Two files drive the memory push-surface. You tune the config at <vault>/projects/agentm/desk/auto-orchestration-config.md, where the AgentKV layout move (task 176 step 3, 2026-09-25) put it — the project root is never read for it (task 176 step 4 locked the root to five files, and the config lists aren't among them), and neither is the retired <vault>/memory/ spelling the memory-root trims (agentm-vault plan 05) moved it out of before that. It holds toggles, thresholds, and cooldowns. It auto-seeds with defaults. Re-seeding never clobbers your edits. The system stores runtime state at ~/.local/state/agentm/auto-orchestration-state.json — the engine state directory ($AGENTM_STATE_DIR overrides it; machine state left the vault in filing-v2 part 2a). This state holds last-fire-per-chain timestamps. It also holds the last-shown snapshot for the shifted-since-last-shown check. All keys live in DEFAULT_CONFIG in auto_orchestration.py. See Auto-orchestration for the why. See Tune auto-orchestration to edit the config.

Question	Answer
Where is the operator config?	<vault>/projects/agentm/desk/auto-orchestration-config.md — auto-seeded, operator-editable. The project root and the retired <vault>/memory/ spelling are never read (task 176 locked the root to five files).
Where is the runtime state?	~/.local/state/agentm/auto-orchestration-state.json (engine state directory; $AGENTM_STATE_DIR overrides) — last_fire timestamps + last_shown snapshot.
Does a re-seed overwrite my edits?	No. seed_config is idempotent and never clobbers an existing file.
What are the four config groups?	Emission toggles · briefing thresholds · nudge thresholds · chain cooldowns.
What does the state file gate?	Cooldown windows (should_fire(chain, now)) and the shifted-since-last-shown briefing check.
Who writes the state file?	auto_orchestration.py alone — via save_state(). No other script in orchestration_*.py may call save_state directly (V5-5 single-writer invariant, LC-2; enforced by verify-v4.sh segment G).
Related pages	Auto-orchestration, Tune auto-orchestration

Each enable_* toggle turns one emission off entirely. A false value short-circuits that chain before any work. You set these with true or false.

Key	Controls	Default
enable_briefing	the SessionStart pending-state briefing	true
enable_idle_chain	the idle-time orchestration chain	true
enable_phase_integration	post-/work reflect + post-/release skill refresh	true
enable_stale_promotion_nudge	the stale-promoted watchlist safety-rail nudge	true

These counts decide whether the SessionStart briefing reports a section. A signal surfaces only at or above its threshold.

Key	Controls	Default
inbox_threshold	The _inbox/*.md count that triggers a briefing flag. Filing v2's write path retired _inbox/ as a capture destination, so this still reads that directory and effectively sees nothing — it does not see the low-confidence filings (filing_confidence: low) that now live at their class directories instead	10
watchlist_high_threshold	_skill-watchlist/ HIGH + pending-review count to surface	1
incubator_pending_threshold	_idea-incubator/<slug>/ dirs pending research to surface	1

This threshold applies inside the one remaining nudge counter before it surfaces. The idea-ledger signal and the "you keep having this idea — promote it?" nudge (idea_ledger_stale_months, promote_mention_threshold, enable_promote_suggest) retired in agentm-vault part 13 with the hand-kept Ideas.md they read — Ideas.md is generated over personal/ideas/ now, and an idea card is never garbage-collected or flagged for recurring. A config file that still carries those three keys is read as before; nothing consults them any more.

Key	Controls	Default
stale_promotion_days	days a _skill-watchlist/ entry can sit status: promoted before the nudge fires	30

This is the minimum window in hours between fires of each chain. The should_fire(chain, now) check honors it. A non-positive value means "always eligible".

Key	Controls	Default
briefing_cooldown_hours	min gap between SessionStart briefing emissions (chain briefing)	8
idle_chain_cooldown_hours	min gap between idle-chain runs (chain idle_chain)	24
phase_reflect_cooldown_hours	min gap for the phase-integration dispatches (chains phase_reflect + phase_release)	1

The idle chain runs these bounded steps in order. The enable_idle_chain toggle and the idle_chain cooldown gate the whole chain. You cannot toggle steps individually. Each underlying script self-no-ops when you give it empty input.

Step	What it runs	Bound
reflect-corpus	reflect.py corpus --execute — mine unseen session transcripts	--batch-size 5 --max-batches 1 (≤5 sessions/pass)
discover-skills	discover_skills.py --cadence-check — refresh discovery sources	cadence-check (self-throttled)
adapt-pass1	adapt_skills.py --limit 3 — stage Pass-1 candidate JSONs	--limit 3

auto_orchestration.py:save_state() writes <engine-state-dir>/auto-orchestration-state.json exclusively (default ~/.local/state/agentm/; $AGENTM_STATE_DIR overrides). This enforces the V5-5 LC-2 single-writer invariant. No other script writes this file directly. Sibling orchestration scripts call through ao.save_state() instead. The vault's git never carried it; since filing-v2 part 2a it lives in the engine state directory, its own git repo the runner commits on cadence. It holds JSON with two top-level objects.

Field	Type	Meaning
last_fire	object	{<chain>: <ISO-8601 timestamp>} of the most recent fire per chain; read by should_fire. Chains: briefing, idle_chain, phase_reflect, phase_release.
last_shown	object	{<signal>: <count>} snapshot of the over-threshold signals the briefing last displayed; state_shifted_since_last_shown compares against it to suppress re-firing until a count changes.

• Auto-orchestration explains what these knobs change. It shows why the surface never nags.
• Tune auto-orchestration provides the recipe for editing the config.
• AgentMemory context payload details the vault folder map. These files reference the _inbox/, _idea-incubator/, _meta/, and memory/ paths.