v4.13.0 — Auto-orchestration: the memory push-surface (V4 #23)

MINOR. The Agent M memory skills were a pull surface — you had to remember to run recall, reflect, discover-skills, adapt-skills, and the watchlist by hand, so pending work piled up unseen. This release turns them into a push surface: open a session and the system already tells you what needs attention; during idle time it runs the right memory chains itself; at phase boundaries it reflects and refreshes. It never blocks a session, never nags (cooldowns plus a "only when state shifted since you last saw it" guard), and never acts on its own — every adoption or write stays operator-gated. Hook/file-based and cross-host (DC-1), entirely agentm-native (DC-3: crickets carries zero AgentM crossover now). This is the last open V4 item — the foundation finish. Single-repo release. The default thresholds and cooldowns are a first guess; the real-use dogfood on the operator's own vault calibrates them.

Added

• SessionStart pending-state briefing (orchestration_briefing.py) — on session boot, a tight 1–3 line block surfaces what's piled up: _inbox/ over threshold · _skill-watchlist/ HIGH-pending count · _idea-incubator/ entries in research · GC-eligible idea-ledger items. Emits ONLY when something shifted since it was last shown AND the cooldown allows; appended to the memory-recall-session-start hook after the always-load recall; non-blocking (any error → empty briefing, never wedges boot).
• Idle-time orchestration chain (orchestration_idle.py) — the memory-reflect-idle hook fires a cooldown-gated chain during idle time: reflect-corpus (≤5 unseen sessions, --max-batches 1) → discover-skills (cadence-checked) → adapt-skills Pass-1 (stages ≤3 candidates, --limit 3). Run detached so it never blocks or gets killed at the hook's 30s SessionStart timeout; results surface the next session via the briefing. Pass-2 (adapt-evaluator) stays operator-gated — a hook can't dispatch a sub-agent.
• Phase-integration auto-dispatch (orchestration_phase.py, via a new harness_memory.py phase-dispatch subcommand) — after /work commits a task, reflect the just-finished session (dedup-guarded against the memory-reflect-stop hook via the .reflected session marker, so a transcript is never reflected twice; works cross-host including Antigravity, which has no Stop hook); after /release, refresh the skill surfaces (index-skills + discover-skills, cadence-checked). Wired into the /work and /release phase specs (§9b), config-gated and non-blocking — extends, doesn't duplicate, the V4 #8 phase context dispatcher.
• Two SessionStart nudges, riding the same briefing block (one consolidated notice): promote-suggest — an idea surfaced ≥3× in the Ideas ledger → "consider /memory promote"; stale-promotion safety-rail — _skill-watchlist/ entries marked promoted for >30 days without action → "author the skill or dismiss."
• Operator-tunable config + runtime state (auto_orchestration.py, stdlib-only) — <vault>/personal-private/auto-orchestration-config.md (thresholds · cooldowns · per-chain toggles in a ```settings fence; auto-seeded once, a re-seed never clobbers operator edits) and <vault>/_meta/auto-orchestration-state.json (per-chain last-fire cooldowns + the last-shown snapshot that drives the anti-fatigue guard).

Changed

• adapt-evaluator sub-agent moved crickets → agentm (harness/agents/), completing the memory-surface consolidation begun in V4 #36 — the whole push-surface is now agentm-native. The paired crickets removal + stale-catalog cleanup already landed on crickets main (ships with crickets' next release).
• memory-reflect-stop hook gained a dedup skip-guard (.sh + .ps1): if a phase-dispatch already reflected the session (the .reflected marker is present), Stop skips — the two cooperate so a session is reflected exactly once.

Internal

• New scripts: auto_orchestration.py, orchestration_briefing.py, orchestration_idle.py, orchestration_phase.py + the harness_memory.py phase-dispatch bridge; adapt_skills.py gained --limit N (bounds per-pass GitHub-enrichment cost). 84 tests across 4 new test files (auto-orchestration 25 · briefing 31 · idle 12 · phase 16); full suite 549 green, 4-OS. Every code task passed an adversarial review that caught a real defect — a UnicodeDecodeError-escapes-catch crashing the hook on non-UTF-8 state/config, a clear-then-refill anti-fatigue suppression hiding live pending work, a wrong-session-under-concurrency reflect that burned the shared cooldown — each fixed and regression-tested.

Deferred

• Pass-2 auto-dispatch from the idle chain — the adapt-evaluator sub-agent stays operator-gated (hooks can't dispatch sub-agents); the idle chain stages candidates + surfaces the count, and the evaluate hand-off rides the phase-dispatch / nudge where dispatch is operator-gated.
• The Anthropic Workflow / mid-conversation-system-message hybrid (DC-5) — a post-V4 research follow-up: analyze whether hooks-for-triggers + Workflow-for-in-session-fan-out optimizes the orchestration.

Cross-references

• ROADMAP-V4 item #23 — auto-orchestration (the last open V4 item; foundation finish).
• agentm v4.12.0 — the immediately prior release (cross-surface vault access).