feat: Ollama provider — local & cloud LLM runtime by 3L0935 · Pull Request #2735 · pingdotgg/t3code

3L0935 · 2026-05-16T20:12:26Z

What changed

Adds Ollama as a first-class provider driver, alongside Codex, Claude, OpenCode and Cursor — backed by a local or cloud Ollama runtime.

Server (apps/server)

ollamaRuntime.ts — HTTP client for the Ollama REST API: /api/chat (streaming + non-streaming), /api/tags, /api/version. Bearer auth via OLLAMA_API_KEY.
OllamaTools.ts — 7 built-in tools the model can call: read_file, write_file, edit_file, bash, list_directory, search_files, web_fetch.
OllamaAdapter.ts — session lifecycle + tool-calling loop, reusing the shared approval flow (request.opened → Deferred → respondToRequest).
OllamaProvider.ts — status probe (version + model list) on a 5-minute refresh.
OllamaDriver.ts — wires adapter + text generation + snapshot into one driver; registered in builtInDrivers.ts.
OllamaTextGeneration.ts — commit message / PR content / branch name / thread title generation.

Contracts (packages/contracts)

OllamaSettings schema (enabled, baseUrl, model, customModels) + patch schema.
ollama driver kind and default models registered for model-selection routing.

Frontend (apps/web)

OllamaIcon SVG, wired into the settings panel and the provider icon map.

Bot-review fixes — issues raised by Cursor Bugbot / Macroscope, addressed across follow-up commits:

interruptTurn no longer permanently closes the session.
A previous turn's fiber is interrupted and awaited before a new turn starts, so two fibers can never mutate the message history concurrently; turn.completed is emitted exactly once per turn via Effect.onExit.
rollbackThread works — per-turn message offsets are tracked, so whole turns (incl. tool messages) splice correctly.
Default model unified to qwen2.5:7b across contracts, adapter and text generation.
The version probe is wrapped in Effect.exit — an unreachable server no longer crashes the whole status check; Ollama Cloud's placeholder 0.0.0 is hidden.
/api/tags parsing reads snake_case modified_at.
Text-generation errors are caught by tag instead of a blanket mapError, avoiding double-wrapped TextGenerationError.

Why

T3 Code only supported cloud agent providers. Ollama adds a local-first option (and Ollama Cloud) so users can run a coding agent against models on their own hardware — with the same approval-gated tool flow as the other providers.

Risk: the bash and file-write tools execute on the host with no sandboxing. They are gated by the standard approval flow in non-full-access runtime modes.

Screenshots

Provider list — Ollama with its icon, version chip and discovered models:

Checklist

This PR is small and focused
I explained what changed and why
I included before/after screenshots for any UI changes
I included a video for animation/interaction changes

Note

High Risk
Adds a new provider runtime that can execute host-side commands and file writes via model tool-calls (approval-gated but still security-sensitive) and introduces new network/streaming code paths for chat, model listing, and status probing.

Overview
Adds Ollama as a first-class provider across server, contracts, and web UI.

On the server, introduces a full Ollama driver stack: an HTTP runtime for /api/chat (plus streaming parser), /api/tags, and /api/version; a session adapter that runs turns, emits runtime events, and supports approval-gated tool calls (including bash/file operations) with interruption/rollback handling; a provider status probe + model discovery snapshot refreshed on an interval; and a text-generation implementation for commit/PR/branch/thread outputs.

In contracts and web, adds OllamaSettings (and patch support), registers default model routing for the ollama driver kind, and updates the UI to surface Ollama in provider pickers/settings plus a dedicated OllamaIcon.

^{Reviewed by Cursor Bugbot for commit 6f18cef. Bugbot is set up for automated code reviews on this repo. Configure here.}

Note

Add Ollama as a local and cloud LLM provider runtime

Adds a full ollama provider driver (OllamaDriver.ts) that manages provider lifecycle, status refresh, and wires into the existing built-in driver registry.
Implements an Ollama adapter (OllamaAdapter.ts) supporting session start/stop, turn execution with streaming, tool call approval flow, interrupt, rollback, and runtime event emission.
Adds tool execution (OllamaTools.ts) for file read/write/edit, directory listing, shell commands, grep search, and web fetch — executed against the local filesystem and shell upon approval.
Adds text generation (OllamaTextGeneration.ts) for commit messages, PR content, branch names, and thread titles using schema-validated JSON output from Ollama.
Adds OllamaSettings to server config contracts and an Ollama icon and settings definition to the web UI.
Risk: tool execution includes shell (bash) and filesystem write access; these are approval-gated but run as the server process with no sandbox.

^{Macroscope summarized 6f18cef.}

Add a full Ollama provider driver integrated with the existing ProviderDriver pattern, enabling T3 Code to use local and cloud Ollama instances as a first-class provider alongside Codex, Claude, OpenCode, and Cursor. ## Backend - — HTTP helpers for Ollama REST API: /api/chat (streaming SSE), /api/tags (model listing), /api/version (health check). Supports Bearer token auth via OLLAMA_API_KEY env variable. - — 7 built-in tools: read_file, write_file, edit_file, bash, list_directory, search_files, web_fetch. Each tool has classify + execute phases matching the tool-calling protocol. - — Full provider adapter with session lifecycle, tool-calling loop, approval flow via Deferred, and fork support (runForkWith). Streams structured runtime events (turn.started, content.delta, item.completed, tool.use, tool.result, etc.). - — Health check provider: pings /api/tags at startup and on 5-minute refresh interval, exposes available models directly from the Ollama server. - — ProviderDriver<OllamaSettings, OllamaDriverEnv> bridging adapter + text generation + snapshot management into a single entry point. - — Commit message, PR content, branch name, and thread title generation via Ollama chat completions. - — OllamaDriver registered in BUILT_IN_DRIVERS array and BuiltInDriversEnv union type. ## Contracts - — OllamaSettings schema: enabled, baseUrl (default http://localhost:11434), model, customModels. Patch schema for runtime updates. - — Ollama driver kind registered in driver model contract for model selection routing. ## Frontend - — OllamaIcon SVG (simple-icons Ollama logo). - — Ollama entry in PROVIDER_CLIENT_DEFINITIONS with icon, label, and settings schema for the settings panel. - — Ollama added to PROVIDER_ICON_BY_PROVIDER mapping. - — Typecheck fix for the added model field. ## Usage Configure Ollama in Settings → Providers: - Server URL: http://localhost:11434 (local) or https://ollama.com - API key via OLLAMA_API_KEY environment variable (for Ollama Cloud) - Models are auto-discovered from /api/tags

coderabbitai · 2026-05-16T20:12:34Z

Important

Review skipped

Auto reviews are disabled on this repository. Please check the settings in the CodeRabbit UI or the .coderabbit.yaml file in this repository. To trigger a single review, invoke the @coderabbitai review command.

⚙️ Run configuration

Configuration used: Repository UI

Review profile: CHILL

Plan: Pro

Run ID: 6f13ecd8-24e2-41d7-b8b5-c342db6fd3da

You can disable this status message by setting the reviews.review_status to false in the CodeRabbit configuration file.

Use the checkbox below for a quick retry:

🔍 Trigger review

✨ Finishing Touches

🧪 Generate unit tests (beta)

Create PR with unit tests

Tip

💬 Introducing Slack Agent: The best way for teams to turn conversations into code.

Slack Agent is built on CodeRabbit's deep understanding of your code, so your team can collaborate across the entire SDLC without losing context.

Generate code and open pull requests
Plan features and break down work
Investigate incidents and troubleshoot customer tickets together
Automate recurring tasks and respond to alerts with triggers
Summarize progress and report instantly

Built for teams:

Shared memory across your entire org—no repeating context
Per-thread sandboxes to safely plan and execute work
Governance built-in—scoped access, auditability, and budget controls

One agent for your entire SDLC. Right inside Slack.

👉 Get started

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.

❤️ Share

_{Comment @coderabbitai help to get the list of available commands and usage tips.}

macroscopeapp · 2026-05-16T20:22:07Z

Approvability

Verdict: Needs human review

This PR introduces a complete new Ollama provider with session management, tool execution (including shell commands and file operations), and LLM integration — substantial new user-facing capability that warrants human review. An unresolved medium-severity comment also identifies inconsistent model selection handling.

^{You can customize Macroscope's approvability policy. Learn more.}

…le, rollback indices ## What changed - **ollamaRuntime.ts**: `Effect.promise` → `Effect.tryPromise` (4 call sites). Network errors are now typed `OllamaRuntimeError` failures instead of unrecoverable defects. `TextDecoder` now flushes on the final chunk to prevent UTF-8 truncation at multi-byte boundaries. - **OllamaProvider.ts**: `checkOllamaProviderStatus` receives `processEnv` per-instance instead of reading `process.env.OLLAMA_API_KEY` globally. - **OllamaDriver.ts**: passes `processEnv` from `mergeProviderInstanceEnvironment` to `checkOllamaProviderStatus` (was missing before — provider status check was broken when OLLAMA_API_KEY came from instance env rather than global env). - **OllamaAdapter.ts**: tracks `Fiber` from `runFork` in `context.activeFiber`. Finalizer and `stopSession` now interrupt the fiber — no more leaks on session shutdown. `rollbackThread` uses per-turn message indices instead of `numTurns*2`, correctly handling tool-call turns. - **OllamaTextGeneration.ts**: default fallback model aligned to `llama3.2` (matches `contracts/DEFAULT_MODEL_BY_PROVIDER`). ## Why The Ollama provider runtime was built on `Effect.promise` which converts errors to `die()` defects (unrecoverable). Network flakiness, timeouts, or DNS failures would crash the Effect fiber instead of being catchable `fail()` errors. Each call site has been migrated to `Effect.tryPromise` so all network-layer errors flow through the standard error channel. The session lifecycle had two leaks: (1) fibers not interrupted on `stopSession` or finalizer (would keep running after session teardown), and (2) `rollbackThread` using `numTurns*2` which breaks when turns contain tool-call exchanges (odd number of messages). Both fixed. The `processEnv` isolation bug meant `checkOllamaProviderStatus` read the global environment, so instances configured via workspace env (not global) would fail provider health checks. ## Checklist - [x] This PR is small and focused - [x] I explained what changed and why - [N/A] I included before/after screenshots for any UI changes - [N/A] I included a video for animation/interaction changes

- interruptTurn no longer permanently closes the session: sendTurn now clears the stopped flag instead of rejecting with SessionClosedError - rollbackThread works: turnMessageIndices is populated at each turn start, so whole turns (incl. tool messages) splice correctly - default model unified to qwen2.5:7b across contracts, adapter and text generation (was llama3.2 in contracts only) - ollamaVersion probe wrapped in Effect.exit so an unreachable server no longer crashes the whole status check - ollamaListModels reads modified_at (snake_case) from /api/tags - OllamaTextGeneration catches OllamaRuntimeError by tag instead of a blanket mapError, avoiding double-wrapped TextGenerationError Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

cursor · 2026-05-16T21:12:22Z

+        if (!text || text.length === 0) {
+          return yield* new ProviderAdapterValidationError({ provider: PROVIDER, operation: "sendTurn", issue: "Ollama turns require text input." });
+        }
+        const model = input.modelSelection?.model ?? context.activeModel;


Empty model string not handled in sendTurn

Medium Severity

The model resolution in sendTurn uses nullish coalescing (??), which won't fall back to context.activeModel when input.modelSelection?.model is an empty string "". By contrast, startSession correctly checks trim().length > 0 before using the provided model. If the client sends a modelSelection with an empty model field, sendTurn will pass "" as the model name to ollamaChat, causing the API call to fail.

Additional Locations (1)

apps/server/src/provider/Layers/OllamaAdapter.ts#L104-L108

^{Reviewed by Cursor Bugbot for commit 6c74fed. Configure here.}

cursor · 2026-05-16T21:12:22Z

+  session: ProviderSession;
+  readonly threadId: ThreadId;
+  readonly messages: OllamaChatMessage[];
+  readonly runtimeEvents: Queue.Queue<ProviderRuntimeEvent>;


Unused runtimeEvents field stored in session context

Low Severity

The OllamaSessionContext interface declares a runtimeEvents field (line 44) and each session stores a reference to the shared queue (line 125), but this field is never accessed through the session context. The emit function at line 98 captures the outer runtimeEvents variable directly via closure. This is dead storage that adds confusion about whether events are per-session or per-adapter.

^{Reviewed by Cursor Bugbot for commit 6c74fed. Configure here.}

Ollama Cloud's /api/version returns {"version":"0.0.0"} as a placeholder. Treat it as no version so the provider list shows no version chip instead of a meaningless "v0.0.0". Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

cursor

Cursor Bugbot has reviewed your changes and found 1 potential issue.

There are 4 total unresolved issues (including 3 from previous reviews).

^{❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.}

^{Reviewed by Cursor Bugbot for commit 6217bd6. Configure here.}

interruptTurn only flipped the stopped flag; if the turn fiber was mid-ollamaChat HTTP call it would not observe the flag before the next sendTurn reset it to false and forked a second fiber, so two fibers mutated context.messages concurrently. - sendTurn now interrupts and awaits any prior activeFiber before resetting the flag and forking, guaranteeing a single live fiber. - interruptTurn interrupts the fiber instead of only setting the flag. - turn.completed is emitted from a single Effect.onExit handler, so it fires exactly once whether the turn completes, fails, or is interrupted (previously skipped when the fiber was interrupted). Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

github-actions Bot added vouch:unvouched PR author is not yet trusted in the VOUCHED list. size:XXL 1,000+ changed lines (additions + deletions). labels May 16, 2026

cursor Bot reviewed May 16, 2026

View reviewed changes

Comment thread apps/server/src/provider/ollamaRuntime.ts Outdated

Comment thread apps/server/src/provider/Layers/OllamaAdapter.ts

Comment thread apps/server/src/textGeneration/OllamaTextGeneration.ts

macroscopeapp Bot reviewed May 16, 2026

View reviewed changes

cursor Bot reviewed May 16, 2026

View reviewed changes

Comment thread apps/server/src/provider/Layers/OllamaAdapter.ts

Comment thread apps/server/src/provider/Layers/OllamaAdapter.ts

3L0935 force-pushed the feat/ollama-provider branch from 2d364b8 to b4ede4e Compare May 16, 2026 20:42

3L0935 force-pushed the feat/ollama-provider branch from b4ede4e to 21acf51 Compare May 16, 2026 20:46

macroscopeapp Bot reviewed May 16, 2026

View reviewed changes

Comment thread apps/server/src/provider/ollamaRuntime.ts Outdated

Comment thread apps/server/src/provider/Layers/OllamaProvider.ts

cursor Bot reviewed May 16, 2026

View reviewed changes

Comment thread packages/contracts/src/model.ts Outdated

Comment thread apps/server/src/provider/Layers/OllamaProvider.ts Outdated

cursor Bot reviewed May 16, 2026

View reviewed changes

Comment thread apps/server/src/provider/Layers/OllamaAdapter.ts

3L0935 and others added 2 commits May 16, 2026 23:33

docs(ollama): note the turnError single-fallible-step invariant

6f18cef

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

Conversation

3L0935 commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

What changed

Why

Screenshots

Checklist

Add Ollama as a local and cloud LLM provider runtime

Uh oh!

coderabbitai Bot commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Review skipped

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

macroscopeapp Bot commented May 16, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Approvability

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

cursor Bot May 16, 2026

Choose a reason for hiding this comment

Empty model string not handled in sendTurn

Uh oh!

Uh oh!

cursor Bot May 16, 2026

Choose a reason for hiding this comment

Unused runtimeEvents field stored in session context

Uh oh!

cursor Bot left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

3L0935 commented May 16, 2026 •

edited

Loading

coderabbitai Bot commented May 16, 2026 •

edited

Loading

macroscopeapp Bot commented May 16, 2026 •

edited

Loading