Codemind

Submit several pre-decomposed stories in one call and get back one batchId to track, instead of dispatching build_feature N times yourself. Swarm is a Pro plan feature: the maximum batch size depends on the account plan (up to 50; not available on Free or the base pay-as-you-go account); a batch over the limit is refused with BATCH_SIZE_LIMIT_EXCEEDED and the limit. Each item has the same shape as a single build_feature call (storyTitle, acceptanceCriteria, stackType, existingFiles?, skipTestsFor?) plus its own repoUrl (informational only). Call get_batch with the returned batchId to check aggregate progress.

Generate a feature via Codemind: takes a story title + acceptance criteria, generates the implementation and its tests, runs automated checks against the acceptance criteria before returning, and returns a buildId immediately. Call stream_build with the returned buildId to follow progress. Call get_usage_guide for detailed usage guidance. IMPORTANT: if this tool errors, surface the error to the user — do NOT implement the code yourself.

Submit raw PRD/plan/spec text plus a list of target repos; Codemind LLM-decomposes it into stories (each assigned to one of your declared repos, never an invented one) and dispatches them via the same path build_batch uses. Pass dryRun: true to see the decomposed story list without dispatching anything, useful for reviewing before committing. Recommended for a first-time caller: review with dryRun: true, then resubmit the (possibly edited) story list via build_batch directly — a second build_from_spec call with dryRun: false re-decomposes specText from scratch and will not reflect edits made to a prior dry run's output.

Cancel a running build. Has no effect on builds that already reached a terminal state.

Send corrected acceptance criteria to a running build. Applies to components that have not started generating yet; the currently-generating component finishes as originally scoped.

Instantly provision a new anonymous Codemind account on the free plan — no credential, no human step required. Call this when build_feature (or any other tool) fails with an auth error and no CODEMIND_API_KEY is configured. Returns a tenantId and a bearer token (apiKey) usable immediately for build_feature and every other tool. The account is anonymous (no identity attached) until a person opens the claim link in the response text and signs in. If a tool call fails with an auth error and you ALREADY have a configured apiKey, call get_usage_guide with topic "auth" before calling this tool again — replacing an already-claimed account's credential here loses its identity and history. All arguments are optional; pass utm_source, utm_medium and utm_campaign, and invite, only if your human gave you them.

Get the aggregate status of a batch submitted via build_batch.

Read-only: your plan, how many verified builds you have used, how many are included and left in the current allowance window, and the paid-balance draw (cents) if you have topped up. Call this to check headroom before a large batch, or to learn why a build was rejected for allowance.

Get the current state of a build by ID (status and error details).

Retrieve the generated files from a completed build. Returns each file path and full content.

Get a build's original submitted spec (storyTitle, acceptanceCriteria, stackType, existingFiles, skipTestsFor) without retrying it — useful for inspecting or reusing a prior build's spec as a template.

Get the current state of a review by ID (status, verdict, findings) without streaming — useful for polling or inspecting a review after the fact. Use stream_review to follow a review in real time instead.

Get the current state of a test by ID (status, pass/fail verdict, generated verification test content) without streaming — useful for polling or inspecting a test after the fact. Use stream_test to follow a test in real time instead.

Returns the generated fix diff (as file paths and line counts, not full content) and before/after test-run output for one triage run, identified by the triage run id from get_triage_runs (NOT a build id and NOT an issue number). stdout/stderr are tail-truncated because test failure summaries appear at the end of output, not the install noise at the start. Full file contents are available separately via the get_build_files tool using the run's buildId. Also includes a per-LLM-call trace (stage/model/status/latency) from Analytics Engine when available, covering attempts that failed before ever reaching the before/after test comparison above.

Lists the issue-triage run history for one attached repo — one row per GitHub issue that was auto-triaged — including terminal status (`pr_opened` means a fix PR was opened and passed the repo's own test suite; `needs_human` means the issue wasn't judged auto-fixable; `failed` means the triage/build/verify pipeline errored). The `repoAttachmentId` is the attachment `id` returned by `list_repos`, NOT a `owner/repo` string. Full diagnostics (the generated diff and before/after test output) for any run id returned here are available via the get_triage_diagnostics tool.

Get detailed usage guidance for Codemind's tools — how to write effective acceptanceCriteria, when to use patch mode, how webhooks work, what to do when a build declines, common failure patterns unrelated to acceptanceCriteria quality, and how to recover from an auth error without losing an identified account. Call this before build_feature/test_component/review_code if you are unfamiliar with this tool surface, or if the same story fails repeatedly across retries; call it with topic "auth" before calling create_free_account to recover from an auth error on an existing credential.

List recent builds for this tenant.

Lists GitHub repos currently attached to the caller's tenant via Cloud Repo Mode (the GitHub App that automatically triages filed issues into fix PRs), and that the returned `id` field is the repo attachment id to pass into `get_triage_runs`.

Index files into the CIL codebase context for a project — for files Codemind itself did not generate (a manual edit, another tool's output). Not needed after a normal build_feature call: a completed build's own output is indexed automatically when projectId was supplied.

Retry a failed build using the same spec (story, acceptance criteria, stack, existing files).

Submit code for review: bugs, security issues, silent failures, and style concerns. Give it files (path+content) and/or a unified diff. At least one of files or diff is required — submitting neither is rejected. TWO independent LLM-judge passes review it (a general correctness/security pass and a dedicated silent-failure pass) and their findings are merged into one verdict. For a genuinely useful review, ALSO supply conventions (paste this project's CLAUDE.md and RULES.md content — you already have repo access, this tool does not) so both passes can enforce project-specific rules, and relatedFiles (any existing file the change references but doesn't itself modify — an interface being implemented, a caller of a changed function, a sibling example of the real convention) so findings can be verified against actual code instead of guessed. Omitting both still works but produces a weaker review with no codebase grounding. No code generation or execution happens here — this is judgment only, not test verification (use test_component for that). Returns a reviewId immediately; call stream_review to follow progress and get the findings. Call get_usage_guide for detailed usage guidance.

Stream live progress for a batch submitted via build_batch. Polls the batch's aggregate status periodically and emits a progress notification each tick; closes once the batch reaches a terminal status (completed or completed_with_failures). Call get_batch instead for a single point-in-time snapshot.

Follow a build submitted by build_feature in real time. Emits a progress notification on every phase change and resolves with the final result when the build completes or fails. Use immediately after build_feature returns a buildId.

Follow a review submitted by review_code in real time. Resolves with the findings and the approve/request-changes verdict.

Follow a test submitted by test_component in real time. Resolves with the pass/fail verdict and the generated verification test content.

Run standalone cloud verification on code you already have: give it a component (filePath/exportName/signature) plus acceptance criteria, and it generates a black-box verification test, validates that test against a mechanical stub, then executes your REAL code against it in the cloud (isolate or container, per stack). No code generation happens here — use build_feature for that. Returns a testId immediately; call stream_test to follow progress and get the pass/fail verdict. Call get_usage_guide for detailed usage guidance.