Firstmate Agent Crew

README and docs/configuration.md describe a least-privilege design: firstmate is read-only over projects except narrow guarded, captain-approved operations under AGENTS.md hard rule 1, and a self-repo boundary blocks upstream pushes without maintainer opt-in or per-task consent. User confirmation exists via install consent, merge approval, and explicit +yolo opt-in. Data-flow transparency is only partial: Relay needs a .env pairing token and posts public X/Discord replies, with dry-run previews mentioned, but no complete data-flow or retention description, so deducted. Sensitive data handling mentions credentials not being copied and 0600 atomic writes for model pins, but lacks systematic handling of tokens, logs, or transcripts, so deducted. Dependency security pins ShellCheck/actionlint/Herdr 0.7.4 in CI, but tasks-axi and the Pi package are unpinned and no vulnerability scanning is shown, so deducted. External effects include PR creation, merges, and public replies; gated by modes and consent but the default surface is broad with no consolidated effect inventory, so deducted. Rollback has guarded /updatefirstmate with restart, worktree isolation, and on-disk state, but no described rollback for merged changes or published public replies, so deducted. Source attribution has MIT copyright to Kun Chen and author links, but publisher identity is unverified and external sources are not systematically attributed, so deducted.

Self-consistency: README, configuration.md, and verification docs agree on FM_HOME, backends, and supervision concepts with stable terminology, scoring 2. Dependency availability: relies on many external tools (tmux, gh, treehouse, herdr, quota-axi, Pi package, tasks-axi); some are pinned in CI but tasks-axi and the Pi package are not, and quota-axi has only a version floor, so deducted. Failure messages: docs describe several failure behaviors (refuse to build on unavailable model, failed write not reported as saved, unresolvable path names the variable), but there is no unified error-code or diagnostics catalog, so deducted.

Audience and scenarios: README clearly targets developers running multiple coding agents with parallel fix, investigation, and audit scenarios, scoring 2. Capability boundaries: explicitly states it is not a model, harness, skill, MCP server, or CLI, and flags codex-app as unsupported and Zellij/Orca/cmux as experimental, scoring 2. Trigger precision: skill triggering depends on AGENTS.md and per-harness hooks, with documented differences (Cursor headless lacks a turn-end hook, Grok needs --trust), making triggers complex and harness-dependent, so deducted. Environment fit: macOS/Linux only, requires tmux and other tools, Windows unaddressed, so deducted.

Information architecture: README has clear features, quick start, how-it-works, skills table, and documentation index, and configuration.md has a find-a-setting table, scoring 2. Install notes: gives gh auth login, clone, and per-harness launch commands with trust prompts, scoring 2. Naming stability: many skill and config file names, some marked experimental, with no versioned naming policy, so deducted. Examples and FAQ: has dialogue and Bearings invocation examples but no standalone FAQ, so deducted. Known limitations: multiple experimental backends, blocked codex-app, and Cursor headless limits are documented, scoring 2. License: full MIT text present, scoring 2. Versioning/changelog: no CHANGELOG or version scheme found, scoring 0. Maintenance responsibility: CONTRIBUTING.md is referenced and the author is named, but there is no governance, support channel, or maintainer commitment statement, so deducted.

Output usability: delivers PRs, local merges, and investigation reports plus /bearings digests with clear output shapes, scoring 2. Marginal value: addresses the real pain of orchestrating parallel agents and session juggling with clear gains over manual tabs, scoring 2. Cost benefit: depends on many external tools and harness subscriptions, complex configuration, and while supervision is zero-token the overall learning and operational cost is high, so deducted.

Claim traceability: verification docs give dates, versions, and command outputs, which is fairly traceable, but many README feature claims are not individually linked to verification records, so deducted. Cross-source corroboration: README, configuration.md, and verification docs corroborate some facts (quota-axi versions, Herdr pin), but most claims rest on a single source, so deducted. Fact/inference separation: verification docs explicitly separate observation from inference (absent fields treated as uncertainty), but README mixes marketing phrasing with facts, so deducted.