meridianlabs-ai

Check out the PR branch for an issue — /checkout <issue-number> finds the issue's PR (linked-PR chip fast path, then agent comments, then the claude/issue-N-* branch convention) and checks its branch out in the current repo clone.

Promote a reviewed inspect_ai fork branch upstream — open the UKGovernmentBEIS PR with the fully-qualified Fixes ref, then do the tracking bookkeeping (Atlas Sign-off stage, Upstream PR field, issue comment, supersede the fork PR). Idempotent — also use it to…

Merge approved upstream inspect_ai PRs from the Atlas board's Merge stage, one at a time — resolve conflicts against main, guard CHANGELOG/submodule invariants, coordinate companion ts-mono PRs and submodule pointer bumps when the viewer schema changed, watch…

Bring the Atlas board current in one shot — dispatch the hourly Atlas sync workflow (hosted in this agents repo) now (external-PR discovery + upstream state sync) and run the local link-upstream-chips sweep (clickable issue↔PR chips). Use when the board looks…

Import an upstream inspect_ai issue into the fork — /import <upstream-issue-number-or-url> mirrors the UKGovernmentBEIS issue as a fork issue (snapshot body + machine-readable upstream link) and adds it to Atlas as Todo, so @auto can work it. Does NOT kick…

Relay an external review upstream — /post-upstream-review <proxy-issue-number> [instructions] takes the review findings from an External proxy issue, posts them as a single review on the contributor's upstream PR (inline comments on the right lines where…

Use for monitoring, diagnosing, and intervening in RUNNING Inspect AI evaluations via the `inspect ctl` command-line control channel. Triggers when: watching a live eval for stalls/errors/retry failures, checking progress of an active eval process, launching…

Use whenever the user needs to read, inspect, or process Inspect AI eval log files (`.eval` or `.json`). Covers the Python API (`read_eval_log`, `read_eval_log_samples`, `read_eval_log_sample_summaries`, `list_eval_logs`), `header_only` mode, the critical…

Use whenever the user wants to analyze or understand what happened in an Inspect AI eval, sample, or set of logs. Routes to the right tool (`inspect_ai.log` for single-log work, `inspect_ai.analysis` for cross-log dataframes, Inspect Scout for transcript…

Use whenever the user is working with Inspect AI or any of its ecosystem packages (Evals, Flow, Scout, Viz, SWE, Harbor, Sandboxes). Maps concerns to packages and points at each package's docs before calling its API.

Realistic customer service application with system prompt, tool schemas, and sample tickets for grounding audit scenarios

Realistic corporate email archive for email-management audit scenarios

Recurring maintenance task — clear the current batch of GitHub dependabot security alerts by adding or updating pnpm override entries in pnpm-workspace.yaml. Use whenever the user mentions dependabot alerts, security advisories, CVEs/GHSAs, vulnerable…

Drive the real inspect log viewer (apps/inspect web UI) against real .eval fixture logs and capture evidence that user-facing behavior works — log list, sample detail, transcript, scores. Use when a change to the viewer needs proof it behaves correctly in the…

Land a PR that requires a coordinated ts-mono submodule change. Use when the check-schema-and-types CI job fails, when a Python model change requires regenerating the OpenAPI spec / TypeScript types, or when any change touches the ts-mono submodule and needs…

Assemble the welcome banner by reading the asset and running the script

Triage the nightly auto-update PR for the Harbor registry — fill in categories, title, repo, arxiv on newly-stubbed `docs/overrides.yml` entries so CI passes and the docs/AISI evals browser surface the dataset with proper branding.