Add super-tasker skill (opt-in, with responsive-agent dep)

[asdigitos]

Summary

Adds two opt-in skills at repo-root skills/:

• skills/super-tasker/ — file-backed multi-track task orchestration.
• skills/responsive-agent/ — pause-and-ask / resume protocol (dependency).

Neither is bundled into ductor_slack/_home_defaults/, so a fresh ductor
install does not pick them up automatically. Install is a deliberate
copy into the live workspace (see skills/README.md).

super-tasker turns the agent into a Lead that walks a file-backed
frontier of tasks and dispatches Executor sessions per leaf via
tools/task_tools/create_task.py. It builds on responsive-agent for
human-in-the-loop continuation — no in-loop budgets, no self-trigger.

What changed since the previous round

Three rounds of review feedback from Victor are folded in:

1. Opt-in, not auto-install. Skills moved out of
   ductor_slack/_home_defaults/workspace/skills/ into top-level
   skills/. skills/README.md documents the manual install path.
2. Activation model corrected. Self-messaging does not re-enter a
   ductor session, so the previous "cron sends self a [super-tasker:lead-pass]
   message" framing was wrong. SKILL.md now enumerates the four real
   external triggers for a lead pass:
   • user Slack message,
   • executor completion callback injected into the parent session by
     create_task.py,
   • responsive-agent resume,
   • cron / webhook (an external scheduler firing into the agent, not
     the agent talking to itself).
     The doc explicitly states the agent cannot wake itself.
3. Split / merge are one-way state transitions. Made explicit in
   SKILL.md: once a → a1, a2, parent a is permanently Splitted and
   stops being tracked; active set replaces it with {a1, a2}. Once
   b1, b2 → b3, sources are permanently Merged and stop being tracked.
   No rollup, no revival. The lead walks the frontier (non-terminal
   tasks whose split_from parent is terminal-or-absent), not "top-level".
4. resume.py surfaces the executor anchor on the first output line.
   Previously the resumed agent had to scan the body for [stask-exec:<id>]
   to know it should re-enter executor mode. Now resume.py does that
   scan itself and re-emits the anchor as the first line of stdout.
   [stask-exec:<id>] is also the canonical anchor in spawn_executor.py
   and SKILL.md (the old [super-tasker mode=executor task=...] form is
   gone).

File layout

skills/
├── README.md
├── responsive-agent/
│   ├── SKILL.md
│   └── scripts/{cleanup.py,pause_and_ask.py,resume.py}
└── super-tasker/
    ├── SKILL.md
    └── scripts/{tasker.py,spawn_executor.py}

Design decisions worth review

• Frontier vs. top-level. is_top_level (strict: no SplitFrom parent,
  not in any merged_from) is kept for the user-facing "list my major
  initiatives" view. is_frontier (non-terminal AND parent is
  terminal-or-absent) is what the lead pass walks. Both predicates are
  exposed via list --top-level and list --frontier.
• Re-evaluation cadence: triggered, not every-pass. Re-eval fires when
  one of: executor appended reeval-request, task was never planned,
  user-note contains a reeval intent, or task is Open for ≥ STASK_STALE_DAYS
  (default 3) with no events. Rationale: each re-eval is a subagent call;
  sweeping every Open task every pass is expensive and noisy.
• MAX_DEPTH = 3 — enforced in tasker.py (depth-3 tasks reject split).
• No plan-before-execute confirmation gate. User intervenes
  asynchronously via user-note events; the lead reads them on the next
  pass. The intervention verb table is in SKILL.md.
• Executor scoping. spawn_executor.py builds a prompt that lists the
  allowed read window (task.json + events.jsonl + context/ + the absolute
  paths in context_files) and rule 1 forbids reading the wider workspace.

Status

Design review, round 2. Not for auto-install — files live at repo-root
skills/ precisely so merging does not push them into anyone's
workspace. Next step after approval: copy into Ethan's running ductor
instance and exercise on a real task.

Test plan

• tasker.py init → add → brief returns the new task as frontier
• split creates depth-1 children; parent becomes Splitted and is
  excluded from the frontier; children become the frontier
• depth-3 task rejects further split
• merge on terminal task is rejected; merge on two Open tasks
  creates a result and sources go to Merged
• --text works before or after the subcommand
• resume.py injects [stask-exec:<id>] as the first stdout line when
  it appears in the saved context (smoke-tested locally)
• Install on Ethan's ductor instance and run a real lead pass