Combines the best of both open PRs, resolving their overlapping approaches
and applying all 6 fixes that are not yet in main:

**mailbox.py**
- broadcast(): resolve logical agent names to inbox directory names before
  exclusion check, so the sender is not delivered their own broadcast when
  inboxes use user-prefixed format (e.g. "alice_worker" vs "worker")
  [from PR HKUDS#15]
- receive() / peek(): replace bare list comprehension with _parse_messages()
  helper that wraps each message in try/except, so a single corrupted file
  no longer crashes the entire fetch [from PR HKUDS#9]

**transport/file.py**
- fetch(consume=True): atomically rename msg file to .consumed before
  reading, so concurrent callers cannot both receive the same message
  [from PR HKUDS#15]

**cli/commands.py**
- config set / config get: derive valid_keys from ClawTeamConfig.model_fields
  instead of a hardcoded 3-item set, making all 7 config keys accessible;
  handle boolean fields (skip_permissions) correctly [from PR HKUDS#15]
- approve-join: return an error and exit(1) when no matching join request
  is found, instead of silently fabricating an agent name [from PR HKUDS#9]
- launch_team: read skip_permissions from config and pass it to be.spawn(),
  matching the behaviour of the standalone spawn command [from PR HKUDS#15]

Note: the plans-cleanup fix (PR HKUDS#9 subdirectory approach, PR HKUDS#15 member-filter
approach) was already merged via PR HKUDS#6 using the subdirectory strategy, so it
is intentionally excluded here.

All 135 existing tests pass.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>

- Fix approve-join proceeding even when no matching join request is
  found, which silently added members with fabricated names. Now
  returns an error and exits.
- Fix receive()/peek() crashing on any corrupted message, which also
  caused valid consumed messages to be lost. Invalid messages are now
  skipped gracefully via per-message try/except.

Note: the plans cleanup fix was dropped from this PR as it was already
resolved upstream via PR HKUDS#6.

Made-with: Cursor

Fix team cleanup deleting plans from other teams

- 8 failing tests covering ArtifactFrontmatterBase field shape,
  register/get/reset/list registry helpers, Pitfall HKUDS#6 None vs
  validation-failure distinction, and PhaseRegistry D-03 collision rule
- Plugin-manager integration test proves Plan 02-01's lazy-import
  guarded loop will wire end-to-end once the registry module lands (GREEN)
- All tests fail at RED with ModuleNotFoundError on clawteam.harness.evidence_schemas

…se pydantic base

- ArtifactFrontmatterBase (pydantic BaseModel) with persona / step_label /
  done / artifact_type / created_at — shape parallel to TurnEnvelope
  (§02-CONTEXT lesson HKUDS#2: one model, two serialization surfaces)
- Module-level _REGISTERED dict + register_schema / get_schema /
  list_registered / reset_registry accessors — structural twin of
  PhaseRegistry (§02-CONTEXT lesson HKUDS#3)
- Duplicate artifact_type raises ValueError naming both existing and
  attempted classes (§02-CONTEXT D-03 collision rule)
- get_schema returns None for unregistered names — Pitfall HKUDS#6 caller
  distinguishes this from pydantic ValidationError
- Activates Plan 02-01's lazy-import guarded loop end-to-end; the Phase 0
  regression matrix stays 12/12 green

…st_command + deploy_url

- 13 failing tests cover presence, frontmatter schema, stub detection (sections + blacklist), test-report rerun + cache-hit + exit-1 + timeout, and ship-notes deploy_url 2xx/non-2xx/URLError paths
- Test-local schema fixtures (DesignDoc/TestReport/ShipNotes) inherit from ArtifactFrontmatterBase; Phase 3 GstackSprintPlugin ships the real subclasses
- Pitfall HKUDS#4 proof: subprocess_runner injection seam asserted with shell-free list cmd + cwd pin + 300s timeout
- Pitfall HKUDS#6 proof: unregistered artifact_type surfaces 'Phase 3 plugin must register' message
- Pitfall HKUDS#10 proof: deploy_url_checker injection seam covers URLError offline-safety path
- RED: all tests fail with ModuleNotFoundError on clawteam.harness.evidence_gate

- 3 new modules (high_impact_gate, conflict, backfill) wired into /learn
- 16 new tests green (6 gate + 5 conflict + 5 backfill)
- 94/94 memory+learn+plugin tests green; plugin registers 13 skills
- MEM-06 + QUALITY-10 closed
- Phase 6 Success Criteria HKUDS#6 + HKUDS#8 realized

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

User request: "用tmux hook 把prompt注入进每一轮对话 包括身份和需要用的
skill 或者用tmux hook去做一些别的事情 总之tmux hook会是一个很好的技术".

Concept: agents drift over long sessions. CEO forgets to delegate (user
just hit this bug), engineer forgets which `clawteam` subcommands are
available, shipper forgets that `/ship` lives in the framework. A
periodic re-injection of "you are X; here's what you should use"
keeps them anchored.

Implementation:

1. `clawteam nudge <pane_id>` CLI command (new clawteam/nudge.py):
   - Looks up pane_id → session → team → agent via the persisted
     tmux_pane_map.json
   - Debounces (60s cooldown per pane) so repeated silence-hook firings
     don't spam the claude TUI
   - Reads the latest sprint state (current_phase + goal) for context
   - Builds a role-specific reminder text from _ROLE_NUDGES table
     (each role gets 2-3 lines: what you ARE + what commands you CAN
     run now). Appends "Goal: X · Phase: Y" when available.
   - Injects via `tmux send-keys -l <text>` + `send-keys Enter` so
     claude processes it as a user turn.
   - Silent-fail on any error (a noisy hook drowns the user's tmux).

2. tmux hook wiring (tmux_backend.py::_configure_nudge_hook):
   - `monitor-silence = 45s` on each window
   - `alert-silence` hook fires `clawteam nudge #{pane_id}` in the
     background (`run-shell -b`)
   - Opt-out via `CLAWTEAM_NUDGE_SECONDS=0` env var
   - Called from both `TmuxBackend.tile_panes()` (happy path) and
     `TmuxBackend.enable_mouse()` (the --windows non-tiled path) so
     every team gets nudge wiring regardless of layout choice.

3. pane_map normalization (tmux_backend.py::tile_panes):
   - Post-layout readback of pane_title returned claude's dynamic
     status strings like "⠐ CEO" / "✢ ENG-MGR" (spinner glyphs prefix).
   - Normalize by matching uppercase title content against the
     pre-merge window roster (roster_upper dict). Any title containing
     a known role (e.g. "⠐ CEO" contains "CEO") maps to the canonical
     lowercase role name.
   - pane_map.json now reliably stores {"0": "ceo", "1": "pm", ...}
     regardless of what claude's currently showing.

Real-env verified:
  clawteam go --no-attach --no-window "test"
  cat ~/.clawteam/teams/<team>/tmux_pane_map.json  → clean lowercase roster
  clawteam status                                    → Panes: 0=ceo 1=pm ...
  # wait for claude to boot
  clawteam nudge --force <pane_id>                   → injects role reminder
  tmux capture-pane                                  → sees "[nudge] Reminder:
                                                      you DELEGATE, never
                                                      implement. Decompose..."

Future uses of this hook infrastructure (brainstormed but not impl):
  - HKUDS#3 pane-died → log death reason, optional respawn
  - HKUDS#5 alert-activity → append pane output to activity.jsonl timeline
  - HKUDS#6 session-created → broadcast phase+goal to all panes
  - HKUDS#7 send-keys filter → intercept dangerous commands (rm -rf, force push)
  - HKUDS#10 pipe-pane → validate agent envelope JSON against schema

Tests: 30 solo/spawn_cli still green (nudge unit tests not added yet
— next commit candidate).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>