MCP tools

Generated from src/registry.ts by pnpm docs:gen. Do not edit by hand. The MCP server and the built-in agents use this same list. Every tool takes an optional repo (a local path or GitHub URL; defaults to AVERT_REPO or the working directory). Results carry a path: the ordered steps from the change to the route, with file and line.

Tools marked write change the graph or call GitHub, so the hosted read-only endpoint does not expose them.

Tool	Access	What it does
`index_repo`	write	Index (or re-index) a repository: routes, functions, call chains and every external API call site.
`sync_changes`	write	Pull provider changes for the SDKs this repo uses (SDK release diffs + model lifecycle data) and create findings.
`list_changes`	read	List provider change events that touch this repo (removed, renamed, deprecated, retired, added). Added = notice only.
`blast_radius`	read	Everything one change breaks: call sites (down the API surface, through rolling aliases), then routes and owner teams (up the call chain), each with its path.
`list_findings`	read	Findings (one per change x call site) with status open \| claimed \| done \| declined.
`explain_path`	read	The evidence path for one finding: change -> surface -> call site -> functions -> route -> owner, with file and line.
`rank_findings`	read	Rank findings by severity, urgency (days to effective date) and reach (routes + PageRank centrality of the call chain).
`migration_batches`	write	Group findings into migration batches: connected components (WCC) of findings sharing an owner and a change or function. One batch = one issue.
`claim_finding`	write	Atomically claim an open finding so no other agent works on it. Fails if already claimed, naming the holder.
`record_decision`	write	Remember a decision about a finding (fix, defer, wontfix, accept_risk, note) so later sessions and other agents see it.
`recall`	read	Recall memory: decisions and sessions (episodic) and migration recipes (procedural, from this repo and the shared graph). With no ids, returns recent history. Call this first.
`sdk_params`	write	Params of a finding’s operation in the provider’s latest SDK, with doc text and deprecation flags. Use it to choose a replacement when a change names none.
`save_recipe`	write	Save how a change was fixed (from -> to: a param name or literal value, with steps) so the next matching migration reuses it.
`publish_recipe`	write	Publish a repo recipe to the shared graph so other repositories hit by the same change are offered it.
`open_issue`	write	Open one GitHub issue for a migration batch, with every path as evidence.
`open_pr`	write	Open a GitHub pull request fixing one finding, when the change names a replacement or a recipe covers it. Never merges.

Arguments

`index_repo`

Index (or re-index) a repository: routes, functions, call chains and every external API call site.

No arguments besides repo.

`sync_changes`

Pull provider changes for the SDKs this repo uses (SDK release diffs + model lifecycle data) and create findings.

No arguments besides repo.

`list_changes`

List provider change events that touch this repo (removed, renamed, deprecated, retired, added). Added = notice only.

kind: removed | renamed | deprecated | retired | added
provider: string — Provider id: openai, anthropic, stripe, gemini, twilio

`blast_radius`

Everything one change breaks: call sites (down the API surface, through rolling aliases), then routes and owner teams (up the call chain), each with its path.

change_id (required): string — ChangeEvent id from list_changes

`list_findings`

Findings (one per change x call site) with status open | claimed | done | declined.

status: open | claimed | done | declined

`explain_path`

The evidence path for one finding: change -> surface -> call site -> functions -> route -> owner, with file and line.

finding_id (required): string

`rank_findings`

Rank findings by severity, urgency (days to effective date) and reach (routes + PageRank centrality of the call chain).

status: open | claimed

`migration_batches`

Group findings into migration batches: connected components (WCC) of findings sharing an owner and a change or function. One batch = one issue.

No arguments besides repo.

`claim_finding`

Atomically claim an open finding so no other agent works on it. Fails if already claimed, naming the holder.

finding_id (required): string

`record_decision`

Remember a decision about a finding (fix, defer, wontfix, accept_risk, note) so later sessions and other agents see it.

finding_id (required): string
kind (required): fix | defer | wontfix | accept_risk | note
reason (required): string

`recall`

Recall memory: decisions and sessions (episodic) and migration recipes (procedural, from this repo and the shared graph). With no ids, returns recent history. Call this first.

finding_id: string
change_id: string

`sdk_params`

Params of a finding’s operation in the provider’s latest SDK, with doc text and deprecation flags. Use it to choose a replacement when a change names none.

finding_id (required): string

`save_recipe`

Save how a change was fixed (from -> to: a param name or literal value, with steps) so the next matching migration reuses it.

change_id (required): string
from (required): string
to (required): string
steps (required): string
callsite_id: string

`publish_recipe`

Publish a repo recipe to the shared graph so other repositories hit by the same change are offered it.

recipe_id (required): string

`open_issue`

Open one GitHub issue for a migration batch, with every path as evidence.

batch_id (required): string — Batch id from migration_batches

`open_pr`

Open a GitHub pull request fixing one finding, when the change names a replacement or a recipe covers it. Never merges.

finding_id (required): string