feat(providers): add Devin CLI provider via ACP - #11239
Closed
adampeterhiggins wants to merge 7 commits into
Closed
adampeterhiggins wants to merge 7 commits into
adampeterhiggins wants to merge 7 commits into
Conversation
Add a first-class Devin provider so T3 Code can run Devin's SWE models through the local `devin` CLI's ACP server. - Contracts: `DevinSettings` (enabled, binaryPath, customModels), provider kind, "adaptive" default model, "swe-1-6-fast" text-generation default. - `DevinAcpSupport` builds `devin --permission-mode <mode> acp` spawn args, maps T3 runtime modes onto Devin's session modes (accept-edits, smart, ask, plan, bypass), and applies model selection through the session's `model` config option. - `DevinAdapter` runs sessions over `AcpSessionRuntime`: prompt streaming, steering, interrupt/stop, permission requests via `request.opened`, and `session/elicitation` mapped to user-input questions (form mode only; URL-mode elicitations are cancelled). - `DevinProvider` probes `devin --version` / `auth status` and discovers models via `devin models list --format json` with a static SWE fallback catalog; `devin update` provides maintenance. - `DevinDriver` registers the provider (multi-instance capable) and `DevinTextGeneration` reuses the ACP runtime in `ask` mode for commit messages, PR content, branch names, and thread titles. - Web/mobile: provider icon, driver metadata, settings search, composer draft persistence, and status ordering. Docs: providers-devin page. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The "Your agents" step only showed Claude Code and Codex. Add Devin as a primary agent card with its install command (cli.devin.ai install.sh / setup.ps1) and `devin auth login` for the sign-in terminal. Since Devin is opt-in, a disabled card now offers an Enable button that flips the providerInstances envelope instead of a dead "Disabled" label. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Replace the placeholder "D" letterform with the official Devin mark (three interlocking hexagons, 24x24, currentColor) sourced from the published vector used by @lobehub/icons. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The picker's display-name helper strips a leading subProvider qualifier
from the model name. Devin variant labels are already fully qualified
("Inkling High", "SWE-2 Max"), so tagging them with the family slug made
rows render as just the effort suffix ("High") with "Devin · inkling" as
the subtitle. Drop subProvider from the catalog rows.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
… options Devin encodes effort, speed, and context in the model uid itself (`claude-opus-5-high-fast`, `glm-5-2-max-1m`), so the picker showed one row per variant — ~215 entries — instead of one row per model with selectable axes like other providers. `devinModelCatalog` now parses each variant uid right-to-left (context, speed, effort suffixes, `-`/`_` separators) and groups variants by base into a single `ServerProviderModel` per family with `effort`/`speed`/ `context` option descriptors. Bare variants contribute a "Default" effort choice; variant uids are kept as aliases so stored flat selections still resolve. The adapter folds dim selections back into a concrete advertised uid by parsed-dims matching rather than string composition, so irregular uids (`MODEL_GPT_5_2_LOW`, `claude-5-fable-*`) and missing combinations relax instead of failing the turn. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
The registry's snapshot merge retained all previously seen models for every driver except codex/antigravity/opencode, so Devin's old per-variant rows survived each probe alongside the grouped family rows. Give Devin the same rule: a ready probe's discovered list replaces the inventory, while warning/error/disabled/unauthenticated snapshots keep the last good catalog. Generated with [Devin](https://devin.ai) Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Families like gpt-5.1 and claude-sonnet-4.5 carry their effort tiers in
uids that share no base (`MODEL_PRIVATE_12..15`, `MODEL_PRIVATE_2/3`), so
uid-suffix parsing left them as separate rows. Variants in one family are
the same model by definition — when the uids can't prove it, parse the
label remainder ("GPT-5.1 Low Thinking" → low) and emit a single effort
select whose choices carry the concrete uid. The resolver now passes a
dim selection that is itself an advertised model value straight through,
so opaque families resolve without a shared base.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
github-actions
Bot
added
vouch:unvouched
|
Caution Review failedThe pull request is closed. ℹ️ Recent review info⚙️ Run configurationConfiguration used: Path: .coderabbit.yaml Review profile: CHILL Plan: Advanced Run ID: 📒 Files selected for processing (34)
📝 WalkthroughWalkthroughThis PR adds Devin as a built-in provider. It adds settings, model catalogs, ACP sessions, provider discovery, text generation, onboarding support, icons, persistence, tests, and documentation. ChangesDevin provider integration
Estimated code review effort: 5 (Critical) | ~90 minutes Change: Feature Sequence Diagram(s)sequenceDiagram
participant T3Code
participant DevinDriver
participant DevinProvider
participant DevinAdapter
participant DevinCLI
T3Code->>DevinDriver: Create Devin provider instance
DevinDriver->>DevinProvider: Check CLI status and models
DevinProvider->>DevinCLI: Run version, auth, and model commands
DevinProvider-->>DevinDriver: Return provider snapshot
T3Code->>DevinAdapter: Start ACP session
DevinAdapter->>DevinCLI: Spawn devin acp
DevinCLI-->>DevinAdapter: Stream ACP events
DevinAdapter-->>T3Code: Emit turns, content, and approval events
Suggested reviewers: ✨ Finishing Touches🧪 Generate unit tests (beta)
Comment |
![macroscopeapp[bot]](https://avatars.githubusercontent.com/in/900172?s=60&v=4){kind=small}
| for (const selection of input.selections ?? []) { | ||
| if (isDevinModelDimOptionId(selection.id)) continue; | ||
| const option = configOptions.find((candidate) => candidate.id === selection.id); | ||
| if (!option || option.id === "model") continue; |
Contributor
There was a problem hiding this comment.
🟠 High acp/DevinAcpSupport.ts:249
An approval-required turn can run with bypass instead of requiring approval when selections contain { id: "mode", value: "bypass" }. applyDevinAcpModelSelection forwards mode through setConfigOption after spawning with normal, while resolveDevinModeId does not replace an already-writable mode; exclude mode here so the runtime-mode resolver remains its sole owner.
- if (!option || option.id === "model") continue;
+ if (!option || option.id === "model" || option.id === "mode") continue;🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/server/src/provider/acp/DevinAcpSupport.ts around line 249:
An approval-required turn can run with `bypass` instead of requiring approval when selections contain `{ id: "mode", value: "bypass" }`. `applyDevinAcpModelSelection` forwards `mode` through `setConfigOption` after spawning with `normal`, while `resolveDevinModeId` does not replace an already-writable mode; exclude `mode` here so the runtime-mode resolver remains its sole owner.
| ]; | ||
| const { enabled: _legacyEnabled, ...legacyConfigRest } = legacyConfig ?? {}; | ||
| updateSettings({ | ||
| providerInstances: { |
Contributor
There was a problem hiding this comment.
🟡 Medium onboarding/WelcomeWizard.tsx:702
Enabling two disabled agents before the first settings update is reflected causes the later update to replace providerInstances with a stale snapshot and silently remove the earlier enabled envelope. enableProvider spreads settings.providerInstances captured during render, while updateSettings replaces the entire map; merge each entry against the latest settings or serialize these updates so concurrent enables are preserved.
🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/web/src/components/onboarding/WelcomeWizard.tsx around line 702:
Enabling two disabled agents before the first settings update is reflected causes the later update to replace `providerInstances` with a stale snapshot and silently remove the earlier enabled envelope. `enableProvider` spreads `settings.providerInstances` captured during render, while `updateSettings` replaces the entire map; merge each entry against the latest settings or serialize these updates so concurrent enables are preserved.
Comment on lines
+504
to
+522
| if (!wanted.hasAny && input.advertisedValues.includes(model)) return model; | ||
|
|
||
| const wantedBase = normalizeDevinModelId(model); | ||
| const candidates = input.advertisedValues.filter( | ||
| (uid) => parseDevinModelUid(uid).base === wantedBase, | ||
| ); | ||
| if (candidates.length === 0) return model; | ||
|
|
||
| if (!wanted.hasAny) { | ||
| // No dims selected: keep the session's current value when it already | ||
| // belongs to this family, else the bare variant, else first advertised. | ||
| if (input.currentValue && candidates.includes(input.currentValue)) { | ||
| return input.currentValue; | ||
| } | ||
| const bare = candidates.find((uid) => { | ||
| const dims = parseDevinModelUid(uid); | ||
| return dims.effort === undefined && dims.speed === undefined && dims.context === undefined; | ||
| }); | ||
| return bare ?? candidates[0]!; |
Contributor
There was a problem hiding this comment.
🟡 Medium provider/devinModelCatalog.ts:504
When no non-default dimensions are selected, resolveDevinModelUid resets a session from currentValue such as glm-5-2-max to the grouped model slug glm-5-2. The early exact-match return at line 504 prevents the later currentValue preservation branch from running; defer the model pass-through until after checking whether currentValue belongs to the family.
- if (!wanted.hasAny && input.advertisedValues.includes(model)) return model;
-
const wantedBase = normalizeDevinModelId(model);
@@
if (input.currentValue && candidates.includes(input.currentValue)) {
return input.currentValue;
}
+ if (input.advertisedValues.includes(model)) return model;
const bare = candidates.find((uid) => {🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/server/src/provider/devinModelCatalog.ts around lines 504-522:
When no non-default dimensions are selected, `resolveDevinModelUid` resets a session from `currentValue` such as `glm-5-2-max` to the grouped `model` slug `glm-5-2`. The early exact-match return at line 504 prevents the later `currentValue` preservation branch from running; defer the model pass-through until after checking whether `currentValue` belongs to the family.
| }); | ||
|
|
||
| const DEVIN_PLAN_MODE_ALIASES = ["plan", "architect"]; | ||
| const DEVIN_READ_ONLY_MODE_IDS = new Set(["plan", "ask"]); |
Contributor
There was a problem hiding this comment.
🟠 High acp/DevinAcpSupport.ts:121
An approval-required turn remains in read-only architect mode after a plan turn, so normal editing requests cannot proceed. architect is accepted by DEVIN_PLAN_MODE_ALIASES but missing from DEVIN_READ_ONLY_MODE_IDS, preventing the restore path from selecting a writable mode; add it to the read-only IDs.
Suggested change
| const DEVIN_READ_ONLY_MODE_IDS = new Set(["plan", "ask"]); | |
| const DEVIN_READ_ONLY_MODE_IDS = new Set(["plan", "ask", "architect"]); |
🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/server/src/provider/acp/DevinAcpSupport.ts around line 121:
An approval-required turn remains in read-only `architect` mode after a plan turn, so normal editing requests cannot proceed. `architect` is accepted by `DEVIN_PLAN_MODE_ALIASES` but missing from `DEVIN_READ_ONLY_MODE_IDS`, preventing the restore path from selecting a writable mode; add it to the read-only IDs.
| // A sendTurn while a prompt is in flight is a steer: the agent folds | ||
| // the new prompt into the ongoing work, so the active turn id is | ||
| // reused instead of opening a new turn. | ||
| const steeringTurnId = ctx.promptsInFlight > 0 ? ctx.activeTurnId : undefined; |
Contributor
There was a problem hiding this comment.
🟠 High Layers/DevinAdapter.ts:807
A concurrent second sendTurn creates a separate turnId and emits another turn.started, leaving the first turn without a terminal event and attributing streamed output to the wrong turn. ctx.promptsInFlight is incremented before ctx.activeTurnId is assigned, so the second call sees an active prompt but no active ID; reserve ctx.activeTurnId immediately after choosing turnId.
🤖 Copy this AI Prompt to have your agent fix this:
In file @apps/server/src/provider/Layers/DevinAdapter.ts around line 807:
A concurrent second `sendTurn` creates a separate `turnId` and emits another `turn.started`, leaving the first turn without a terminal event and attributing streamed output to the wrong turn. `ctx.promptsInFlight` is incremented before `ctx.activeTurnId` is assigned, so the second call sees an active prompt but no active ID; reserve `ctx.activeTurnId` immediately after choosing `turnId`.
Contributor
ApprovabilityVerdict: Not approved Macroscope's review found this PR not approvable — This 4,038-line change introduces a new production ACP provider with external process execution, authentication, permissions, sessions, model discovery, onboarding, and text generation, along with new product defaults. It also adds a static-analysis suppression and retains unresolved Medium/High findings involving permission enforcement, concurrency, model selection, and settings updates. Not approved because:
Adjust the Minimum Blocking Severity for this repo — including turning it Off — in Settings. You can add or adjust custom eligibility rules. Learn more. |
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
devin acpover Agent Client Protocol, same architecture as the Cursor provider): sessions, streaming, steering, interrupt/stop, permission requests, elicitation handling, and text generation.claude-opus-5-high-fast,glm-5-2-max-1m,MODEL_GPT_5_2_LOW), sodevinModelCataloggroupsdevin models list --format jsoninto one picker row per family witheffort/speed/contextoption descriptors. Families whose variant uids share no base (MODEL_PRIVATE_*) are grouped via label-derived dims, with the concrete uid carried in the option choice. At dispatch the adapter resolves{ model, effort, speed, context }back to a concrete advertised uid by parsed-dims matching — irregular uids and missing combinations relax instead of failing the turn.providers.devin, opt-in like Cursor/Grok), onboarding wizard entry with install/sign-in actions, official Devin mark on web and mobile, docs page.Details
devin --model <uid> --permission-mode <mode> acp; mode picks the closest advertised ACP mode (accept-edits / smart / ask / plan / bypass).devin --version+devin auth status; ACP advertisedevin-browserauth.session/request_permissionmaps torequest.openedwith risk classification;session/elicitationmaps touser-input.requested(URL-mode elicitations are cancelled — chat can't answer them).askmode withswe-2family default.swe-2-high) resolve to grouped rows via aliases.Test plan
apps/serverfocused tests: catalog grouping/resolution, ACP support (spawn args, mode mapping, model selection), adapter prompt flow, provider snapshot/auth parsing — 39+ testsProviderRegistrymerge tests for stale Devin rowsdevin acp: session start, model apply, streaming deltas, permission promptGenerated with Devin
Summary by CodeRabbit
New Features
Documentation