Implemented trailboss CLI with jump-next and skip commands that
navigate to panes returned by the daemon's /next endpoint.
- bin/trailboss: main CLI with jump-next, skip, popup stubs
- jump-next: queries /next, navigates via switch-client/select-window/select-pane
- skip: POSTs /skip, then jumps to new head
- test-navigation.sh: synthetic event test verifying exit criterion
Exit criterion passed: jump-next lands operator on pane returned by /next.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Phase 1: Probed PermissionRequest hook, confirmed firing + payload shape.
Documented in docs/research/claude-code-mechanics.md.
Phase 2: Built trailboss-emit.sh that forwards hook payloads to collector
and injects $TMUX_PANE. Wired all hooks in .claude/settings.json.
Verified: stub collector logs POSTs with session_id, cwd, and pane_id.
Demonstrated bare curl drops $TMUX_PANE, proving wrapper required.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The launcher sourced env into its own process, but the loop runs in a separate
tmux shell that didn't inherit ANTHROPIC_BASE_URL/AUTH_TOKEN — so claude ran
unauthenticated ("Not logged in"). Source the env file inside the loop command
so the proxy/model config is present in the shell that launches claude.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
The plan defined deliverables but no definition of done. Add a "Testing &
validation" section: a per-phase exit-criteria table, seven acceptance
scenarios (AS-1..AS-7, incl. reconcile self-correction, dropped-event
recovery, skip/cooldown, no-forced-focus-steal, pane reuse), a test harness
(synthetic event injection, throwaway-tmux isolation, transcript fixtures,
mock hook emitter, navigation assertion, invariant checks), and a quality
gate. Point the marathon instruction at these as the definition of done.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Model the setup on existing marathon instances (FORGE/AIVC/NEEDLE):
- .marathon/instruction.md — per-iteration autonomous prompt, plan-driven off
docs/plan/plan.md with the 7-phase order, settled invariants, and commit/push
loop protocol
- .marathon/launch-marathon.sh — self-contained `while true` loop piping the
instruction to `claude --print`, sourcing model/proxy config from an
untracked .marathon/env (keeps internal endpoints out of this public repo)
- .marathon/env.example — config template
- .gitignore — ignore .marathon/env, logs/, *-config/, and .beads/*.db
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Add a "Switching & keybindings" section: global prefix bindings that call the
daemon via run-shell — Next (jump to head-of-queue pane), Popup (display-popup
picker), Skip (tail+cooldown then advance) — plus a status-line segment.
Documents the prefix-not-bare-key constraint and why the jump is the keypress,
never automatic. Resolves open questions 3 and 4 down to polish.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- MEDIUM: reason enum stated as exactly {permission, stopped} (drop stray
"stopped/needs-next" third label)
- MEDIUM: specify oldest-READY-stuck-first — a cooldowned skipped item is not
eligible to be the head; advance to next ready item, else present empty
- MEDIUM: SubagentStop downgraded from "Confirmed" to "exists; not probed"
(probe only saw SessionStart/Stop)
- LOW: agent-inbox reuse reworded (Trail Boss actions are navigate-and-reply /
skip, not accept/edit/ignore); note permission_mode is display-only
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
- HIGH: settings.json example wired PermissionRequest to a bare curl that drops
the env-only $TMUX_PANE; route all hooks through trailboss-emit.sh (which
injects $TMUX_PANE) and note SessionStart is not special
- MEDIUM: remove the stray trailboss-register.sh (contradicted "registry
self-heals on every emit"); single emitter script
- LOW: add struck-through SubagentStop row to plan detection table; clarify
disler's store is append-only (not "read-only")
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Resolve open questions from the design session:
- Stuck is stuck: no permission-vs-stopped priority; reason is display-only;
queue is a flat FIFO dead-letter queue (Stop AND PermissionRequest still
both required — permission blocks emit no Stop)
- Drop Notification entirely
- Auto-advance depletion loop: next stuck session loads on resolve/skip;
saturation is a non-issue by construction
- New primary open question: harness-coupled detection vs harness-agnostic
core, via a normalized stuck/unstuck adapter contract (switching is already
tmux-level/harness-agnostic)
- Reboot: operator re-invokes manually (no auto-resurrection in v1)
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Emphasize that agents run autonomously and the human is engaged only by
exception — the failure-mode handler / dead-letter-queue processor for a fleet
of agents, not a gate on every step.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Single-pane attention router for interactive AI coding agents: stuck sessions
surface in one prioritized queue (most-stuck first) via Claude Code hooks, and
your reply is delivered back into the exact session.
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
