feat(agent): add Devin (Cognition AI) as the 12th provider, plus 2 ACP shared-transport fixes

[mskutlu]

Summary

Adds Devin (Cognition AI's terminal coding agent) as Multica's 12th provider, plus two related fixes that surfaced while building and exercising the integration.

Devin's ACP server (devin acp) speaks the same protocol Hermes / Kimi / Kiro already use, so the new backend is a thin shim around the existing hermesClient transport. The two fixes harden the shared ACP path for any future provider — they're not Devin-specific.

Commits

1. feat(agent): add Devin (Cognition AI) as the 12th provider

• New server/pkg/agent/devin.go (~340 LOC) modeled on kiro.go. Spawns devin acp and reuses the shared hermesClient for initialize / session/{new,load} / session/prompt / auto-approval of session/request_permission.
• Devin does not implement session/set_model (model selection lives in Devin's own config). Rather than calling and failing, the backend logs a warning when opts.Model is set and continues with Devin's default. ModelSelectionSupported("devin") returns false so the runtime UI hides the model dropdown for Devin runtimes.
• agent.New switch + launchHeaders updated; daemon config.go auto-detects devin on PATH (configurable via MULTICA_DEVIN_PATH / MULTICA_DEVIN_MODEL).
• Skill / context surfaces wired:
  • Project-level skills: .devin/skills/ (matches Devin for Terminal's native convention)
  • User-level skills: ~/.config/devin/skills/
  • Shares the AGENTS.md context-injection branch with the other agents-spec providers.
• Frontend: DevinLogo SVG + case "devin" in provider-logo.tsx; "devin" added to the ProviderName union in step-welcome.tsx.
• Tests: devin_test.go covers New() returns *devinBackend, tool-name normalization, acp spawn, custom-args filtering, prompt → output round-trip, the Model option warning path, and session/load resume. agent_test and models_test extended to lock LaunchHeader coverage and ModelSelectionSupported("devin") = false.
• Docs (English + Chinese): README, providers.mdx, index.mdx, daemon-runtimes.mdx, how-multica-works.mdx, cloud-quickstart.mdx, CLI_INSTALL.md, CLI_AND_DAEMON.md, SELF_HOSTING.md — all bumped 11 → 12 with a Devin section + the MULTICA_DEVIN_PATH / MULTICA_DEVIN_MODEL env vars.

2. fix(api): demote cold-start 401 noise to warn / info

• packages/core/api/client.ts already special-cases 404 → warn. Add 401 to that list. The thrown ApiError still propagates to the caller, so callers that consider a particular 401 fatal can still react; only the log level on the response itself changes.
• packages/core/platform/auth-initializer.tsx: introduce logAuthInitFailure that demotes 401 to info ("no active session") and keeps error for everything else (network, 5xx, malformed response). Both cookie-mode and token-mode catch blocks route through it.

Anonymous page loads probe /api/me and /api/workspaces to bootstrap the session; both return 401 by design — they're answering "no active session, please log in," not signalling a real failure. Logging that at error level on every cold start trains developers to ignore real auth breakage.

3. fix(agent): pick ACP permission optionId from agent's actual options

hermesClient.handleAgentRequest hardcoded optionId: "approve_for_session" in its session/request_permission reply. That works for kimi / kiro because both happen to use that exact string — but it's a kimi-flavoured optionId, not the canonical ACP one. Devin's ACP server presents options whose IDs follow the spec's kind names (allow_once / allow_always / reject_once) and treats unknown IDs as a rejection, logging "Tool execution was rejected by the user: Unknown permission option" and aborting the turn without emitting any final assistant text — surfaced to the daemon as a misleading "devin returned empty output" / blocked task.

Fix: parse the request's options array, pick the most permissive accept option by canonical kind (allow_always > allow_once), and echo back whichever optionId the agent itself advertised. Provider-agnostic: kimi / kiro still resolve to approve_for_session (their kind=allow_always entry), Devin resolves to allow_session (its label for the same kind). A defensive fallback to the legacy approve_for_session literal preserves behaviour if a future agent ever emits an empty options array.

New unit tests cover Devin-style optionIds, kimi/kiro-style optionIds, allow_always > allow_once priority, reject-only options, unknown kinds, and empty options. Existing kimi/kiro/hermes tests pass unchanged (regression guard kept the literal approve_for_session assertion — fix still resolves to that for kimi-flavoured option arrays).

Verification

• go build ./..., go vet ./..., gofmt -d on every changed file — all clean.
• go test ./... (full server suite, ~25 packages) — green.
• pnpm typecheck (6 tasks) + pnpm test (33 view test files / 241 tests) — green.
• Live end-to-end against a real devin acp session on macOS:
  • Daemon detected devin on PATH and registered the runtime: Agents: claude, codex, opencode, gemini, devin.
  • The Runtimes page rendered the new "Devin" runtime row with the new DevinLogo icon.
  • Sent the chat prompt "Run pwd and whoami and report the output as a single line" to a Devin agent.
  • Daemon log showed auto-approved agent permission request method=session/request_permission option_id=allow_session (Devin's actual kind=allow_always optionId, not the kimi-flavoured literal).
  • Terminal tool executed; Devin emitted final agent_message_chunk: /Users/.../workdir | msk.
  • Task status flipped from blocked: devin returned empty output (pre-fix) to completed.

Caveats

• Devin's ACP server emits a few forest autosave failed: database disk image is malformed warnings on stderr during sessions. Those are Devin-side issues with their internal session DB and aren't actionable from Multica's side; just noise in the daemon log.
• Devin's agentInfo.name reports affogato (their internal codename) — so any user-facing strings derived from agentInfo (none today, but worth a flag) won't say "Devin".
• Model selection is currently a no-op for Devin per the note above. If Devin adds session/set_model support later, ModelSelectionSupported("devin") should flip to true and devinStaticModels should be replaced with a CLI-discovery function (TODO comment in models.go).

Happy to split into three separate PRs (the two fixes apply standalone) if reviewers prefer.

[vercel]

@mskutlu is attempting to deploy a commit to the IndexLabs Team on Vercel.

A member of the Team first needs to authorize it.

[Copilot]

Pull request overview

Adds Devin (Cognition AI) as a new ACP-based provider and hardens shared ACP transport behavior, while reducing noisy auth logs on expected 401s during cold-start session probing.

Changes:

• Add a new devin agent backend that spawns devin acp, reuses the shared ACP transport, and wires skills + runtime config support.
• Fix ACP auto-approval to select the agent-advertised permission optionId based on canonical kind (prefer allow_always), improving provider portability.
• Demote expected cold-start 401s from error logging to warn/info in the core API client and auth initializer.

Reviewed changes

Copilot reviewed 31 out of 31 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
server/pkg/agent/devin.go	New Devin backend (spawns devin acp, ACP lifecycle, tool-name normalization, model-selection warning behavior).
server/pkg/agent/devin_test.go	Unit tests for Devin backend behaviors (spawn args, prompt round-trip, resume via session/load, no session/set_model).
server/pkg/agent/hermes.go	Shared ACP fix: choose permission approval optionId by advertised kind rather than hardcoding.
server/pkg/agent/hermes_test.go	Regression + unit tests for permission option selection logic (Devin-style and kimi/kiro-style).
server/pkg/agent/models.go	Add Devin static placeholder model list + ModelSelectionSupported("devin") = false.
server/pkg/agent/models_test.go	Tests covering Devin model-selection unsupported and placeholder model defaults.
server/pkg/agent/agent.go	Register devin in backend factory and launch header mappings + docstrings.
server/pkg/agent/agent_test.go	Ensure launch header coverage includes devin.
server/internal/daemon/config.go	Auto-detect devin on PATH; support MULTICA_DEVIN_PATH / MULTICA_DEVIN_MODEL.
server/internal/daemon/execenv/context.go	Add project-level Devin skills path .devin/skills.
server/internal/daemon/local_skills.go	Add user-level Devin skills path ~/.config/devin/skills.
server/internal/daemon/execenv/runtime_config.go	Include devin in AGENTS.md context injection path.
packages/views/runtimes/components/provider-logo.tsx	Add Devin provider logo and routing.
packages/views/onboarding/steps/step-welcome.tsx	Extend provider union type to include devin.
packages/core/api/client.ts	Demote 401 responses to warn logging (while still throwing ApiError).
packages/core/platform/auth-initializer.tsx	Log expected cold-start 401 auth init failures at info instead of error.
apps/docs/content/docs/providers.mdx	Docs matrix updated to include Devin and “12 tools” count.
apps/docs/content/docs/providers.zh.mdx	Chinese docs matrix updated to include Devin and “12 tools” count.
apps/docs/content/docs/index.mdx	Docs homepage updated to list Devin as supported runtime tool (12 total).
apps/docs/content/docs/index.zh.mdx	Chinese docs homepage updated to list Devin as supported runtime tool (12 total).
apps/docs/content/docs/how-multica-works.mdx	Update architecture explanation to include Devin among supported tools.
apps/docs/content/docs/how-multica-works.zh.mdx	Chinese architecture explanation updated to include Devin.
apps/docs/content/docs/daemon-runtimes.mdx	Update daemon startup docs to reference 12 tools incl. Devin.
apps/docs/content/docs/daemon-runtimes.zh.mdx	Chinese daemon startup docs updated to reference 12 tools incl. Devin.
apps/docs/content/docs/cloud-quickstart.mdx	Update quickstart prerequisite tool list to include Devin.
apps/docs/content/docs/cloud-quickstart.zh.mdx	Chinese quickstart prerequisite tool list updated to include Devin.
README.md	Top-level README updated to include Devin in supported tools list and runtime diagram.
README.zh-CN.md	Chinese README updated to include Devin in supported tools list and runtime diagram.
CLI_INSTALL.md	CLI install guide updated to mention Devin in detected agents list.
CLI_AND_DAEMON.md	CLI/daemon guide updated for Devin detection + env var overrides.
SELF_HOSTING.md	Self-hosting prerequisites updated to include Devin CLI.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

In the ACP lifecycle goroutine, msgCh/resCh are deferred to close, but on early returns (e.g. initialize/session/load/session/new failures) there’s no synchronization with the stdout reader goroutine. cmd.Wait() does not guarantee the reader has finished draining stdout, so the reader can still call trySend(msgCh, ...) after the channels are closed, causing a send-on-closed-channel panic. Consider explicitly waiting for readerDone before closing the channels in all exit paths (including early failures), e.g. by moving channel closes to a shared cleanup block that joins the reader goroutine first.

Suggested change

	defer cancel()
	defer close(msgCh)
	defer close(resCh)
	defer func() {
	stdin.Close()
	_ = cmd.Wait()
	defer func() {
	cancel()
	stdin.Close()
	_ = cmd.Wait()
	<-readerDone
	close(msgCh)
	close(resCh)

[Copilot]

The Devin ACP dialect notes still say auto-approval works by replying with a hardcoded "approve_for_session" optionId. hermesClient.handleAgentRequest now selects the agent-advertised optionId by kind, so this comment is no longer accurate (and is the root cause this PR is fixing). Please update the note to describe the new portable behavior (select allow_always/allow_once by kind and echo back the corresponding optionId).

Suggested change

	// entirely by hermesClient.handleAgentRequest replying
	// "approve_for_session" to every session/request_permission request,
	// identical to how kimi / kiro are driven.
	// entirely by hermesClient.handleAgentRequest selecting the
	// agent-advertised allow_always / allow_once option by kind for each
	// session/request_permission request and echoing back the matching
	// optionId.

[Copilot]

The comment says Hermes/Kimi/Kiro route model selection through session/set_model “before each prompt”, but the current implementations set the model once after session creation (and before the first prompt). Consider adjusting the wording to match the actual behavior so future readers don’t assume model switching is performed per-turn.

Suggested change

	// ACP `session/set_model` RPC before each prompt, which means the
	// UI's dropdown choice is carried all the way down to the LLM call.
	// ACP `session/set_model` RPC after session creation and before the
	// first prompt, which means the UI's dropdown choice is carried all
	// the way down to the LLM call.

[mskutlu]

All three review points addressed in 9c71672 — pushed to the same branch.

1. Channel-close race in the ACP lifecycle goroutine

You're right, and the same fragile pattern exists in kiro.go / kimi.go / hermes.go too — fixed all four so the next ACP shim doesn't inherit it. Replaced the four separate defer cancel() / defer close(msgCh) / defer close(resCh) / defer { stdin.Close(); cmd.Wait() } with a single deferred cleanup block whose ordering is documented inline:

defer func() {
    cancel()        // unblock anything bound to runCtx
    stdin.Close()   // ask the agent to exit gracefully
    _ = cmd.Wait()  // wait for the process; stdout closes here
    <-readerDone    // wait for the reader goroutine to fully drain
    close(msgCh)    // safe: the reader was the only msgCh sender, now finished
    close(resCh)    // safe: only this goroutine sends to resCh
}()

The happy paths still call stdin.Close() / cancel() / <-readerDone explicitly before reading shared state — those become idempotent no-ops with the deferred cleanup. Existing tests pass with -race on a 2× run.

In current code the race is largely gated by streamingCurrentTurn (early-return paths happen before that flag flips, so trySend is never invoked), but the pattern is fragile — any future early-return added after streamingCurrentTurn.Store(true), or any notification handler that doesn't gate on it, would make the race live. Defense-in-depth.

2. Stale Devin docstring on auto-approval

Updated the comment block in devin.go to describe the post-#d23e9c3 portable behaviour:

- Devin's `acp` subcommand does NOT accept the root-level
  --permission-mode flag. Daemon-mode auto-approval is handled
  entirely by hermesClient.handleAgentRequest, which inspects the
  `options` array advertised on each session/request_permission
  request and selects the most permissive accept-style optionId
  by canonical ACP `kind` (allow_always > allow_once). For Devin
  that resolves to optionId="allow_session"; kimi / kiro resolve
  to "approve_for_session" — same code path, agent-specific IDs.

3. ModelSelectionSupported "before each prompt" wording

Reworded to match what the code actually does — Hermes / Kimi / Kiro call session/set_model once after session/new and before the session's first (and only) prompt; the model is fixed for the life of the session. Mid-task switching is not wired today.

Verified end-to-end: go build ./..., go vet ./..., gofmt -l clean, go test ./... (full server suite) green, go test -race ./pkg/agent/... green.