Celmis

Draft review rules for one repository from its code, for approval. Refused where the installation cannot generate them.

HTTP routes of a repo found from framework syntax (FastAPI, Flask, Express, Laravel, Symfony, Go): method, path, handler, file:line. Optional path_glob. Heuristic; supported=false means it could not look.

Return the cached architecture summary (markdown) for `repo_slug`. This is a compact orientation doc — entry points, main flows, ownership, integrations. Regenerate from UI or via POST /api/intel/architecture/{slug}/rebuild.

The workspace's monthly spend cap and how much of it is used: cap_usd, spent_usd, used_pct, alert_pct, hard_stop, over_cap.

Status and summary of a dependency audit. Omit run_id for the latest run. status is queued|running|done|error — poll until it leaves queued/running, then read list_dep_findings.

Return the caller's effective research access for `repo_slug`: whether it is researchable, whether source code is visible, and which path globs are denied (creds/crypto/db). Use to explain to the user why some detail is out of bounds.

Return owner info for a file (or fallback ancestor dir) in `repo_slug`. Data: top git-blame authors from the last N days, matched CODEOWNERS entries, and the single 'primary_owner' — the top committer identity. Use when you need to know who to page / cc on a change.

Return one project's full detail: repos (slugs + roles), descriptions, chat count. Use this to enumerate the repos you should also query when writing code for one of them.

Fetch the latest review run for a PR reference like `gitlab:owner/repo#42` or a full URL. Returns verdict, findings (severity, agent, file:line, suggestion), and — for the workspace owner or admin only — cost. Use to see whether a PR passed, or to trigger a fresh run.

Read the effective review policy for `repo_slug`: enabled flag, target branches, natural-language rules, folder rules, per-agent model overrides, per-agent prompt overrides. Use to know what the AI reviewer will enforce before opening the PR.

One review run: summary, verdict, agents and findings (severity, file, line, title), bounded by limit (default 20, max 40; `truncated` says it was cut). Give run_id, or repo_slug + number for the latest run of that pull request.

The code-review settings in force: workspace defaults or one repository (repo_slug), where each value comes from (repo | workspace | install), which agents run, guidelines, rules and which repositories override the defaults.

LLM spend of the workspace over a period: totals, tokens, cache hit, top surfaces/models/agents/repos/operations and a series. days 1-365 (default 30); optional surface, model, repo_slug; bucket hour|day|week|month. Owner or admin only.

Fetch full details about a symbol by id. Use after find_symbol() to get the extra fields (signature, docstring, exports and so on).

Code-review run usage over a period: runs (completed / failed), tokens, cost and a daily series. Needs owner, admin or editor on the workspace — the same gate as the Analytics page.

Queue a (re-)index of repos, the code graph that search, Q&A and reviews read. repo_slugs names them; omit for the workspace (owner narrows; existing graphs are skipped unless force=true). At most 50. Needs `review` on each repo.

List repos the CURRENT caller may research, with visibility level (metadata|code) and any deny-path patterns. Use this first to know what you are allowed to investigate — repos you cannot research are omitted. Mirrors the human UI + REST /api/access/my so agents and people see the same boundaries.

Incoming monitoring alerts of the workspace, newest first: id, title, severity, status (new | acked | fixed), repo hint. Optional status filter; limit up to 50. Use the id with ack_alert.

Findings of one audit run, worst severity first: package, installed vs latest, severity, recommendation. Optional severity filter. This is the material a report is written from.

List every tracked deprecated symbol with its known consumers, replacement (if any) and target removal date. Optional filter `repo_slug` narrows to one repo.

List all repository groups defined in the workspace. Each group represents related repos for cross-team analysis (for example, frontend + backend + mobile of one product).

Tracked review issues, worst severity first, with counts per status. status open|fixed|dismissed|resolved, comma-separated (default open); optional severity, repo_slug, pr number, text q; limit default 15, max 25. A row's id is what update_issue takes.

Background jobs of this workspace (indexing, documentation, audits, reviews) with counts per status. Optional status (pending | running | done | failed | dead | cancelled) and kind filters; limit up to 50. Use the id with retry_job / cancel_job.

Members of the workspace with their roles and teams. Read only — inviting people or changing roles is done in the app.

List all projects (multi-repo bundles). Each project groups related repositories that share domain / owning team so cross-repo lookups make sense. Returns id, name, description, repo count.

List repositories: either all indexed repos (group_name=None) or repos in a specific group. Returns slug + indexing status + symbol count.

The newest review runs of one repository (repo_slug) or of the whole workspace: pull request, status, verdict, finding counts by severity, when. status filters (complete|failed|running|queued|partial|skipped); limit defaults to 10, at most 25. Each row carries the run_id for get_review_run.

Operational state of every workspace repo: indexed, documented, auto review on/off, branch. Run before any verb that takes repo_slugs. Differs from list_accessible_repos, which says what the CALLER may research.

Bulk-apply a text replacement across all consumers of `symbol` in the project: opens branch `celmis-migrate/<symbol>-<n>` per consumer and commits old_text to new_text at the consumer's line. Repos whose provider cannot apply fixes (only github can) are skipped.

Propose review rules for one repository (or the workspace where the review-rules list exists). rules: a list of {title, instructions, path_glob?, severity? info|warning|error|critical, agents?}; at most 10 at a time. Where the rules list exists they are saved PENDING for an editor's approval.

Execute read-only Cypher query against per-repo (repo_slug) or cross-repo (group_name) graph. Write keywords blocked: CREATE/DELETE/SET/MERGE/REMOVE/DROP. Returns rows as a list of dicts. Use this for complex queries that the tools above do not cover.

Put a dead, failed or cancelled background job back in the queue. job_id from list_jobs. Needs owner or admin on the workspace, and the job must belong to it.

Queue an AI review of ONE pull request (repo_slug + number), or all_open=true for up to 25 open ones (numbers, branch narrow it). post_comments (default true) posts findings. Needs `review` on the repo. Returns run ids for get_review_run.

Given a stack trace (Sentry-style or plain text) referencing files inside `repo_slug`, return the owner(s) most likely to be responsible. Uses the ownership snapshot. Optional `title` is used only in the response payload.

Find things in code. kind=search (default): symbols and doc notes matching `query`, optionally in repo_slug. kind=usages: callers of symbol `query`. kind=owner: owners of `path`. kind=architecture: cached summary. Bounded by limit (default 15, max 50).

Search the code of the project this token belongs to: symbol names from the index plus a text search over the files (works for any language, Unicode names included). Optional repo_slug narrows to one repository of the project. Free; bounded by limit (default 15, max 50).

Search a function/class by name across the repos of a project (or all your repos when project_id is omitted). mode: auto (exact, then substring), exact, prefix, fuzzy. Returns repo, file, start/end line, kind. Use before writing new code so you do not reimplement an existing symbol.

Turn automatic PR review on or off for a set of repos (repo_slugs, owner prefix, or none for the whole workspace) and optionally pin the branch every surface then reads. Refuses above a cap, since it changes every future pull request.

Set the workspace's monthly spend cap in USD (0 = no cap), the alert percentage (1-100) and hard_stop (block calls once the cap is reached). Needs owner or admin on the workspace.

Queue a dependency and vulnerability audit. repo_slugs empty = every workspace repo; owner filters by 'owner/' prefix; branch overrides for this run. report_engine none|api|claude_code picks who writes the prose. Returns run_id.

Ordered checklist for a cross-team integration against `target_repo_slug`: read the architecture summary, list owners, pick endpoints, draft a client. Use at the start of an integration task instead of ten separate calls.

Set the status of one or several review issues (issue_ids, at most 25): open | fixed | dismissed | resolved. Needs the member role or above on the workspace. Reopening clears the close marks; any other status records a manual resolution.

Change ONE review setting: scope workspace|repo (+ repo_slug). key: run_on_drafts, approve_when_clean, request_changes_on_critical, committable_suggestions, comment_min_severity, max_inline_comments, summary_enabled, review_language, disabled_agents, agent_prompt_guidelines. null value inherits.