@devries-cameron20 / composio Mirror · read-only
Fork 0

Composio powers 1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench to help you build AI agents that turn intent into action.

Mirrored from https://github.com/Camo67/composio.git

Synced Nov 21, 06:00 UTC

next
md 5.62 KB

Verification Map

The maintained argv and working directories are in scripts/agent-profiles.json. Run from the repository root with its pinned mise toolchain, Bun and the normal or Python development environment prepared. Profiles run existing checks and their normal environment setup (including nox environment creation); they do not choose a model.

The CLI and TypeScript SDK profiles use package-scoped Turbo tasks so prerequisite workspace builds run before tests and type checks, including in a fresh checkout.

Profile Existing checks What remains outside its proof
workflow Bun task integration tests, their TypeScript check and both agent-guidance validators Agent judgment, product behavior, hosted CI
cli CLI package tests, boundary validators and TypeScript checks Built-binary behavior, real backend behavior
sdk-typescript Core SDK unit tests and type checks Changed providers, runtime packaging, live API behavior
sdk-python Existing make chk and make tst nox sessions in python/ Built artifacts, live API behavior, providers outside those sessions
release-metadata Existing release-workflow checks Hosted checks, merges, tags, registries and clean consumer installation

Use the package skills for narrower tests or additional affected providers. For packaging or runtime changes, use existing Node/Cloudflare or CLI Docker E2E checks after reading typescript-testing or cli-e2e. For publication, follow cli-release or python-release; a local receipt never establishes publication.

Carry a task between sessions

The TypeScript CLI runs with Bun and uses effect/cli for argument validation and generated help. Start with bun run agent:task --help or bun run agent:task init --help to inspect commands and required flags. Task commands print JSON; verification streams check output to stderr.

bun run agent:task init cli-output --goal "Reproduce the reported CLI output bug and verify the fix" --profile cli --source https://github.com/ComposioHQ/composio/issues/123 --authority "Implement locally; publication requires the current request"
bun run agent:task status cli-output
bun run agent:task verify cli-output
bun run agent:task list

The id names .artifacts/agent-tasks/<id>/task.json. Update its nextStep and blocker fields as work progresses. Keep the goal observable; record the real source issue or PR URL when one exists. The authority field preserves context; it is never permission to exceed the current user request or repository rules.

Receipts retain the task, profile, source-content fingerprint and exact HEAD at verification time. A changed task acceptance contract, tracked or untracked source file, or profile makes previous passing evidence stale. Ignored artifacts and Claude Code's .claude/scheduled_tasks.lock are excluded, so a receipt does not cover ignored binaries or environment state changing later. Fresh external observations are still required for production and registry claims. Changing only nextStep or blocker preserves verification evidence.

Each attempt has its own receipt and a private output log capped at 64 KiB. Longer output keeps its first 16 KiB and last 48 KiB around an omission marker, so final errors survive. The log is saved while commands run, so it keeps recent output when the runner is killed. Attempts are numbered in order, not by clock time. Failed attempts survive subsequent success. Missing tools, failed commands, interruption and source edits during verification cannot produce a passing attempt. Remaining commands after the first failure were not run. An abruptly killed runner reads as interrupted, including when its latest attempt has no finished receipt. A retry recovers a dead runner's lock once its old command group is gone. If that group remains alive, or the runner died before saving the command's process group, the error identifies the command and receipt to inspect before retrying. Live-runner locks are never recovered automatically. An empty or corrupt verify.lock is replaced only when no task has a running attempt whose runner or command group is alive. A command that exits while workers it started are still running does not end its step: the runner stops those workers before the next command or lock release.

Verification is serialized per worktree because profiles share build and nox outputs. Use separate worktrees for parallel tasks. An interrupted lock acquisition can leave .artifacts/agent-locks/acquire.lock; inspect its PID and remove it only after confirming that process has ended.

Learn from interventions

test/agent-task.test.ts turns concrete workflow failure classes into runnable checks: context survives resumption; edits invalidate proof; failures preserve evidence; missing tools fail rather than skip. These exercise the tool's behavior, not an LLM's planning quality. For a new recurring mistake, add the smallest behavioral check at the real failure boundary and demonstrate its failure before the fix. Keep unrelated passing cases as retention checks. Assess judgment-heavy changes on new tasks before claiming they reduce supervision.