Practice skill: spawn one-shot workers unnamed; shut teammates down when their result is sealed - #1142

Scope

• What this PR changes, in one or two sentences: the practice skill's workers rows (subagent-definitions, agent-teams) now state the agent lifecycle from the official docs at 2.1.300. One-shot workers spawn without a name and resume by agent id. A name is only for a worker that must message others. Each teammate is shut down once its result is sealed, because an idle teammate stays running and addressable.
• Base commit: 0f4ee075b
• Lane: lane:foundation
• Owned paths touched:
  • .claude/skills/claude-native-practice/reference/slots.json and workers.md;
  • the slot manifest;
  • evidence/artifacts/agent-lifecycle-20261015/sources.json (new) and tests/test_native_practice_agent_lifecycle.py (new);
  • manifests/evidence.json.

Why: the command center found 48 idle evidence-reviewer teammates in its session (2026-10-15). Each was a one-shot micro spawned with a name, and the command center stopped them. The owner item of 2026-10-15, paraphrased by the command center, is to resolve the agent team's full lifecycle.

SOTA sources

Official Claude Code docs at client 2.1.300, fetched 2026-10-15T04:08:42Z. Each page's sha256 and the quoted lines are in sources.json.

• agent-teams:
  • :60 and :430: while agent teams are enabled, a subagent that Claude names launches as a teammate, and Claude may name a subagent on its own.
  • :62: -p and Agent SDK sessions keep it a subagent.
  • :86 and :420: an idle teammate's row hides, but the teammate stays running and addressable.
  • :202-210: the lead ends a teammate with a shutdown request.
• sub-agents :937-939: a name makes a subagent addressable. In an interactive session with teams on, a named spawn from the main conversation launches as a teammate, unless it is a fork or passes isolation on the call.
• hooks :2853-2858: TeammateIdle with exit 2 keeps a teammate working, and {"continue": false, ...} stops it.
• anthropics/claude-code CHANGELOG at 94506ff0, 2.1.178: with CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1, every session has one implicit team, and the Agent tool's name spawns a teammate.

Evidence-class table

Local commands run

$ python3 -m unittest tests.test_native_practice_agent_lifecycle    # before the change
FAILED (failures=10, errors=1)
$ python3 -m unittest tests.test_native_practice_*  (8 modules)
Ran 105 tests; OK (skipped=3)
$ python3 scripts/validate.py
rc 0; {"components": 71, "hashed_files": 12041, "profiles": 4, "receipts": 240, "status": "passed"}

Decision record

The slot rows are the record (dated notes on agent-teams). The host switch question (whether CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS stays on host-wide) is the owner's call. The measured case goes to the command center separately. A TeammateIdle hook design follows as a reviewed host diff.

Fix round 1 (c28c14b)

From GPT full1 at b4c4bba (full1-pr1142/GPT-VERDICT-b4c4bba3b.md), blocker 1 (P1). The command center released this commit at 2026-10-15T10:49Z.

python3 -m unittest tests.test_native_practice_agent_lifecycle tests.test_published_digest_privacy tests.test_provenance_digest_custody: OK (1 skip). scripts/validate.py: rc 0. The earlier commit keeps the removed field in its history; no history rewrite.

Follow-ups

• GPT P3: some detailed lifecycle citations and native observations remain partly unverified.
• GPT P3: the text guard records, but does not enforce, shutting a teammate down once its result is sealed.
• GPT P3: refresh the body's run receipts for the reviewed head.

Checklist

• No GitHub Actions changed.
• No workflows changed.
• No secrets are printed, logged or committed.
• No new paid hosting, subscription or billing surface.
• Peer-owned untracked files and worktrees were preserved.

🤖 Generated with Claude Code