footnote

Dispatch and message background agent workers from a runner-less surface (phone / Happy app). One router over the shipped `fno agents` mesh: spawn (launch a build worker), handoff (continue a doc without re-deriving a plan), discuss (open a daemon-managed chat thread), send (message a peer over the bus), watch/list/logs (observe), stop (terminate). Normalizes messy input (smart quotes, name, provider, mode), confirms billed launches, and reports the real receipt - never a fabricated one. Works for claude, codex, and gemini (handoff/discuss are claude-only in v1). Use when: 'spawn a worker for ab-XXXX', 'hand off this doc to a worker', 'open a discussion thread', 'send a peer a message', 'run /target in the background', 'ask codex <question>', 'list/watch/stop my agents'.

Loaded at every SessionStart so the agent knows the two footnote surfaces (slash-command workflows + fno CLI primitives) exist from turn one. Mirrors the using-superpowers pattern.

Create implementation blueprints (plans) as a single .md doc. Given a /think design doc, mutate it in place; given an idea, create one. 'quick' scales the sections down for bugs and 1-session work; every plan is one .md == one PR == one node. Use when: 'create plan', 'implementation blueprint', 'break this down', 'how should we build'.

Use when: build this feature, get it done end-to-end, or execute a plan from idea to PR.

Execute a plan. Routes between a lightweight single-session executor (flat, default) and full wave orchestration (waves, alias operator). Use when: 'do this plan', 'execute the plan', 'run the waves'.

Reason about a design before building. Routes to design+BDD exploration (think, default), scenario/failure-mode stress testing (what-if), or a multi-persona expert debate (panel). Use when: 'think through this', 'brainstorm', 'what could go wrong', 'stress test this idea', 'convene a panel', 'should we build X'.

Drive a PR through its lifecycle. Routes to create (open a PR via a Haiku worker), check (poll for external review and implement it), or merged (the post-merge ritual). Use when: 'create pr', 'open pr', 'submit pr', 'check pr', 'get review', 'post merge', 'process the merged PR'.

Propose and apply optimal ordering for pending specs on the graph backlog. LLM proposes, human approves. Use when: 'triage the backlog', 'what should I work on next', 'reorder my specs', 'find duplicates in graph'.

Review a diff or a research brief. Routes to the internal six-agent Claude panel (sigma, default), a cross-model second opinion (peer), the advisory research-verify panel for a doc deliverable (research), or a self-cert attestation for the config.review.reviewers gate (declare). Use when: 'review this', 'code review', 'is this ready', 'get a second opinion', 'have codex review this PR', 'review this research brief', 'declare this reviewed'.

Multi-perspective feature completeness analysis and planning loop. Use when: 'audit feature', 'what's missing', 'feature completeness', 'gap analysis', 'discover features to build', 'plan all features'.

Send and receive cross-project messages between agents in different projects (send via `fno mail send --to-project`, drain/read via `fno mail`). Use when an LLM in one project (e.g. acme-backend) wants another project's agent (e.g. acme-web, acme-docs, acme-blog, marketing) to know about or react to something. Recipient daemons drain mail autonomously; only kind:question reaches a human.

Schema-driven setup wizard for footnote settings. Derives its question set from the Pydantic config model (the single source of truth) instead of a hardcoded list; asks only the real per-project decisions and defaults the rest.

Generate and maintain project documentation - architecture docs, how-to guides, API references. Use when: 'document this', 'architecture docs', 'API contract', 'runbook', 'user guide', 'how-to', shipping documentation for a feature.

Run N parallel variations of the same feature for comparison. Use when: exploring design alternatives, comparing architectures, A/B testing implementations, 'give me 3 takes on this'.

Test-Driven Development: Write test first, watch it fail, implement minimal code. Use when: implementing any feature or bugfix, before writing production code, during /do workflow, whenever acceptance criteria exist.

Repair a broken state. Routes to the fast one-fix-per-iteration loop with auto-revert (fix, default) or the scientific-method hypothesis loop (investigate). Use when: 'fix all errors', 'make tests pass', 'fix the build', 'debug this', 'investigate this failure'.

Drive any deliverable to its finish line. The umbrella over delivery terminals: 'ship pr' is the PR lifecycle (= today's /pr), 'ship doc' ships a research brief to output_dir and grades it. Use when: 'ship this', 'ship a PR', 'ship the doc', 'ship the brief', 'deliver this'. Not for ongoing areas (budget, community) - those have no finish line; use /target or /megawalk.

Self-improvement loop for the toolkit. Captures corrections passively (git post-commit, pre-commit verifiers, /insights tags) into ~/.fno/corrections.log, then runs a monthly review via a fresh Claude API call against the current rule text and surfaces patches the user triages in 20 minutes. Use when the user asks to review corrections, run an autocorrect review, triage proposed patches, install the autocorrect schedule, ingest /insights, check autocorrect status, or audit recurring mistake classes in the toolkit.

Keep prompt cache alive during idle. Prevents 10x cost spike when cache expires. Use when: 'keep cache warm', 'cache keepalive', or auto-activated at session start when project opts in.

Message background agent workers and projects from a runner-less surface (phone / Happy app). One front door over the shipped `fno mail` durable mailbox: send (message a peer or a project), reply (correlated response), unread / list / view / status (read your inbox), ack (advance your read cursor), drain (batch-consume at a loop boundary). Normalizes messy input (smart quotes, recipient, body), refuses an empty recipient or body before writing anything, runs the genuine `fno mail` command, and reports the real msg-id receipt - never a fabricated one. Messaging is free and async, so it never confirms. Use when: 'send tgt-foo a message', 'mail target about X', 'check my unread', 'reply to msg-abc', 'tell everyone on project Y', 'what's in my inbox'.