trail-boss/docs/notes/decisions.md
jedarden 675c1e531d docs: gap-review round 1 — reconcile supporting docs + close spec gaps
Fix 10 gaps found by fresh-eyes review:
- CRITICAL: mechanics doc still documented send-keys-relay/Agent-SDK delivery
  and the Notification hook; rewrote DELIVER to the navigator model, dropped
  Notification throughout
- HIGH: decisions.md + related-work.md still referenced priority "ranking",
  /reply dispatch, send-keys/SDK delivery, and the dead plan/question/idle
  reason taxonomy; aligned to FIFO + navigation + permission/stopped
- collector and daemon stated as one process; pull-not-broadcast presentation
- added trust boundary (loopback-only ingest, single-host trust assumption)
- resolved auto-advance focus-steal hazard (operator-initiated jump) and
  specified skip re-ordering (move to tail + cooldown) to avoid livelock
- downgraded SessionEnd to "not yet probed" (probe only saw SessionStart/Stop)
- dropped forked collector's WebSocket layer; most-stuck → oldest-stuck (FIFO)

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
2026-05-25 15:18:56 -04:00

90 lines
5.3 KiB
Markdown

# Decisions & rationale
## Naming
**Trail Boss** — on a cattle drive, the trail boss is in overall command: sets the direction,
makes the calls, and rides in when a steer bogs down or strays. The product runs a *herd* of
agent sessions; when one gets stuck it reports in, and you — the trail boss — ride over and set
it right. The metaphor maps cleanly onto the mechanism:
- **the herd grazing the range** → sessions working autonomously
- **a steer bogs down or strays** → a `Stop` / `PermissionRequest` hook fires; the collector
flags the session stuck
- **the trail boss rides over and sets it right** → you read the context and give the order
(reply) or wave it on (skip); the queue surfaces stuck sessions oldest-first (flat FIFO, no
priority ranking)
### Names considered and rejected
- **`agent-inbox`** — clearest literal description, but collides head-on with
[`langchain-ai/agent-inbox`](https://github.com/langchain-ai/agent-inbox), an existing
human-in-the-loop inbox for LangGraph agents. Would read as derivative and lose every search.
- **`agent-attention`** — names the value prop (your attention is the scarce resource being
routed), but risks reading as the ML "attention" mechanism.
- **`agent-central`** — self-explanatory but generic, and "central" reads like a passive
dashboard/hub rather than an act-on-the-stuck-one tool.
Trail Boss keeps a memorable, distinctive identity; the tagline carries the legibility for
newcomers.
## Design decisions
### Hooks, not polling
Detection is event-driven via Claude Code hooks. A session emits a signal the moment control
returns to a human: while actively working it emits `PreToolUse`/`PostToolUse`, never `Stop`.
A session counts as waiting only once `Stop` or `PermissionRequest` has fired and no
`UserPromptSubmit` has come since. **Confirmed by probe (2026-05-25):** both `SessionStart` and
`Stop` fire in interactive and `-p` modes, the `Stop` payload carries `last_assistant_message`
(queue context for free), and hook commands inherit the ambient environment.
### Stuck = needs attention, and stuck is stuck
A session that has stopped or is waiting at a permission prompt cannot progress until the human
responds — so it needs intervention by definition. Two collapses follow: there is no
"finished but fine" state (every stop is a queue item), and there is no permission-vs-stopped
priority (it doesn't matter *why* it's stuck). `Stop` and `PermissionRequest` are both required
detection triggers — a permission-blocked session is mid-turn and emits no `Stop` — but they're
treated identically; `reason` is display-only and the queue is a flat FIFO dead-letter queue.
`Notification` is dropped (it adds nothing those two miss). The operator simply depletes the
queue, and the next stuck session auto-loads.
### Navigator, not relay (the delivery model)
Trail Boss routes *attention*, it does not inject *input*. Sessions stay as long-running live
CLIs in tmux panes (Model A), and delivery happens by navigating the operator to the live pane
(`switch-client`/`select-window`/`select-pane`, optionally `link-window` to co-display) where
they interact with the real prompt directly. This dissolves the send-keys fidelity problem,
makes "edit before allow" native (you just type), and means no synthesized input ever reaches a
session.
Rejected delivery alternatives:
- **Resume-to-deliver** (`claude --resume <id>` in a second process): a live interactive CLI
holds in-memory state and does not re-read its transcript, so a resumed process's reply never
reaches the original pane; concurrent attach risks transcript divergence. `--fork-session`
confirms plain `--resume` reuses the session. Only viable in a no-resident-process model
(Model B), which we rejected for v1.
- **`send-keys` relay as the primary path:** retained only as a secondary plain-text option
(basic submission confirmed working); native interaction is preferred.
- **`claude --remote-control`:** routes to the claude.ai / desktop / mobile surface, not a local
channel — useless for a same-host tool.
- **Agent SDK `canUseTool` (Model B):** programmatic permission gating with `updatedInput` is
attractive, but requires running sessions under the SDK instead of the terminal — deferred;
the tmux-navigator model fits the existing workflow and the durability requirement.
### Same-host daemon, durable via tmux
Trail Boss does not need to live *inside* tmux to drive it — tmux is client/server, so any
same-user process issues `tmux` commands to the server (pane ids are server-global). The
control plane is an always-on daemon; presentation is transient (`display-popup` + keybinding).
But for durability across SSH disconnect the daemon must survive SIGHUP, so it runs **in its own
tmux window** (simplest) or under **`systemd --user`** (also survives reboot; tmux does not).
Agents already persist because the tmux server is host-side. While disconnected, the daemon and
hooks keep running, so the queue accumulates the backlog and disconnecting becomes a non-event.
### The transcript is ground truth
Hooks are a low-latency notification; the transcript JSONL is authoritative. A reconcile loop
corrects dropped hook POSTs, daemon restarts, and "answered directly in the pane" by checking
whether a session's transcript has advanced past its last `Stop`.