Tools, examples and evidence limits for connected agents.
Loading task examples…
First find where your code imports and calls the library. Then ask about that call and name the library file it reaches. You get excerpts cited by commit and line. Add a base commit to compare revisions, or a public PR number to inspect its source and GitHub check reports.
Retrieval covers up to four files and reports missing evidence. Source excerpts and passing CI do not establish compatibility with your application; verify the affected behavior locally. Use source_ranges with the cited commit to request missing context, up to 40 lines per range.
1 · Before
Your agent evaluates the exact version against your project’s package facts and keeps the returned verification context for step 3.
2 · Local work
You approve the change; your agent makes it and runs the checks. RepoPilot never installs packages or runs commands.
3 · After
RepoPilot checks the resolved versions, lockfile and advisories. Check results your agent reports are recorded as reported, not re-run.
check_dependencyWhen you want a quick yes-or-no on a repository or package, without details of your own project.
Repository recommendation, receipts, confidence, and evidence gaps.
evaluate_dependency_changeBefore adding or upgrading an exact npm version in this workspace.
A policy decision, advisories for that exact version, install-script and provenance evidence, checks to run, and a rollback plan.
verify_dependency_changeAfter the manifest, lockfile, and local checks have changed.
Checks of the resolved versions, lockfile and advisories, plus the check results your agent reports.
analyze_repoWhen investigating a specific source question or entering an unfamiliar repository.
With a question: commit-pinned source excerpts, optional before/after revisions and observed PR checks. Otherwise: the same repository brief copied from the report, including purpose, main parts, starting files, setup, revisions and gaps.
plan_repo_taskBefore editing an unfamiliar repository at a known checkout SHA.
Ranked files, the code that imports the dependency, checks to run, how the checkout relates to the analyzed commit, and a plan ID for check_change_risk.
get_artifactBefore substantial code work, after the short brief.
The full saved report: key files, entry points and a reading order.
compare_reposWhen choosing between two repositories or npm packages.
Per-use-case preference, trade-offs, confidence, and unknowns.
check_change_riskBefore merging a pull request or after producing a local diff.
A score for how risky the diff looks from the files it changes, and what to review first.
Compact by default
Briefs preserve context. The full repository artifact is opt-in.
Unknown means unknown
A cache miss or missing evidence is never presented as approval.
Don't auto-retry two tools
Tools do not edit your repository. Each plan_repo_task call starts a new plan, and each check_change_risk call on that plan counts as an attempt. Retrying either in a loop creates duplicates.
agent.summary is the action-oriented answer; confidence describes the available repository evidence; and evidence_gaps states what the agent must not assume.
A dependency result with evidence_source: live_fallback uses bounded GitHub/OpenSSF evidence because no full analysis is cached. It has limited confidence, keeps tests, CI, published dependency advisories, and code structure unknown. It does not start a full repository analysis or consume the uncached-analysis budget. These live checks share one limit of 500 a day across all callers; results cached in the last five minutes and full cached analyses do not count.
Exact dependency evaluation combines immutable npm metadata, deps.dev advisories, project runtime/peer/license constraints, and repository evidence. It still does not certify safety or production fit. Check results your agent reports are labelled as reported by the caller.
Task text is sent to the model provider to rank files at the requested revision and is never persisted raw. Project snapshots, manifest/lock contents, diffs, command strings, and proof output are not retained.
check_change_risk accepts bounded source-free file facts: relative paths, status, additions, deletions, and optional previous paths. Filenames are still metadata. Omit the repository slug when identity is sensitive and inspect payloads before sending. The @repopilotapp/cli package, run on your machine, reads one project folder, collects these facts, and shows the exact outbound payload in --dry-run mode.
Tools your agent calls over MCP when it needs facts it cannot see from the code alone.
evaluate_dependency_change
Checks the exact version your agent is about to install against deps.dev and OSV. A version with a high or critical advisory comes back as avoid, with the advisory ID; a clean or already-patched version is confirmed in one plain sentence.
Measured: With RepoPilot, Claude Haiku caught and named the advisory on 39 of 41 vulnerable versions published after its training, against 29 without it.
evaluate_dependency_change
Resolves the dependency tree a fresh install would bring in and checks every package in it, naming any vulnerable one, the chain that pulls it in, and how to override it.
Measured: On 12 packages that were clean themselves but pulled in a vulnerable dependency, Claude Haiku caught 11 with RepoPilot, against 2 without it, and kept all 8 clean packages either way.
verify_dependency_change
After the install, checks every package the lockfile change added, up to 1,000, including the ones a fresh resolution misses, such as automatically installed peers, and names each vulnerable one.
Measured: On 10 packages whose vulnerable dependency only appeared in the installed lockfile, Claude Haiku caught all 10 with this check, against none with the pre-install check alone, and kept all 6 clean packages.
evaluate_dependency_change
Compares the calls your code makes against both published versions of a package and names the ones the upgrade breaks, even when the release notes don't mention it.
plan_repo_task
Turns an issue into the files most likely to change, ranked from the code and the repository's own history of similar fixes, with the evidence behind each pick.
Measured: Put the right file in its top three on 11 of 16 large repositories it had never seen.
Measured results come from tests whose plans were fixed before they ran, on fresh cases with Claude Haiku 4.5, in September and October 2026. They describe those test sets and are not a guarantee for your project; see the terms.
Also over MCP: check_dependency, compare_repos, analyze_repo and check_change_risk. Connect your agent →