Flip radio wake defaults: workers auto-submit and stay in the background; the PM box is never auto-submitted

[martin-conur]

Closes #254

What changed

	before	now
task-pm	auto-submit on (#189)	off; --auto-submit / --auto opts in
task-work (claude + kiro, --plan too)	auto-submit only with --auto / --auto-submit	on; --no-auto-submit opts out
focus after task-work / task-recreate-worker	jumps to new tab unless --auto	stays on caller; new --focus jumps
--auto	permission mode + auto-submit + focus	permission mode only (claude); no-op on kiro

#246's tri-state is unchanged: every launcher still injects an explicit TASK_FORCE_AUTO_SUBMIT=0/1, so a relaunched PM drops the 1 it recorded under #189 (pinned by a test). The PM/planner prompts already say plain task-work <slug> <url>, so they needed no edit. They now get the wanted behaviour with no flag.

Risk from the spec: --no-auto-submit is documented as the opt-out for a worker you mean to step into and type in (task-work --help, the workflow docs, README).

Docs: README, the seven steering/*.example.md workflow docs and this repo's .claude/gh-workflow.md: the flag lists, the PM wake paragraph (inverted), and the "radio check sitting unsubmitted" symptom row (inverted). CHANGELOG has an entry. Upgrading: re-run task-init <loadout>. Live sessions keep what they registered with.

Verification (working tree = this PR's diff over cc0523d)

• ./run_tests.sh: 1321/1321 ok (bats, full suite).
• tools/check-drift.sh: 15 groups clean. shellcheck -x clean on every changed shell file.
• New and updated tests feed the launch env to radio register and read AUTO_SUBMIT back from the session file, for task-work on every impl and for task-pm on claude and kiro. Focus is pinned for the default, --plan, --auto and --focus paths, and for task-recreate-worker.
• Mutation check: I broke each of the 9 below one at a time on the working tree and re-ran the named suites. Failing-test counts:
  • task-work default auto-submit → off: task_work_agents 8, task_work_impls 1, radio_runbook 1
  • task-work --no-auto-submit ignored: 6 + 1
  • task-work focus default → jump: zellij_tab_stay_focused 3, task_work_agents 1
  • task-work --focus ignored: 2
  • task-pm default → on: task_pm 4
  • task-pm --auto-submit ignored: 3
  • task-recreate-worker focus default → jump: 1
  • task-recreate-worker --focus ignored: 1
• Not covered: no live zellij session was driven. Focus is verified only against the stubbed zellij action call sequence.

🤖 Generated with Claude Code

[martin-conur]

Spec compliance

Issue #254 had four deliverables; all are present:

Requirement	Status
task-work: auto-submit on by default (all agents, --plan included)	✅
task-work: focus stays on caller by default; --focus jumps	✅
task-pm: auto-submit off by default; --auto-submit / --auto opts in	✅
--auto is permission mode only on claude; no-op on kiro's task-work	✅

Code correctness — the logic flip in bin/task-work:
Old path wired AUTO_SUBMIT to AUTO_MODE, so a plain task-work injected TASK_FORCE_AUTO_SUBMIT=0. The new path inverts: =1 is the default and only AUTO_SUBMIT_FLAG="0" (from --no-auto-submit) flips it to =0. Injection is always explicit (never unset), which is what the #246 tri-state requires.

task-pm — safe transition across a live sidecar:
The old code set AUTO_SUBMIT=1 unconditionally. The new default is "" (→ TASK_FORCE_AUTO_SUBMIT=0). The spec's risk note is addressed: a PM launched under the old default wrote AUTO_SUBMIT=1 into its sidecar, and radio reads an unset env as "restore sidecar" — so a bare relaunch without the explicit 0 would silently re-enable auto-submit. The PR exports TASK_FORCE_AUTO_SUBMIT=0 (never unset), which overrides the stale sidecar correctly. The new test claude task-pm's default overrides a PM that recorded auto-submit under #189 (#254) pins this end-to-end through the session file.

Tests — end-to-end, not just env injection:
The new _session_auto_submit() / _pm_session_auto_submit() helpers register against the exact env the launch line injects and read AUTO_SUBMIT back from the session file. This catches a regression where the env is injected but radio register ignores it. Mutation check covers all 9 breakage scenarios with specific failing-test counts.

Docs — all required surfaces updated:
README (PM auto-submit paragraph, kiro section, symptom table), .claude/gh-workflow.md, all 7 steering/*.example.md templates, CHANGELOG with upgrading note. The radio check sitting unsubmitted symptom row is correctly inverted for both PM (default now) and worker (non-default).

---

Code-review findings

Nit — task-recreate-worker --auto description is accurate but could confuse a reader who just learned task-work --auto is a no-op on kiro:

In both .claude/gh-workflow.md and kiro-gh/steering/gh-workflow.example.md, task-recreate-worker's --auto option still reads: "on kiro it governs auto-submit only (#206)." That phrasing echoes the old task-work language. For task-recreate-worker it is technically correct — the command's default is sidecar-restore (not auto-submit-on), so --auto actively sets TASK_FORCE_AUTO_SUBMIT=1 on kiro even after #254 — but a reader who just processed "task-work --auto is a no-op on kiro" will be surprised. A one-clause parenthetical like "(unlike task-work, whose default is already on)" would close the gap. Not a blocker.

Nit — kiro --auto-submit flag is missing from task-recreate-worker's kiro workflow doc:

The kiro-gh template's task-recreate-worker option list was updated to add --focus but still omits --auto-submit from the documented options. --auto-submit exists and is tested — it just isn't listed. The claude-gh template documents all three variants; the kiro one only shows --auto and --auto-submit / --no-auto-submit in the descriptive text but not as list items. Minor gap.

---

Verdict: clean-with-nits

The PR delivers everything #254 asked for. Both code paths (auto-submit flip in task-work, explicit-off in task-pm) are correct, the tri-state invariant is preserved, tests are thorough and end-to-end, and all required doc surfaces were updated. The two nits above are documentation clarity issues — neither affects runtime behavior or tests. Merge when ready.

[martin-conur]

PM: fix both of the reviewer's doc points in this PR (tight-PR norm), then re-request review:

1. task-recreate-worker --auto description (.claude/gh-workflow.md and the kiro templates): add the clause that sets it apart from task-work. Something like "unlike task-work, whose default is already on, this command defaults to the auto-submit setting radio recorded for the role".
2. --auto-submit missing from the kiro task-recreate-worker option list. Check all three kiro templates (kiro-gh, kiro-notion, kiro-local), not just the one the reviewer cited, and make the option lists match the claude templates.

Docs only. Confirm a CI run exists for the new latest commit.