Resolve AI

Post a message to a Resolve chat. Provide IDs based on what you're doing: - Continuing an existing chat inside an investigation: pass BOTH `chat_id` AND `investigation_id`. - Continuing a standalone chat (not in any investigation): pass `chat_id`. - Starting a new chat inside an investigation: pass `investigation_id` (no `chat_id`). - Starting a standalone chat: pass neither. Returns immediately (non-blocking) with `chat_id` (and `investigation_id` when scoped), a `url`, and usually a `stream_command` — a self-contained `curl` that streams the reply. Run `stream_command` (when present) as a background process that stays alive across steps so you can read its streamed stdout, and follow the reply live; it exits when the turn completes. Treat an in-band `[done]` marker as success and `[error: ...]` as failure, even if the process exits normally. Use `get_chat` to read the conversation, condensed — including the final result once it has finished. To attach uploaded files/images, pass `attachments` with IDs returned by `upload_attachment`. If the host cannot expose raw bytes for upload, describe the image/file directly in `message` instead.

ask

List recent alert events for the caller's org — the same data the Resolve `/alerts` page shows. Returns each alert's `alert_id`, `title`, `severity`, `action` (`fire` / `warn` / `resolve`), `time`, target `entity_key`, attribution, alert-rule reference, full `labels` map, `is_noisy` flag, `skip_investigation_reason` (why an alert was not auto-investigated, or null), `is_auto_investigated` flag, and `investigation_id` (drill-in slug; pass to `get_investigation` / `ask` / `steer_investigation`; absent if no non-stub canvas exists even when `is_auto_investigated` is true). Sort: `time` descending (newest first). `time_range` defaults to `last_24h`; use `custom` together with `from_iso` (required) and `to_iso` (optional, defaults to now). Filters compose with AND semantics. `team_id` scopes to a team's configured alert rules; `alert_rule_key` narrows to a specific rule's composite key (e.g. `"High CPU::12345"`); `only_auto_investigated=true` returns only alerts that triggered an automatic investigation; `alert_id` fetches a single alert by id. `label_filters[]` mirrors Resolve's label-filter expression: each entry is `{ label_name, operator, values }`. Filters across the array are AND'd; `values` within a filter are OR'd. Operators: `MATCHES` (exact), `NOT_MATCHES`, `CONTAINS` (substring), `NOT_CONTAINS`, `EXISTS` (label is present, `values` ignored), `NOT_EXISTS`. Examples: `[{label_name:"severity",operator:"MATCHES",values:["critical","high"]}]`, `[{label_name:"team",operator:"NOT_EXISTS",values:[]}]`. `limit` defaults to 50 with a hard ceiling of 5000 (requests above 5000 are rejected). Use to triage which alerts to drill into via `get_investigation` / `ask`.

list_alerts

List Resolve chats. Pass `investigation_id` to list threads scoped to that investigation; omit to list your recent chats (legacy chat threads and Reclaw agent threads, treated the same). Returns each chat's `chat_id`, `name`, current `status` (`running` / `complete` / `errored`), and `created_at` / `updated_at`. When `investigation_id` is set, `time_range` / `from_iso` / `to_iso` are ignored — the investigation is the scope. For the standalone listing, `time_range` defaults to `last_7d`; use `custom` together with `from_iso` (required) and `to_iso` (optional, defaults to now) for a precise window. Sort order: `updated_at` descending. Use between turns to find chats that have settled — running chats indicate the agent is still working.

list_chats

List recent Resolve investigations for the caller's org. Returns each investigation's `investigation_id`, `title`, `use_case` (ALERT / CHAT_INVESTIGATION / INCIDENT), `run_type` (AUTO / MANUAL), `status` (live position: initializing / running / grouped / stopped, or null), `stop_reason` (raw last stop reason; `awaiting_*` = paused, not done; null if none), `triage_only` flag, `is_noisy` flag, the resolved `team` (display name), primary `alert` summary, the linked `slack_channel`, and `created_at` / `updated_at`. `time_range` defaults to `last_24h`; use `custom` together with `from_iso` (required) and `to_iso` (optional, defaults to now) for a precise window. Filters compose with AND semantics — `team_id` scopes to a team, `run_type` / `use_case` / `triage_only` narrow further. Sort: `created_at` descending (newest investigations first). `limit` defaults to 50 with a hard ceiling of 500 (requests above 500 are rejected). Use to discover an `investigation_id` before calling `get_investigation`, `ask`, or `list_chats`.

list_investigations

Fetch the current state of a Resolve chat — `status` (`running` / `complete` / `errored`), `total_messages`, and the (possibly sliced) transcript of user and assistant turns. Pass `investigation_id` when the `chat_id` refers to a thread inside an investigation (you got both from `list_chats` or `ask`); omit it for standalone chats. For polling: store `total_messages` from the previous response and pass it as `after_index` to fetch only newer messages — when the chat is still running with no new content, the response's `messages` array will be empty.

get_chat

Fetch a Resolve investigation by ID. Returns its metadata (title, team, phase, linked `slack_channel`) alongside its alerts, report, theories, and mitigations inline. Use `read_file` to get more details on files not returned. If the user gave a Resolve canvas URL, the ID is the path segment after /chat/ — extract it before calling. Also returns a `stream_command` for following the investigation's live trace (theories, evidence, phases) without steering.

get_investigation

Read a full text file scoped to a Resolve investigation or standalone chat. Pass BOTH `chat_id` and `investigation_id` for a chat/thread inside an investigation; pass only `chat_id` for a standalone chat; pass only `investigation_id` for investigation-level files. Investigation threads share their parent investigation's workspace files, so `investigation_id` selects the workspace when both IDs are present. Use `/workspace/agents/<agent>/citations/...` paths from Resolve citations.. The file content is returned inline. Very large files fail with a size error instead of being truncated.

read_file

Kick off a new Resolve investigation from a markdown problem statement. Reuses the backend signal of the UI's `Start RCA` short-circuit. Compose `prompt` from the user's problem description, any local observations, and relevant context (paste alert text, log snippets, suspected components inline — there is no separate alert_ids parameter). Creates server-side state and runs against the org's action / credit budget — confirm with the user before calling. Returns `investigation_id`, a `url`, and a `stream_command` for following the live trace (theories, evidence, phases). Call `get_investigation` for the condensed state.

start_investigation

Send a steering directive to an active Resolve investigation. Influences the investigation's direction — confirm with the user before calling. Reuses the backend signal of the UI's Steer button — the directive flows to the agent teams. Returns a `message_id` for this steer; the trace tags this steer's progress and reply lines with it. Also returns a `stream_command`, but only run it if you aren't already following this investigation — one subscription carries every steer's tagged updates. To attach uploaded files/images, pass `attachments` with IDs returned by `upload_attachment`. If the host cannot expose raw bytes for upload, summarize the image/file directly in `message` instead.

steer_investigation

Submit explicit human feedback to Resolve. Provide exactly one of `investigationFeedback`, `chatFeedback`, or `productFeedback`, matching the target the human is judging: an RCA investigation, a chat answer, or the Resolve product / MCP experience. For product feedback, include only session-observed facts in `details`; never invent diagnostics.

submit_feedback