feat(agent): add native Ollama backend with tool-use loop

[sanjay3290]

What

Adds a first-class ollama agent provider alongside claude/codex/opencode/openclaw/hermes/gemini. Models with tools capability (gemma3+, qwen2.5+, llama3.1+) drive the Multica CLI via a single shell tool; chat-only models still work for conversational tasks.

Closes #769.

Why

Issue #769 closed with the env-var workaround (#846) — pointing ANTHROPIC_BASE_URL at Ollama. That works for users who've installed Claude Code, but Ollama-only users still need a supported CLI on their machine before anything works. This ships the real native path.

End-to-end verified

• LOC-3 (in-workdir): gemma4:latest ran a 6-tool-call sequence on a Multica issue — read context, wrote a file, updated status to in_review, posted a summary comment. 29s.
• LOC-5 (out-of-sandbox): same model wrote $HOME/Desktop/multica-ollama-test.txt via absolute-path printf, verified with ls -l, posted output as comment. 4 calls / 20s.

Configuration

MULTICA_OLLAMA_HOST              default http://localhost:11434
MULTICA_OLLAMA_MODEL             required; empty = provider skipped
MULTICA_OLLAMA_SHELL_TIMEOUT     default 5m; per-command cap
MULTICA_OLLAMA_MAX_OUTPUT_BYTES  default 16384; tool-result cap

Daemon probes /api/tags for reachability, /api/version for version, /api/show for tool capability (prints stderr WARNING if the configured model doesn't declare "tools"). All probes are best-effort and do not block daemon startup.

Architecture

• stream:false when tools active — tool_calls arrive on the final chunk anyway; single-shot is simpler and equally responsive at task granularity.
• Single shell tool, executed via bash -c (NOT -lc; login shells re-source /etc/profile and wipe the PATH injection the daemon makes for the multica CLI). Matches Claude's bypassPermissions and Gemini's --yolo.
• tool_call_id threaded on tool-role messages so strict models (qwen2.5, llama3.1) can correlate results to calls.
• MaxTurns cap (default 25) prevents runaway tool loops.
• Context status (DeadlineExceeded/Canceled) takes precedence over side-effect errors when determining Result.Status — matches the pattern claude.go and gemini.go use after PR fix(daemon): correct Gemini backend status on timeout and cancellation #920.

v1 non-goals (deliberate, documented)

• Streaming during tool turns (chat-only non-tool turns still stream the final text as one event)
• Per-task model selection (single configured model wins)
• Session resume (Ollama is stateless; ResumeSessionID ignored with debug log)
• Dynamic /api/tags model discovery
• Granular tools (file_read/file_write/etc. — single shell covers the same ground via standard Unix utilities)

These match the limitations Gemini #755 shipped with.

Test plan

• go test ./pkg/agent/... ./internal/daemon/... — 22 new tests, all pass in 10s
• go vet ./... clean
• gofmt -w applied
• pnpm --filter @multica/views --filter @multica/core typecheck clean
• End-to-end smoke test: gemma4:latest on local Docker stack, 2 live tasks completed (in-workdir file write + out-of-sandbox Desktop write)
• Daemon registers ollama runtime, reports online, appears in /api/runtimes
• Regression tests for: context cancel mid-loop, timeout mid-loop, malformed response (both empty-everywhere and OOM-case {role:assistant, content:"", tool_calls:[]}), MaxTurns cap, shell exit code forwarding, unknown tool, tool_call_id wire propagation, env-var fallback on malformed values, /api/show capability probe (true/false/404/unreachable)

Files changed

• server/pkg/agent/ollama.go — new (609 LOC)
• server/pkg/agent/ollama_test.go — new (712 LOC)
• server/pkg/agent/agent.go — register case "ollama"; DetectVersion now takes provider name to dispatch HTTP probe for HTTP-based backends
• server/internal/daemon/config.go — HTTP probe of MULTICA_OLLAMA_HOST; capability check with WARNING; references agent.DefaultOllamaHost
• server/internal/daemon/execenv/runtime_config.go — ollama added to the AGENTS.md case
• server/internal/daemon/daemon.go — one-line change to pass provider name into DetectVersion
• packages/views/runtimes/components/provider-logo.tsx — OllamaLogo SVG + switch case

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

2 Skipped Deployments

Project	Deployment	Actions	Updated (UTC)
multica-web	Ignored		Apr 15, 2026 1:50pm
multica-web-production	Ignored		Apr 15, 2026 1:50pm