You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
This commit was created on GitHub.com and signed with GitHub’s verified signature.
Added
fnl-coding-routerreferences/non-code-work.md, read on demand for queries on live data, infrastructure, dashboards and forecasts; the router stays code by default, and DEEP now also covers compounding calculations, predictions and ambiguous input.
fnl-coding-router "Proportion": build what the acceptance criteria require at the product's current stage, trace each hunk to a criterion and run tests at checkpoints, not after every edit or across the whole suite. Handoff briefs, verification and the review gate carry the limit.
Changed
A judge named in the host config (for example typesafe-ai for TypeSafe's Jev) now makes the grade, route, depth and next-activity calls first; without one, or when it is inconclusive, the agent decides itself. The detail moved into references read only when a judge is configured.
fnl-delegate: plain content writing goes to Codex first; a "Never delegated" rule lists what a delegate never does or decides and what the coordinator itself does only with the user's authority; a same-family review fallback asks the user first on a security-triggered pull request.
fnl-review-gate: severities are defined (CRITICAL, HIGH and MEDIUM block; LOW never causes another round), rework is capped at two rounds, and findings outside the change stay in the report.
fnl-security-review-gate: adds a data-leaving trigger, a short report when no trigger matches, negative tests only through entry points that reach the touched resources, an exploit scenario for MEDIUM or higher, a NEEDS-VALIDATION outcome that blocks until the owner validates a concern the repository cannot settle, and a rule that security-audit in focused mode supplies method, not findings, so a file counts as engine-only only when the engine says it read it.
fnl-open-source-review: reruns before each MINOR release, strongly recommends a secret scanner, caps the verdict at "not ready" while a secrets, leaks or license check is BLOCKED, and splits repository settings into blocking and noted.
fnl-verification: names the gap and proposes a test for an untested runtime change, proposes a full suite for widely shared code, points to the non-code reference, and lets only the owner waive a check.
fnl-commit-messages: a commit is one whole seam with a body of about six lines; agents commit when a seam passes its targeted tests, on their own branch, without pushing; pull request text stays short.
fnl-handoff: findings go back to the same run while it is light and to a new run after about 60% context or two failed rework rounds; a prose brief names language and tone, and the editing pass runs only for social content or documents.
fnl-advisor and fnl-committee: both are proposed and wait for the user's yes (an advisor may run without it while the user is away; a committee never does); tier follows difficulty; a same-family advisor stands in only when both other families are out; a committee adds a third member only after two rounds of disagreement, settles it with one reversible experiment, and waits for the user before its plan is dispatched.
fnl-coordinator-sessions: coordinate from three user stories or ten tasks; the coordinator writes no application code; standby covers the next story plus one fix session; independent stories still run one after another.
fnl-status: flags a group only from five runs (first-pass acceptance below 60%, or timeouts plus fallbacks at 25%); a coordinator closes an orphaned ledger record only when its log or the repository shows how the run ended.
fnl-knowledge covers two layers, the knowledge base and the repository, and leaves an agent's native memory to the agent (the per-host references/hosts.md is removed); a decision is written as soon as the user settles it, a lesson at the end of a task or at once after an incident, and a project-state entry carries a review-by date. fnl-setup no longer writes a native-memory pointer.
fnl-setup lets the user choose a detailed or a quick mode (the quick mode asks every decision once with its recommended default, and a skill that later needs an unsettled item asks the user once), reruns only when an agent is added or changed or the host template gains a field, grants Antigravity only exact test commands recorded in the host config and kept to one repository, and offers the low-risk lane line only to a repository with one owner. fnl-review-gate asks the user once to adopt the gate where the repository has not.
fnl-setup-instructions treats about 150 lines in the shared body as a soft limit and offers to move detail into a skill or reference, adds a contract line that lets an agent propose one entry (at most one per session, written only on the user's yes) after the user corrects the same behavior twice, and recommends the Claude Code file as the source, with an include where a host has one and a managed copy that warns about manual sync where it has none.
fnl-plugin-packaging runs the trigger suite only for skills whose description changed (the whole suite before a MINOR release), applies the icon section only when a directory asks for an icon, and says that a catalog tracking the default branch serves -dev builds that change with every merge.
fnl-ask adds a row for queries on live data, infrastructure changes, dashboards and forecasts, and notes on the advisor and committee rows that they are proposed first, wait for the user's yes and spend quota.
fnl-reviewer, fnl-security-reviewer and fnl-verifier: the two reviewers are always the entry and run the host's native review (/code-review, codex review) as their engine, with the effort set by depth. The security reviewer runs security-audit in focused mode on the diff when a trigger matches or the diff is large, and the native security review otherwise; where an instrument is missing it reviews by hand and tells the caller that /fnl-setup guides installing it. Reviewers use the CRITICAL to LOW scale, trace each hunk to an acceptance criterion, name engine-only coverage, and need a concrete scenario for a MEDIUM or higher security finding; fnl-security-reviewer runs on opus and fnl-verifier on sonnet. fnl-delegate uses the verifier for hard and very hard runs, and the coordinator checks easy and medium runs itself. fnl-setup gains a review tools step that strongly recommends installing missing instruments.
fnl-session-hygiene: a finished user story with its review gate closed is a boundary for a fresh session, knowledge is written back only when it would change what an agent does next time, and session titles use English for the activity unless the user asks for their own language.
fnl-versioning: an agent never cuts a release on its own and proposes one when a spec is merged, a spec usually bumps the last digit, a release is two pull requests, and changelog entries stay to one or two sentences.
What to do
Run /fnl-setup again: its new review tools step finds the review instruments the reviewer agents use (/code-review, /security-review, codex review, Cloudflare's security-audit) and recommends installing any that are missing.
If your host config names a route judge, nothing changes in it; the coordinator and the router now ask that judge first.