Immutable. This exact content is served forever at /api/v1/blob/58a21935bf2f552f.
--- name: agent-fleet description: >- Run a fleet of research/writer agents (codex sol/terra/luna at 1M context, claude opus) invar-conductor style from this repo: mechanical context preload (never "please read X"), two-step dispatch (brief then /goal gate), DONE CHECKLISTs with observed evidence, reports born as artifact files, conductor-side verification. Use whenever work fans out to multiple nested agents (venue research, article drafting, parallel investigation). Companion: the agent-tmux skill (the transport verbs); this skill is the ORCHESTRATION. --- # agent-fleet — conductor-style multi-agent runs in the ivue repo Adapted from invar's conductor skill (read `../invar/.claude/skills/conductor/SKILL.md` for the full doctrine when orchestration grows past one evening). This file carries what THIS repo needs: launch recipes, the preload law, the two-step dispatch, and the verification duties. Transport mechanics (send/state/peek/wait) live in `.claude/skills/agent-tmux/SKILL.md` — read it before driving sessions. ## Launch recipes (verified 2026-09-01) Codex, 1M context window (sol = strongest, terra, luna = cheapest; model economy: research on luna, hard synthesis on sol): ```bash codex --dangerously-bypass-approvals-and-sandbox \ -m gpt-5.6-sol \ -c model_reasoning_effort=high \ -c model_context_window=1000000 \ -c model_auto_compact_token_limit=900000 ``` Claude opus with IBR as the SYSTEM PROMPT (the strongest seat — same shape the user runs the main session in): ```bash claude --dangerously-skip-permissions --model opus \ --system-prompt USE_IBR_FOR_REASONING \ --append-system-prompt-file=.claude/skills/ibr/IBR.md ``` Sessions are tmux, named `ivue/<agent>`, launched detached (`tmux new-session -d -s ivue/<name> -x 220 -y 50 -c <repo> "<cmd>"`). The worked dispatch script: `scripts/fleet-dispatch.sh` beside this file — read it before writing a new one; it encodes every lesson below. ## THE PRELOAD LAW — dump mechanically, never instruct to read Agents skip reading when told "read X first" — sometimes heavily. So context is INJECTED, not requested: - Build one deterministic preload file per agent by CONCATENATION (common brief + angle brief + inventories + IBR + the ivue skill), fixed section order, headed `# PRELOADED SECTION n/N`. - **argv is capped at 128KB per single argument on Linux** (E2BIG hit live at 147KB) — a big preload cannot ride the CLI positional prompt. The channel that has no cap: launch the agent bare, then `tmux load-buffer -b <buf> <file>` + `tmux paste-buffer -d -b <buf> -t <session>`. - **After a huge paste the TUI can stay INSIDE bracketed-paste mode**: every later keystroke (Enter included) is eaten as paste content — observed live: each Enter appended another [Pasted Content] chunk. Close the paste explicitly with the bracketed-paste END sequence, then submit: `tmux send-keys -t <s> -H 1b 5b 32 30 31 7e` → sleep → `C-m`. - Verify submission by the BUSY FOOTER appearing (`esc to interrupt`), never by the text being visible (an unsubmitted composer looks identical — the agent-tmux ladder). - For claude agents, IBR goes in via `--append-system-prompt-file` (stronger than in-turn) and the rest of the preload via paste. ## Two-step dispatch (invar law): the BRIEF, then the GOAL 1. **The brief** is the preload's head: `## In plain words` (2–3 sentences a ten-year-old could follow), the task, a STRICT output format (machine-mergeable tables), the agent's LANE (disjoint angles per agent — overlap only when convergence-validation is wanted), and `## THE DONE CHECKLIST` — numbered checkpoints (CP1..CPn), each requiring OBSERVED EVIDENCE (counts, strings read back) in a `## Checkpoint report` table in the output file. Honest NOT DONE rows accepted; silent omissions are not. 2. **The goal**, sent SEPARATELY after launch: `/goal You may not write DONE until EVERY checkpoint ... passes` pointing at the brief's checklist (one source of truth — never restate the checklist in the goal). Sent second so it's what the agent holds when deciding whether to stop. `/goal` works in BOTH codex and claude CLIs. 3. Goal delivery while the agent is mid-turn: codex queues with **Tab**; claude queues with **Enter**. VERIFY: codex shows `shift + ← edit last queued message` / a `↳` marker; claude shows `Press up to edit queued messages`. Text visible in the composer proves NOTHING (it sat unsubmitted on both claude agents tonight until an explicit Enter). ## Reports and verdicts - Every agent writes its report to a DURABLE artifact file named in the brief (`tmp/press-briefs/out-<name>.md` tonight); last line is the literal `DONE`. Verdicts come from artifacts, never from panes. - Poll with `tmux capture-pane` for the busy footer + check the out file's existence and tail. An interactive agent never exits when finished — it goes idle. - **Conductor verification is mandatory before using results** (two-arms): row counts vs the checklist's claims, spot-check a sample of URLs yourself, dedupe across agents, and treat every delegated finding as a HYPOTHESIS until checked. Convergence between independent agents is supporting evidence, not proof. ## The wave pattern Fan-out happens in WAVES, synthesis between them, conductor builds while agents run: wave 1 discover (research lanes) → conductor merges + verifies → wave 2 produce (writers get the merged map + house voice skills preloaded: write-article title doctrine, ste-expression, blog voice) → conductor assembles the deliverable (data + UI + docs). Never idle-wait on a wave: the conductor's own build work (components, skills, synthesis scaffolding) runs between polls. ## Hygiene - `tmux kill-session -t ivue/<name>` after reaping; never leave a fleet running unattended past its task. - Preloads/briefs/outputs live under `tmp/` (gitignored); the skill and dispatch script are the committed knowledge. - Steering mid-flight: one small message via the queue mechanics above; a file is not a channel after launch (agent-tmux law). - Model quota: codex sessions are cheap; claude opus sessions bill the interactive bucket — spawn only what the task earns.