aictrl.dev

List the first-party aictrl tasks that can run for one live-authorized repository. It reports only tasks aictrl provisions — Code Review and PR Explanation — and never templates authored inside the organization. An empty result means no first-party task is available for this repository right now; it is not a statement about any other aictrl capability. Listing never starts work.

list_tasks

Read all recorded findings for a PR (any lifecycle status) so verdicts can be recorded by exact id. Returns the recorded `id` and the platform-owned `githubCommentId` to join against GitHub review comments (exact for findings the platform delivered inline; null for older / un-anchored findings), plus the full claim (description, suggestion, snippets), lifecycle, and the latest verdict and verification. Set `include_history` for the append-only judgment chains. Judgments written by YOUR OWN run are never returned — a run must not be shown its own prior conclusion as evidence. Findings the obsolescence sweep has resolved are excluded unless `include_obsolete` is set. Read-only — records nothing.

list_pr_findings

Universal entry point for exploring project context. Query code, architecture, issues, backlog, agent skills, and stack layers. Use this tool whenever you need to DISCOVER or READ information about the project.

query_context

Apply a post-hoc triage verdict to a finding that was previously recorded via record_finding. Used by the /reply-to-code-review flow to persist whether a finding is TRUE / FALSE / UNCERTAIN and the FIX / DEFER / IGNORE action taken. The finding is located only by the exact `finding_id` returned by list_pr_findings, on ANY lifecycle status (open, fixed, or dismissed — so verdicts still apply after a PR is merged and findings were swept to dismissed). Only findings recorded into the data layer (the aictrl-dev MCP path) can receive a verdict. Idempotent: re-applying overwrites the prior verdict. A not_found response is an application error; call list_pr_findings to refresh stale IDs and do not treat the write as successful. `judged_at_sha` is required: send the full 40-hex commit SHA of the code you actually read. A SHA that is neither the review SHA nor a later commit of the same PR is rejected with validation_error. `reasoning` is required too (#5364): a verdict whose reasoning cannot be audited later is not worth storing. Each write appends to the finding's judgment history; a re-judge from the same run at the same commit updates that run's own row rather than appending a duplicate. A FALSE verdict on an OPEN finding of severity MINOR or NIT (medium/low/info) also dismisses it as a false positive, with your `reason` as the note; on MAJOR or BLOCKER (high/critical) it stays a suggestion a person confirms. The result reports `autoDismissed` (and `autoDismissReason` when false). A dismissal is not reopened by this tool; a person can restore the finding from the pull-request page.

record_finding_verdict

Start one first-party task returned by list_tasks for an exact pull-request revision. PR Explanation reads the pull request and stores a private explanation in aictrl; it posts nothing to GitHub. Code Review reads the pull request and posts its findings as review comments on that pull request. Call only after the user explicitly asks to run that task in the current turn. Retrying the same request with the same idempotency key returns the same execution.

start_task

Start a published workflow with validated workflow-level inputs. The selected workflow may perform its declared external side effects, such as creating a pull request, only within its configured gates and policy.

start_workflow