OpenAgents

Commit

Add Ollama local model support to openagents coder

586edf191fb2 · AtlantisPleb · · parent b50ab6ccea2f 

Add Ollama local model support to openagents coder

Adds an Ollama backend so openagents coder can chat with locally-served
models. The --model flag accepts ollama:<name>; the CLI then uses the
ollama-js client against http://127.0.0.1:11434 and yields the same
ReplySource chunks as the thread-backed path.

Closes #21.
Closes
#21

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 packages/openagents-cli/package.json
  • modified packages/openagents-cli/src/cli.ts
  • added packages/openagents-cli/src/coder-ollama.ts
  • modified pnpm-lock.yaml

Diff

4 files changed, +172 -53

packages/openagents-cli/package.json modified +1

@@ -55,6 +55,7 @@

55 55 
    "@effect/platform-node": "catalog:",
56 56 
    "@effect/platform-node-shared": "4.0.0-beta.94",
57 57 
    "effect": "catalog:",
58 
    "ollama": "^0.6.3",
58 59 
    "ws": "8.21.1"
59 60 
  },
60 61 
  "devDependencies": {
packages/openagents-cli/src/cli.ts modified +65 -31

@@ -27,6 +27,7 @@ import { CoderSession, DummyReplySource } from "./coder-session.js";

27 27 
import { CoderTaskRegistry } from "./coder-tasks.js";
28 28 
import { runCoderUi } from "./coder-ui.js";
29 29 
import { backendIds } from "./coder-backends.js";
30 
import { OllamaReplySource, isOllamaModelFlag, parseOllamaModelFlag } from "./coder-ollama.js";
30 31 
import { openThread, ThreadUnavailable, type ThreadReplySource } from "./coder-thread.js";
31 32 
import { delegateTool } from "./coder-tools.js";
32 33 
import { describeWorkspace } from "./coder-workspace.js";

@@ -1415,13 +1416,16 @@ const coderReasoningFlag = Flag.choice("reasoning", [

1415 1416 
  Flag.optional,
1416 1417 
  Flag.withDescription("Reasoning effort recorded on the thread as its admitted execution shape"),
1417 1418 
);
1418 
// The accepted values are still the chat API's published backends. A thread's
1419 
// grant pins its own model and `POST /api/v3/threads` publishes no model
1420 
// parameter, so naming one here cannot change which model answers; the session
1421 
// says so rather than letting the flag look like it worked.
1422 
const coderModelFlag = Flag.choice("model", backendIds() as string[]).pipe(
1419 
// `--model` can name an `ollama:<model>` local model or a chat API backend.
1420 
// For a chat API backend a thread's grant still pins its own model and
1421 
// `POST /api/v3/threads` publishes no model parameter, so naming one cannot
1422 
// change which model answers. For `ollama:<model>` the local Ollama server is
1423 
// used directly and the named model is the one that runs.
1424 
const coderModelFlag = Flag.string("model").pipe(
1423 1425 
  Flag.optional,
1424 
  Flag.withDescription("A chat API backend. The thread's grant pins its own model"),
1426 
  Flag.withDescription(
1427 
    "A model name. Use `ollama:<model>` for a local Ollama server, or a chat API backend id",
1428 
  ),
1425 1429 
);
1426 1430
1427 1431 
/**

@@ -1640,35 +1644,58 @@ const coderCommand = Command.make(

1640 1644 
      const workspace = describeWorkspace();
1641 1645 
      const endpoint = yield* resolveApiEndpoint(endpointOverrides(flags));
1642 1646
1647 
      // A `--model` value that starts with `ollama:` goes to the local Ollama
1648 
      // server and needs no account credential.
1649 
      const wantsOllama = Option.isSome(model) && isOllamaModelFlag(model.value);
1650 
      const ollamaName =
1651 
        wantsOllama && Option.isSome(model) ? parseOllamaModelFlag(model.value) : undefined;
1652
1653 
      // Any other `--model` value still has to name a published backend. The
1654 
      // flag takes a string so an `ollama:` prefix can reach the local server,
1655 
      // which costs the enum `Flag.choice` used to enforce, so the check moves
1656 
      // here rather than disappearing.
1657 
      if (Option.isSome(model) && !wantsOllama && !backendIds().includes(model.value)) {
1658 
        return yield* new InputError({
1659 
          message: `Unknown model ${model.value}. Use ollama:<model> for a local Ollama server, or one of: ${backendIds().join(", ")}.`,
1660 
        });
1661 
      }
1662
1643 1663 
      // The session opens a thread of its own and spends that thread's grant,
1644 1664 
      // so the CLI still holds no provider key and nothing typed here reaches
1645 1665 
      // the account's conversation. Without a credential it falls back to the
1646 1666 
      // stand-in and says so rather than failing.
1647 
      const stored = offline
1648 
        ? Option.none()
1649 
        : yield* findToken(endpoint.origin).pipe(
1650 
            Effect.catchTag("OpenAgentsCli.CredentialPersistenceUnavailable", () =>
1651 
              Effect.succeed(Option.none()),
1652 
            ),
1653 
          );
1667 
      const stored =
1668 
        offline || wantsOllama
1669 
          ? Option.none()
1670 
          : yield* findToken(endpoint.origin).pipe(
1671 
              Effect.catchTag("OpenAgentsCli.CredentialPersistenceUnavailable", () =>
1672 
                Effect.succeed(Option.none()),
1673 
              ),
1674 
            );
1654 1675
1655 
      const thread = Option.isSome(stored)
1656 
        ? yield* Effect.tryPromise({
1657 
            try: () =>
1658 
              openThread({
1659 
                origin: endpoint.origin,
1660 
                token: Redacted.value(stored.value.token),
1661 
                objective: `openagents coder in ${workspace.repository} on ${workspace.branch}`,
1662 
                reasoning: Option.getOrUndefined(reasoning),
1663 
              }),
1664 
            // The server's own code and sentence, which is what turns a ninth
1665 
            // concurrent session from an obscure failure into an instruction
1666 
            // naming the ceiling and how many threads the account is holding.
1667 
            catch: (cause) => coderRefusal(endpoint.origin, cause),
1668 
          })
1669 
        : undefined;
1676 
      const thread =
1677 
        Option.isSome(stored) && !wantsOllama
1678 
          ? yield* Effect.tryPromise({
1679 
              try: () =>
1680 
                openThread({
1681 
                  origin: endpoint.origin,
1682 
                  token: Redacted.value(stored.value.token),
1683 
                  objective: `openagents coder in ${workspace.repository} on ${workspace.branch}`,
1684 
                  reasoning: Option.getOrUndefined(reasoning),
1685 
                }),
1686 
              // The server's own code and sentence, which is what turns a ninth
1687 
              // concurrent session from an obscure failure into an instruction
1688 
              // naming the ceiling and how many threads the account is holding.
1689 
              catch: (cause) => coderRefusal(endpoint.origin, cause),
1690 
            })
1691 
          : undefined;
1670 1692
1671 
      const source = thread ?? new DummyReplySource();
1693 
      // A `--model ollama:<name>` session answers from the local Ollama server,
1694 
      // so it takes neither a thread nor the stand-in.
1695 
      const source =
1696 
        wantsOllama && ollamaName !== undefined
1697 
          ? new OllamaReplySource({ model: ollamaName })
1698 
          : (thread ?? new DummyReplySource());
1672 1699
1673 1700 
      // Children get their own thread on their own model. The conversation
1674 1701 
      // stays on the model it opened with, and a fan-out spends a budget the

@@ -1720,17 +1747,24 @@ const coderCommand = Command.make(

1720 1747 
        session.notice(`This session cannot delegate: ${childThread.reason}`);
1721 1748 
      }
1722 1749
1723 
      if (Option.isNone(stored) && !offline) {
1750 
      if (Option.isNone(stored) && !offline && !wantsOllama) {
1724 1751 
        session.notice(
1725 1752 
          "No stored credential, so replies come from the built-in stand-in. " +
1726 1753 
            "Run `openagents auth login` to reach a real model.",
1727 1754 
        );
1728 1755 
      }
1729 1756
1757 
      if (wantsOllama && ollamaName === undefined) {
1758 
        session.notice(
1759 
          "`--model ollama:` is missing a model name. Use `ollama:<model>`, " +
1760 
            "for example `ollama:qwen3.8:27b-mtp-q8_0`.",
1761 
        );
1762 
      }
1763
1730 1764 
      // A grant pins the model the proxy will use, and the thread route takes
1731 1765 
      // no model parameter, so a named backend cannot reach this turn. Saying
1732 1766 
      // nothing would leave a reader with a flag that appeared to work.
1733 
      if (thread !== undefined && Option.isSome(model)) {
1767 
      if (thread !== undefined && Option.isSome(model) && !wantsOllama) {
1734 1768 
        session.notice(
1735 1769 
          `This thread's grant pins ${thread.model}. \`--model\` names a chat API ` +
1736 1770 
            "backend, which the inference proxy does not route to, so it had no effect.",
packages/openagents-cli/src/coder-ollama.ts added +92

@@ -0,0 +1,92 @@

1 
/**
2 
 * A reply source that calls a local Ollama server.
3 
 *
4 
 * The `ollama` client is used directly, not through the OpenAgents proxy, so
5 
 * the caller's machine must already be running an Ollama server. The default
6 
 * endpoint is `http://127.0.0.1:11434` and the model is read from the
7 
 * `ollama:<name>` shape of the `--model` flag.
8 
 *
9 
 * Ollama chat messages are kept locally in this source; nothing is sent to the
10 
 * OpenAgents chat API. A local model spends no metered budget, so `budget` is
11 
 * left undefined.
12 
 */
13
14 
import { Ollama } from "ollama";
15
16 
import type { ReplyChunk, ReplySource } from "./coder-session.js";
17
18 
const DEFAULT_HOST = "http://127.0.0.1:11434";
19
20 
export interface OllamaOptions {
21 
  /** The Ollama model name, without the `ollama:` prefix. */
22 
  readonly model: string;
23 
  /** The Ollama server endpoint. Defaults to `http://127.0.0.1:11434`. */
24 
  readonly host?: string | undefined;
25 
}
26
27 
/** True when `--model` names an Ollama source. */
28 
export const isOllamaModelFlag = (value: string): boolean => value.startsWith("ollama:");
29
30 
/** Extract the Ollama model name from an `ollama:<name>` flag value. */
31 
export const parseOllamaModelFlag = (value: string): string | undefined => {
32 
  const match = /^ollama:(.+)$/.exec(value);
33 
  return match?.[1]?.trim();
34 
};
35
36 
interface WireMessage {
37 
  readonly role: "user" | "assistant";
38 
  readonly content: string;
39 
}
40
41 
export class OllamaReplySource implements ReplySource {
42 
  private readonly client: Ollama;
43 
  private readonly modelName: string;
44 
  private readonly transcript: WireMessage[] = [];
45
46 
  get model(): string {
47 
    return `Ollama ${this.modelName}`;
48 
  }
49
50 
  constructor(options: OllamaOptions) {
51 
    this.client = new Ollama({ host: options.host ?? DEFAULT_HOST });
52 
    this.modelName = options.model;
53 
  }
54
55 
  async *reply(prompt: string, signal: AbortSignal): AsyncIterable<ReplyChunk> {
56 
    this.transcript.push({ role: "user", content: prompt });
57
58 
    let assistant = "";
59 
    const stream = await this.client.chat({
60 
      model: this.modelName,
61 
      messages: this.transcript,
62 
      stream: true,
63 
    });
64
65 
    const onAbort = () => stream.abort();
66 
    signal.addEventListener("abort", onAbort, { once: true });
67
68 
    try {
69 
      for await (const chunk of stream) {
70 
        if (signal.aborted) break;
71
72 
        const thinking = chunk.message.thinking;
73 
        if (typeof thinking === "string" && thinking.length > 0) {
74 
          yield { type: "reasoning", value: thinking };
75 
        }
76
77 
        const content = chunk.message.content;
78 
        if (typeof content === "string" && content.length > 0) {
79 
          assistant += content;
80 
          yield { type: "text", value: content };
81 
        }
82
83 
        if (chunk.done) break;
84 
      }
85 
    } finally {
86 
      signal.removeEventListener("abort", onAbort);
87 
      if (assistant.length > 0) {
88 
        this.transcript.push({ role: "assistant", content: assistant });
89 
      }
90 
    }
91 
  }
92 
}
pnpm-lock.yaml modified +14 -22

@@ -2217,6 +2217,9 @@ importers:

2217 2217 
      effect:
2218 2218 
        specifier: 4.0.0-beta.94
2219 2219 
        version: 4.0.0-beta.94
2220 
      ollama:
2221 
        specifier: ^0.6.3
2222 
        version: 0.6.3
2220 2223 
      ws:
2221 2224 
        specifier: 8.21.1
2222 2225 
        version: 8.21.1

@@ -8426,6 +8429,9 @@ packages:

8426 8429 
  ofetch@1.5.1:
8427 8430 
    resolution: {integrity: sha512-2W4oUZlVaqAPAil6FUg/difl6YhqhUR7x2eZY4bQCko22UXg3hptq9KLQdqFClV+Wu85UX7hNtdGTngi/1BxcA==}
8428 8431
8432 
  ollama@0.6.3:
8433 
    resolution: {integrity: sha512-KEWEhIqE5wtfzEIZbDCLH51VFZ6Z3ZSa6sIOg/E/tBV8S51flyqBOXi+bRxlOYKDf8i327zG9eSTb8IJxvm3Zg==}
8434
8429 8435 
  on-exit-leak-free@2.1.2:
8430 8436 
    resolution: {integrity: sha512-0eJJY6hXLGf1udHwfNftBqH+g73EU4B504nZeKpz1sYRKafAghwxEJunB2O7rDZkL4PGfsMVnTXZ2EjibbqcsA==}
8431 8437 
    engines: {node: '>=14.0.0'}

@@ -13361,13 +13367,14 @@ snapshots:

13361 13367 
    dependencies:
13362 13368 
      '@testing-library/dom': 10.4.1
13363 13369 
      '@testing-library/user-event': 14.6.1(@testing-library/dom@10.4.1)
13364 
      '@vitest/browser': 4.1.10(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))(vitest@4.1.10)
13370 
      '@vitest/browser': 4.1.10(vite@8.1.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(yaml@2.9.0))(vitest@4.1.10)
13365 13371 
      vitest: 4.1.10(@opentelemetry/api@1.9.0)(@types/node@24.13.1)(@vitest/browser-preview@4.1.10)(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))(happy-dom@20.10.6)
13366 13372 
    transitivePeerDependencies:
13367 13373 
      - bufferutil
13368 13374 
      - msw
13369 13375 
      - utf-8-validate
13370 13376 
      - vite
13377 
    optional: true
13371 13378
13372 13379 
  '@vitest/browser-preview@4.1.10(vite@8.1.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(yaml@2.9.0))(vitest@4.1.10)':
13373 13380 
    dependencies:

@@ -13380,24 +13387,6 @@ snapshots:

13380 13387 
      - msw
13381 13388 
      - utf-8-validate
13382 13389 
      - vite
13383 
    optional: true
13384
13385 
  '@vitest/browser@4.1.10(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))(vitest@4.1.10)':
13386 
    dependencies:
13387 
      '@blazediff/core': 1.9.1
13388 
      '@vitest/mocker': 4.1.10(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))
13389 
      '@vitest/utils': 4.1.10
13390 
      magic-string: 0.30.21
13391 
      pngjs: 7.0.0
13392 
      sirv: 3.0.2
13393 
      tinyrainbow: 3.1.0
13394 
      vitest: 4.1.10(@opentelemetry/api@1.9.0)(@types/node@24.13.1)(@vitest/browser-preview@4.1.10)(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))(happy-dom@20.10.6)
13395 
      ws: 8.21.1
13396 
    transitivePeerDependencies:
13397 
      - bufferutil
13398 
      - msw
13399 
      - utf-8-validate
13400 
      - vite
13401 13390
13402 13391 
  '@vitest/browser@4.1.10(vite@8.1.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(yaml@2.9.0))(vitest@4.1.10)':
13403 13392 
    dependencies:

@@ -13415,7 +13404,6 @@ snapshots:

13415 13404 
      - msw
13416 13405 
      - utf-8-validate
13417 13406 
      - vite
13418 
    optional: true
13419 13407
13420 13408 
  '@vitest/expect@4.1.10':
13421 13409 
    dependencies:

@@ -16094,6 +16082,10 @@ snapshots:

16094 16082 
      node-fetch-native: 1.6.7
16095 16083 
      ufo: 1.6.4
16096 16084
16085 
  ollama@0.6.3:
16086 
    dependencies:
16087 
      whatwg-fetch: 3.6.20
16088
16097 16089 
  on-exit-leak-free@2.1.2: {}
16098 16090
16099 16091 
  on-finished@2.3.0:

@@ -17436,8 +17428,8 @@ snapshots:

17436 17428 
    dependencies:
17437 17429 
      '@oxc-project/types': 0.138.0
17438 17430 
      '@oxlint/plugins': 1.68.0
17439 
      '@vitest/browser': 4.1.10(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))(vitest@4.1.10)
17440 
      '@vitest/browser-preview': 4.1.10(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))(vitest@4.1.10)
17431 
      '@vitest/browser': 4.1.10(vite@8.1.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(yaml@2.9.0))(vitest@4.1.10)
17432 
      '@vitest/browser-preview': 4.1.10(vite@8.1.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(yaml@2.9.0))(vitest@4.1.10)
17441 17433 
      '@vitest/expect': 4.1.10
17442 17434 
      '@vitest/mocker': 4.1.10(@voidzero-dev/vite-plus-core@0.2.4(@types/node@24.13.1)(esbuild@0.28.1)(jiti@2.7.0)(terser@5.49.0)(tsx@4.20.6)(typescript@6.0.3)(yaml@2.9.0))
17443 17435 
      '@vitest/pretty-format': 4.1.10

This page updates live while a promote is in flight · changelog