container reaches on its host.
        provider, _, name = self.model_name.partition("/")
        if provider == "ollama":
            return f"ollama:{name}"
        return name or provider

    @property
    def _local_lane(self) -> bool:
        return bool(self.model_name) and self.model_name.startswith("ollama/")

    @override
    async def install(self, environment: BaseEnvironment) -> None:

            "exit $status"
        )

        env = {"OPENAGENTS_TOKEN": self._token}
        if self._local_lane:
            env["OLLAMA_HOST"] = os.environ.get(
                "OPENAGENTS_CODER_OLLAMA_HOST", "http://host.docker.internal:11434"
            )

        await self.exec_as_agent(
            environment,
            command=command,
            env={"OPENAGENTS_TOKEN": self._token},
            env=env,
        )

      // coder without asking for one.
      const localModel =
        local && named === undefined && !offline && !resume
          ? yield* Effect.promise(() => discoverOllamaModel())
          ? yield* Effect.promise(() => discoverOllamaModel(process.env["OLLAMA_HOST"] || undefined))
          : undefined;

      if (local && named === undefined && localModel === undefined && !offline && !resume) {

      // already a real name and needs no round trip.
      const resolved =
        wantsOllama && askedFor !== undefined && named !== undefined
          ? yield* Effect.promise(() => resolveOllamaModel(askedFor))
          ? yield* Effect.promise(() =>
              // Same OLLAMA_HOST honor as the reply source below: resolution
              // must ask the server the session will actually talk to.
              resolveOllamaModel(askedFor, process.env["OLLAMA_HOST"] || undefined),
            )
          : undefined;

      if (resolved !== undefined && resolved.model === undefined) {

          : wantsOllama && ollamaName !== undefined
            ? new OllamaReplySource({
                model: ollamaName,
                // The standard Ollama env var, honored so a session in a
                // container can reach the Ollama server on its host —
                // 127.0.0.1 inside a container is the container.
                ...(process.env["OLLAMA_HOST"] ? { host: process.env["OLLAMA_HOST"] } : {}),
                ...(Option.isSome(reasoning) ? { reasoning: reasoning.value } : {}),
              })
            : (thread ?? new DummyReplySource());

58	58		`@property`
59	59		`def _token(self) -> str:`
60	60		`token = os.environ.get("OPENAGENTS_TOKEN", "")`
61		-	`if not token:`
	61	+	`if not token and not self._local_lane:`
62	62		`raise ValueError(`
63	63		`"OPENAGENTS_TOKEN is not set. The coder's thread lane needs a "`
64	64		`"chat:account token for the server at OPENAGENTS_CODER_API_URL."`

70	70		`if not self.model_name:`
71	71		`return None`
72	72		`# Harbor spells models provider/name; the catalog id is the name.`
73		-	`return self.model_name.split("/", 1)[-1]`
	73	+	# An `ollama/<name>` model selects the coder's local lane instead:
	74	+	# the CLI's `--model ollama:<name>` shape, answered by an Ollama
	75	+	`# server the container reaches on its host.`
	76	+	`provider, _, name = self.model_name.partition("/")`
	77	+	`if provider == "ollama":`
	78	+	`return f"ollama:{name}"`
	79	+	`return name or provider`
	80	+
	81	+	`@property`
	82	+	`def _local_lane(self) -> bool:`
	83	+	`return bool(self.model_name) and self.model_name.startswith("ollama/")`
74	84
75	85		`@override`
76	86		`async def install(self, environment: BaseEnvironment) -> None:`

136	146		`"exit $status"`
137	147		`)`
138	148
	149	+	`env = {"OPENAGENTS_TOKEN": self._token}`
	150	+	`if self._local_lane:`
	151	+	`env["OLLAMA_HOST"] = os.environ.get(`
	152	+	`"OPENAGENTS_CODER_OLLAMA_HOST", "http://host.docker.internal:11434"`
	153	+	`)`
	154	+
139	155		`await self.exec_as_agent(`
140	156		`environment,`
141	157		`command=command,`
142		-	`env={"OPENAGENTS_TOKEN": self._token},`
	158	+	`env=env,`
143	159		`)`

1965	1965		`// coder without asking for one.`
1966	1966		`const localModel =`
1967	1967		`local && named === undefined && !offline && !resume`
1968		-	`? yield* Effect.promise(() => discoverOllamaModel())`
	1968	+	`? yield* Effect.promise(() => discoverOllamaModel(process.env["OLLAMA_HOST"] \|\| undefined))`
1969	1969		`: undefined;`
1970	1970
1971	1971		`if (local && named === undefined && localModel === undefined && !offline && !resume) {`

2016	2016		`// already a real name and needs no round trip.`
2017	2017		`const resolved =`
2018	2018		`wantsOllama && askedFor !== undefined && named !== undefined`
2019		-	`? yield* Effect.promise(() => resolveOllamaModel(askedFor))`
	2019	+	`? yield* Effect.promise(() =>`
	2020	+	`// Same OLLAMA_HOST honor as the reply source below: resolution`
	2021	+	`// must ask the server the session will actually talk to.`
	2022	+	`resolveOllamaModel(askedFor, process.env["OLLAMA_HOST"] \|\| undefined),`
	2023	+	`)`
2020	2024		`: undefined;`
2021	2025
2022	2026		`if (resolved !== undefined && resolved.model === undefined) {`

2194	2198		`: wantsOllama && ollamaName !== undefined`
2195	2199		`? new OllamaReplySource({`
2196	2200		`model: ollamaName,`
	2201	+	`// The standard Ollama env var, honored so a session in a`
	2202	+	`// container can reach the Ollama server on its host —`
	2203	+	`// 127.0.0.1 inside a container is the container.`
	2204	+	`...(process.env["OLLAMA_HOST"] ? { host: process.env["OLLAMA_HOST"] } : {}),`
2197	2205		`...(Option.isSome(reasoning) ? { reasoning: reasoning.value } : {}),`
2198	2206		`})`
2199	2207		`: (thread ?? new DummyReplySource());`