acp-extension-devin

Lody's ACP compatibility adapter for the official Devin CLI.

ACP client → acp-extension-devin → official `devin acp`

Devin already speaks ACP, so this is an ACP-to-ACP adapter. It launches the official runtime named by DEVIN_PATH, forwards standard ACP traffic in both directions, and translates the private cognition.ai/* wire into acp-extension-core contracts. The adapter does not bundle, build, or patch Devin, and it does not depend on Lody workspace packages. Devin owns models, tools, sessions, permissions and MCP connections.

Run

Requires Node.js 22.14 or newer and pnpm 10.20.

pnpm install
pnpm build
DEVIN_PATH=/path/to/official/devin node dist/index.js

DEVIN_PATH is required and is used exactly as given; the adapter never searches PATH. It starts $DEVIN_PATH acp with the inherited environment and injects no configuration of its own. Authenticate and choose models through Devin on the execution machine.

stdout carries ACP only; diagnostics and Devin's stderr go to stderr. When the client closes the connection, the adapter ends Devin's stdin, then exits with the runtime's exit code or re-raises its terminating signal.

Runtime contract

runtime-manifest.json pins the official runtime the adapter is built against and every private field it relies on: _meta keys such as cognition.ai/subagent_context, the cognition.ai/subagentSupport client capability, the _cognition.ai/compaction notification method, the sk:: sidekick tool-call prefix, and the reserved agent IDs. Tests assert against these values.

The manifest declares Devin 3000.11.1 as the minimum supported version; from that version Devin accepts stdio, HTTP and SSE MCP servers in session requests. The pinned and end-to-end validated version is 3000.11.3.

Capabilities

The adapter advertises only what it translates. Compaction lifecycle notifications are translated directly, while subagent events are translated after bilateral negotiation. Other ACP traffic passes through.

• Standard ACP from Devin, including sessions, prompts, tool calls, permission requests, modes and configuration options, passes through as the runtime reports it.
• Core subagentEvents v1, described below.
• Core compaction v1, described below.
• Core usage v1, described below.
• MCP servers supplied by the client, forwarded verbatim.

A dedicated Plan Mode translation is not implemented, and the adapter does not advertise it.

Subagent events

The adapter always advertises subagentEvents v1 in agentCapabilities._meta.lody. Only when the client also advertises clientCapabilities._meta.lody.subagentEvents ({ version: 1 }) does it request Devin's private subagent stream with cognition.ai/subagentSupport and translate it; otherwise Devin's native output passes through untouched.

Once negotiated, delegated work becomes independent Core runs on _lody/subagents/event instead of appearing in the root transcript:

• A run_subagent child becomes a run from Devin's explicit started and completed rows. Its text, thoughts, plans and tool steps are attributed to that run, and its context usage is reported as run progress. A nested child records its parent run.
• A Fusion Sidekick handoff becomes a run named Sidekick that carries its tool steps. Devin reports no lifecycle for sidekicks, so the adapter infers when a run starts and ends from the sidekick tool calls and root activity. It never invents a description or summary.
• Permission requests raised by a run's tool calls still go to the client through the root session, with a run-scoped tool-call ID and _meta.lody.subagentRunId / subagentToolCallId identifying the owner.

Run IDs are owned by the adapter and differ from Devin's agent IDs; an agent ID reused after its run has ended starts a new run. Each run's snapshot precedes its content, and content arriving after termination is dropped. Child output is never re-emitted to the root stream, including events the adapter cannot attribute. Runs do not support cancellation or output reads.

session/load and session/resume replay is forwarded unchanged and creates no runs.

Usage accounting

Devin's private cognition.ai/* token counters on usage_update notifications become Core SessionUsageUpdate on _lody/session/usage_update, advertised as agentCapabilities._meta.lody.usage { version: 1 }. The raw update still passes through unchanged.

Each runtime row reports one inference request. Devin's inputTokens includes cache reads AND cache writes (verified on gpt-6-luna: inputTokens = fresh + cachedRead + cachedWrite), so it is split into disjoint Core buckets: inputTokens minus both cache shares becomes inputTokens, reads become cacheReadInputTokens, and cachedWriteTokens becomes cacheCreationInputTokens. SWE-2 rows omit cachedWriteTokens. The request size is the model context window and lands on the aggregate usage only.

Devin emits each request twice — once untagged and once tagged subagent_context — so identical counters deduplicate by signature. Rows for run_subagent children are billed to that child's own accounting identity and are skipped; root-agent and sidekick rows count under the session, the latter under a sidekick model bucket. Per-model modelUsage tracks the ACP model config option, and every counted row produces a delta alongside the cumulative totals under a per-session usageScopeId. Rows replayed during session/load and session/resume create no accounting.

Private cost fields (totalCreditCost, totalAcuCost) and responseDimensions pass through verbatim under their cognition.ai keys; the adapter never maps them into costUSD because the units are unverified.

Compaction

Devin's _cognition.ai/compaction notifications become Core context-compaction activities on standard ACP tool_call and tool_call_update messages. Native started, completed and failed states drive the lifecycle, and native summaries are retained in the activity details.

The native compact command is available through the standard ACP prompt path. Its response can arrive before compaction finishes; the native notifications determine completion. Manual compaction, automatic compaction, and cancellation have been verified on Devin 3000.11.3. Manual compaction has also been verified in the Lody UI.

MCP

mcpServers from session/new and session/load are forwarded to Devin verbatim. Devin supports stdio, Streamable HTTP and SSE servers in these requests, so the adapter needs no MCP-specific translation and never writes an MCP configuration file.

Through the adapter on Devin 3000.11.3:

• stdio, Streamable HTTP and SSE servers each completed real tool calls, and configured HTTP/SSE request headers reached the server;
• after session/load in a new adapter process, a newly supplied HTTP server was available to the loaded session;
• Sidekick and foreground subagents called MCP tools, and those steps were attributed to their runs;
• in Code mode, approving the permission request let the MCP call proceed.

In Lody Nightly 0.103.0-nightly.1, with Devin's global configuration isolated, a session created from the UI called a tool on Lody's built-in HTTP MCP server.

Validation

pnpm check
pnpm build

pnpm check runs type checking, the Vitest suite and Prettier. Tests use synthetic ACP messages and an injected spawnImpl; they never start a real Devin process or contact a model provider. They cover negotiation, run attribution and lifecycle, permission routing, replay handling, the runtime launch command, the manifest pin, compaction lifecycle, summaries and cancellation, and verbatim forwarding of non-empty stdio, HTTP and SSE MCP configurations in session/new and session/load, with and without subagent negotiation.

The end-to-end MCP results above were observed on Devin 3000.11.3. They do not establish behavior on other runtime versions or real-provider quality.