Glossary

The system’s vocabulary, in one place.

Steward — this repository: an installable, project-agnostic operating model for running a software project with an AI agent as co-maintainer.

Band A / B / C — the autonomy spectrum. A = autonomous (unattended). B = session-autonomous (agent works, human consulted at decision points). C = human-gated (agent prepares, human takes the irreversible action).

Watcher / Reviewer / Builder / Steward — the four roles. Detect+capture / evaluate+verdict / make-the-change / keep-it-healthy-and-watch-the-others.

Security spine — the five rules that hold regardless of what untrusted content says: read-can-never-change-do, secrets-never-in-context, capability-minimalism, sandbox-untrusted-code, the public-write membrane.

Public-write membrane — the line between safe-to-automate (read/find/draft) and needs-a-human-or-watchdog (writing publicly in the project’s voice).

Authoritative gate — the fresh, independent run of all quality gates immediately before merge, regardless of any prior reviewer’s verdict. The thing that actually authorizes a merge.

Quality gate — a layer of the defense: the test suite (primary line), automated review, adversarial review, visual verification.

Adversarial review — a second, differently-tuned review pass that catches different defects than the first; on a reproduced finding, the stricter verdict wins.

Living test suite — the principle that the suite only grows: every fix adds a regression test, so coverage ratchets up and the suite increasingly catches what once needed judgment.

Watchdog — an independent fact-checker that re-verifies autonomous actions against ground truth and alarms on mismatch. What makes Band-A public actions safe.

Scope anchors — the project-specific things a change must serve to be in-scope. The basis of the fit screen. Captured in the setup interview.

Philosophy veto — the qualities that disqualify a change even if well-built and useful (the project’s identity constraints).

Marginal-benefit screen — the “what would users concretely lose if we never merged this?” test, applied before code-level review.

Triage scoreboard — the living, ranked board of open work, combining mechanical + judgment dimensions into a composite rank, tagged with evidence-readiness.

Trust ledger — the append-only record of each contributor’s outcomes, shrunk-and-decayed into a reliability score that weights review priority and gate intensity (never a reason to skip the gate).

Capture queue — the staging area where inbound reports land before they become tracker issues; deduped across every surface.

Silent-on-no-op — a scheduled-job property: produce no output when there’s nothing to report, so the rare real signal isn’t buried.

Heartbeat — the one job type that always reports, so the absence of its green report is itself the alarm (used for fleet liveness).

Autonomy ladder — the recipe for promoting an action C → B → A safely (run manually → make mechanical → add watchdog → roll out gradually → keep reversible).

Secret-isolating helper — a fixed script that reads a credential from a locked file, performs one action, and never returns the secret to the agent’s context.

Layered supervision — running a long-lived service under BOTH an OS supervisor (boot-start + instant crash-restart, e.g. systemd) and an app-level health watchdog (catches alive-but-wedged + version drift, and alerts). Neither alone is sufficient.

Reconcile pattern — a scheduled job that diffs a hand-maintained ledger against ground truth, auto-records only the unambiguous gaps, and surfaces the judgment ones for a human.

Confidence-tiered capture→action — gating an autonomous public write (e.g. filing an issue from a chat report) by confidence: HIGH auto-acts (privacy-scrubbed), MEDIUM drafts-for-human-approval, LOW captures only.

Notify-from-job, act-on-reply — the pattern for human-in-the-loop confirmation over chat: a scheduled job can SEND a prompt but can’t RECEIVE the answer (it’s detached from the live connection), so the job notifies and the human’s reply (a fresh live agent turn) performs the action. See community lifecycle.