Tune Auto Orchestration

How to tune auto-orchestration

Note

Goal: Adjust the auto-orchestration toggles, thresholds, and cooldowns so the SessionStart briefing and idle-time memory chain fire on a cadence that fits how you work. Prereqs: MEMORY_VAULT_PATH set; harness with the auto-orchestration push-surface installed (V4 #23). See Auto-orchestration config for every key this page edits.

The tunables live in <vault>/personal-private/auto-orchestration-config.md, auto-seeded with sensible defaults the first time the push-surface runs. The file is yours to edit — a re-seed never clobbers your edits. Every key sits in one settings fence; you change a value, save, and the next run picks it up. This page is the recipe for editing a threshold, a cooldown, or a toggle, then verifying the change with --dry-run.

Steps

1. Locate (or seed) the config. Open <vault>/personal-private/auto-orchestration-config.md. If it doesn't exist yet, seed it with the defaults:

   python3 harness/skills/memory/scripts/auto_orchestration.py --vault-path "$MEMORY_VAULT_PATH" seed-config

   It prints seeded (wrote the file) or kept (one already existed — your edits are safe). The values to edit live in the fenced block:

   ```settings
   enable_briefing = true
   enable_idle_chain = true
   inbox_threshold = 10
   briefing_cooldown_hours = 8
   idle_chain_cooldown_hours = 24
   ...
   ```

2. Raise or lower a briefing threshold. In the settings fence, change the threshold for a signal the briefing reports too often (or not often enough) — for example, only flag the inbox once it's larger:

   - inbox_threshold = 10
   + inbox_threshold = 25

   The other thresholds are watchlist_high_threshold, incubator_pending_threshold, idea_ledger_stale_months, plus the two nudge thresholds promote_mention_threshold and stale_promotion_days.

3. Change a cooldown so a chain fires more or less often. Cooldowns are in hours; a non-positive value means "always eligible". To let the idle chain run twice a day instead of once:

   - idle_chain_cooldown_hours = 24
   + idle_chain_cooldown_hours = 12

   The others are briefing_cooldown_hours and phase_reflect_cooldown_hours.

4. Toggle an emission on or off. Each enable_* key turns one emission off entirely. To silence the SessionStart briefing while keeping the idle chain:

   - enable_briefing = true
   + enable_briefing = false

   The toggles are enable_briefing, enable_idle_chain, enable_phase_integration, enable_promote_suggest, enable_stale_promotion_nudge. (There is no per-step toggle inside the idle chain — enable_idle_chain gates the whole chain.)

5. Verify the change took effect. Confirm the config parses to the values you expect:

   python3 harness/skills/memory/scripts/auto_orchestration.py --vault-path "$MEMORY_VAULT_PATH" show-config

   Then dry-run the affected chain without touching state — the idle chain prints its resolved step plan and cooldown_ok:

   python3 harness/skills/memory/scripts/orchestration_idle.py --vault-path "$MEMORY_VAULT_PATH" --dry-run

   For the briefing, run the generator directly — it prints the block only when something is over threshold and the cooldown allows:

   python3 harness/skills/memory/scripts/orchestration_briefing.py --vault-path "$MEMORY_VAULT_PATH"

   To see when each chain last fired, inspect the last_fire timestamps in <vault>/_meta/auto-orchestration-state.json.

See also

• Auto-orchestration config — every config key and the state-file shape.
• Auto-orchestration — the memory push-surface — what the briefing and idle chain do, and why they never nag.
• Use auto-context in harness phases — tuning the complementary phase-boundary pull surface (recall budgets, save mode).