Skip to main content
Glama

devin_request_async

Destructive

Run Devin CLI requests as durable background jobs, poll for status, and collect the final result.

Instructions

Start a Cognition Devin CLI request as a durable background job. Poll with llm_job_status, collect with llm_job_result.

Input Schema

TableJSON Schema
NameRequiredDescriptionDefault
modelNoModel name or alias (e.g. opus, latest)
configNoConfig file path (Devin --config <PATH>)
promptNoPrompt text for Devin CLI. Required in practice; promptFile is additive (loads an initial prompt from a file).
sandboxNoRun the Devin session in a sandbox (Devin --sandbox). Safety control: never defaulted on; pass true to opt in.
worktreeNoSlice λ: run this request inside a dedicated git worktree owned by the gateway. `true` creates a fresh worktree at `<repoRoot>/.worktrees/<uuid>` branched from HEAD. `{ name?, ref? }` lets the caller supply a sanitized name and/or a git ref (default: HEAD). When the request carries a sessionId and the session already has a worktree, reuse requires same-host ownership metadata and a matching live Git registration. Named path collisions never reuse manager state. Gateway-managed worktrees work with file-backed and PostgreSQL session managers. Shared PostgreSQL rows do not transfer filesystem ownership; reuse and cleanup remain limited to the owning host. The Grok, Devin, and Mistral adapters require an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests fail closed because they cannot durably reselect the worktree. The request must select a registered workspace explicitly, through caller-owned session metadata, or through the configured default. Worktrees never fall back to the gateway process cwd and cannot be combined with local workingDir, addDir, or includeDirs paths. The gateway spawns the child CLI with `cwd: <worktree-path>`; no `-w`/`--worktree` flag is ever emitted to the underlying CLI. On worktree materialization, the gateway suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. On session_delete, session_clear_all or file-backed TTL eviction, both managers hide a durably owned worktree session while cleanup runs. Failed Git removal retains a cleanup-pending tombstone, blocks reuse, and is retried when a manager is registered on the owning host; the record is finalized only once Git no longer registers the worktree, which is read back rather than inferred from the recorded path being absent. A worktree that moved, or whose owner marker cannot be read, is not a removal. Deletion processed by another host removes no worktree and leaves the owning host's record intact. The database-side cleanup_expired_sessions function stages the same tombstone instead of deleting a worktree-bearing session; it invokes no gateway observer and so attempts no removal itself. A tombstone is not bounded by retention. Successful responses are prefixed with `[gateway] worktree=<absolute-path>\n` so callers can use the path. For Claude approvalStrategy:mcp_managed, requesting or reusing a worktree requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1. Other adapters reject mcp_managed before launch. NOTE: callers should `.gitignore` the `.worktrees/` directory in their repo (the gateway does NOT auto-gitignore; see slice λ spec Q4).
sessionIdNoDevin session ID to resume (--resume <id>; use resumeLatest for --continue). Note: the gw-* id minted for a brand-new session is not resumable via sessionId; continue with resumeLatest:true.
workspaceNoRegistered workspace alias for remote HTTP/OAuth provider calls. Do not use this field, workspace_list, or workspace_register_existing_repo as a fallback for stdio/local provider path access; pass workingDir/addDir/includeDirs directly instead.
promptFileNoLoad the initial prompt from a file (--prompt-file)
workingDirNoLocal Devin process working directory. Stdio/local callers may pass local paths directly. Remote HTTP/OAuth callers must use relative paths inside a selected registered workspace. Do not call workspace_* tools to fix stdio/local provider path access.
agentConfigNoAgent config file path (Devin --agent-config <FILE>). devin 3000.4.16 and newer no longer advertise this flag and will reject it; the gateway passes it through rather than refusing it, because older devin installs still accept it and the binary is the authority on its own flags.
forceRefreshNoBypass dedup and force a fresh CLI run even if a recent identical request exists
resumeLatestNoResume the most recent Devin session in cwd (--continue). Note: the gw-* id minted for a brand-new session is not resumable via sessionId; continue with resumeLatest:true.
correlationIdNoRequest trace ID (auto if omitted)
exportSessionNoExport the conversation (Devin --export [<PATH>]). ON BY DEFAULT to a gateway-owned path under ~/.llm-cli-gateway/devin-transcripts: devin writes only its final text to stdout, so without the export a run leaves no reconstructable record. Pass false to opt out, true for a bare --export (devin picks the path), or a string for your own path.
idleTimeoutMsNoTotal-runtime bound, not an idle timer: this provider emits no output until it exits, so the child is killed after this duration even while healthy. Min 30s, max 1h, omit=1h default.
providerFlagsNoFlags passed to the provider binary verbatim, keyed exactly as the binary spells them (e.g. {"--best-of-n": "3", "--verbatim": true, "--rules": ["a", "b"]}). Use this for any flag your installed CLI accepts that this schema does not name: the binary decides what it supports, not the gateway. true emits the flag alone; a list REPEATS the flag once per item (pass a joined string if your CLI wants a comma-separated value). Values may not start with '-', and a flag the gateway is already emitting for this request is refused rather than duplicated. LOCAL stdio callers only: remote HTTP/OAuth callers are refused every flag here and should use this tool's declared parameters, which carry their own host-path and approval gates.
approvalPolicyNoapprovalPolicy is unavailable for this provider. It has no effect with legacy, and mcp_managed is rejected before launch because ambient MCP configuration cannot be isolated.
optimizePromptNoOptimize prompt before execution
permissionModeNoDevin CLI permission mode (--permission-mode). auto, accept-edits, smart, or dangerous. When omitted, Devin uses its own headless default; choose an explicit mode for unattended runs.
approvalStrategyNoApproval strategy: legacy is supported. mcp_managed is rejected before launch because this adapter cannot isolate ambient MCP configuration.legacy
compressResponseNoCompress the response display text when collected via llm_job_result (native compressor; default: [compression].enabled).
createNewSessionNoForce a new session
respectWorkspaceTrustNoRespect workspace trust (Devin --respect-workspace-trust <bool>). Devin defaults true for interactive and false for print mode; set explicitly to override.

Schema Changelog

Changes observed during successful MCP inspections.

  1. Changed1 schema field changedv3.2.1
    • changedInput schema / properties / worktree / description
      Previous value: -"Slice λ: run this request inside a dedicated git worktree owned by the gateway. `true` creates a fresh worktree at `<repoRoot>/.worktrees/<uuid>` branched from HEAD. `{ name?, ref? }` lets the caller supply a sanitized name and/or a git ref (default: HEAD). When the request carries a sessionId and the session already has a worktree, reuse requires same-host ownership metadata and a matching live Git registration. Named path collisions never reuse manager state. Gateway-managed worktrees require the local file-backed session manager and fail closed with PostgreSQL sessions. The Grok, Devin, and Mistral adapters require an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests fail closed because they cannot durably reselect the worktree. The request must select a registered workspace explicitly, through caller-owned session metadata, or through the configured default. Worktrees never fall back to the gateway process cwd and cannot be combined with local workingDir, addDir, or includeDirs paths. The gateway spawns the child CLI with `cwd: <worktree-path>`; no `-w`/`--worktree` flag is ever emitted to the underlying CLI. On worktree materialization, the gateway suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. On session_delete or TTL eviction the gateway hides the session and runs `git worktree remove --force`. Failed removal retains a durable cleanup-pending tombstone which blocks reuse and is retried when the file store is registered on the owning host. The tombstone is finalized only after verified Git removal. Successful responses are prefixed with `[gateway] worktree=<absolute-path>\\n` so callers can use the path. For Claude approvalStrategy:mcp_managed, requesting or reusing a worktree requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1. Other adapters reject mcp_managed before launch. NOTE: callers should `.gitignore` the `.worktrees/` directory in their repo (the gateway does NOT auto-gitignore; see slice λ spec Q4)."New value: +"Slice λ: run this request inside a dedicated git worktree owned by the gateway. `true` creates a fresh worktree at `<repoRoot>/.worktrees/<uuid>` branched from HEAD. `{ name?, ref? }` lets the caller supply a sanitized name and/or a git ref (default: HEAD). When the request carries a sessionId and the session already has a worktree, reuse requires same-host ownership metadata and a matching live Git registration. Named path collisions never reuse manager state. Gateway-managed worktrees work with file-backed and PostgreSQL session managers. Shared PostgreSQL rows do not transfer filesystem ownership; reuse and cleanup remain limited to the owning host. The Grok, Devin, and Mistral adapters require an explicit provider-native sessionId; fresh, createNewSession, and resumeLatest-only worktree requests fail closed because they cannot durably reselect the worktree. The request must select a registered workspace explicitly, through caller-owned session metadata, or through the configured default. Worktrees never fall back to the gateway process cwd and cannot be combined with local workingDir, addDir, or includeDirs paths. The gateway spawns the child CLI with `cwd: <worktree-path>`; no `-w`/`--worktree` flag is ever emitted to the underlying CLI. On worktree materialization, the gateway suppresses repository, system, and global Git hooks and configured clean, smudge, and process checkout filters, sparse checkout, and lazy object fetching. Filter-dependent content such as Git LFS remains in its repository representation instead of executing host commands. On session_delete, session_clear_all or file-backed TTL eviction, both managers hide a durably owned worktree session while cleanup runs. Failed Git removal retains a cleanup-pending tombstone, blocks reuse, and is retried when a manager is registered on the owning host; the record is finalized only once Git no longer registers the worktree, which is read back rather than inferred from the recorded path being absent. A worktree that moved, or whose owner marker cannot be read, is not a removal. Deletion processed by another host removes no worktree and leaves the owning host's record intact. The database-side cleanup_expired_sessions function stages the same tombstone instead of deleting a worktree-bearing session; it invokes no gateway observer and so attempts no removal itself. A tombstone is not bounded by retention. Successful responses are prefixed with `[gateway] worktree=<absolute-path>\\n` so callers can use the path. For Claude approvalStrategy:mcp_managed, requesting or reusing a worktree requires approval and LLM_GATEWAY_APPROVAL_ALLOW_BYPASS=1. Other adapters reject mcp_managed before launch. NOTE: callers should `.gitignore` the `.worktrees/` directory in their repo (the gateway does NOT auto-gitignore; see slice λ spec Q4)."
  2. Addedv3.2.0
  3. Removedv3.0.0
  4. First observedv2.16.0

TDQS

A4/5.0
Behavior4/5

Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?

Annotations already signal destructiveHint=true and readOnlyHint=false, so the mutation risk is covered. The description adds substantial behavioral value beyond annotations: durable background job semantics, polling/collection workflow, worktree ownership and cleanup tombstones, child process killed after idleTimeoutMs, approvalStrategy mcp_managed rejected, remote HTTP/OAuth callers refused for providerFlags, worktree stdout prefix, and export on by default. Some behaviors are buried in the worktree and export parameters, but the description is remarkably transparent. There is no contradiction with annotations.

Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.

Conciseness3/5

Is the description appropriately sized, front-loaded, and free of redundancy?

The tool description itself is a single front-loaded sentence with clear lifecycle guidance: "Start a Cognition Devin CLI request as a durable background job. Poll with llm_job_status, collect with llm_job_result." That's concise and structured. However, the overall definition includes an extremely long worktree parameter description that is dense and hard to parse; while it is located in the schema, its length affects the usability of the definition as a whole, making 3 rather than 5 appropriate.

Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.

Completeness4/5

Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?

Given 23 parameters, 100% schema coverage, no output schema, and complex destructive behavior, the description plus schema cover startup, polling, result collection, worktree edge cases, concurrency implications, and provider-specific restrictions well. It lacks an explicit return-value/output shape description (since there is no output schema), but the description does route callers to llm_job_result for collection. The annotation destructiveHint=true plus the child-kill timeout and approval rejection details make this adequate for a high-risk async tool. Missing a bit on exact response format of the job launch itself, hence 4.

Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.

Parameters3/5

Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?

Schema description coverage is 100%, so the baseline is 3. The description of the tool itself adds almost no parameter semantics beyond the schema; instead the schema fields carry rich descriptions. The tool description's main added value is the job lifecycle routing (llm_job_status/llm_job_result) rather than adding semantics to parameters. Therefore a 3 is appropriate: the schema does the heavy lifting, and the description correctly doesn't duplicate it.

Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.

Purpose5/5

Does the description clearly state what the tool does and how it differs from similar tools?

Description starts with a specific verb and resource: "Start a Cognition Devin CLI request as a durable background job." It clearly distinguishes this from the synchronous devin_request sibling and gives the polling/collection flow: "Poll with llm_job_status, collect with llm_job_result." The name, title, and description all align.

Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.

Usage Guidelines4/5

Does the description explain when to use this tool, when not to, or what alternatives exist?

The description tells the agent when to use this vs. alternatives: async background work, poll with llm_job_status, collect with llm_job_result. Sibling tools llm_job_status and llm_job_result are named explicitly. However, it doesn't explicitly say 'use devin_request instead when you need synchronous behavior' or list exclusions like when NOT to use async, so it earns 4 rather than 5.

Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.