feat(workspace): add idle-loop reflection pattern (Hermes/Letta shape, opt-in, ~90 LOC)

[HongmingWang-Rabbit]

Why

Observed team utilization today: ~0.5% — agents idle 99.5% of the time. Only cron fires and CEO-typed A2A wake them. CEO's north-star is 24/7 iteration, current cadence falls short.

Research baseline

Survey of 8 agent frameworks done this session (Hermes, Letta, Trigger.dev, Inngest, AG2, Rivet, n8n, Composio, SWE-agent). Findings in the conversation record and docs/ecosystem-watch.md. Nobody runs while(true) per agent. The working patterns are:

• Event-driven + hibernation (Hermes, Letta, Trigger.dev, Inngest) — wake on external event, sleep otherwise, zero idle cost
• Cron/user-triggered ephemeral runs (AG2, Rivet, n8n, SWE-agent) — what we already do

The winning unlock for Molecule AI is the Hermes reflection-on-completion pattern combined with the Letta backlog-pull pattern, implemented at the workspace layer so it applies uniformly to all 12 roles without per-workspace YAML changes.

What this PR does

Adds a ~90 LOC workspace-template change:

config.py — two new fields on WorkspaceConfig:

• idle_prompt: str = "" — prompt to self-send when idle
• idle_interval_seconds: int = 600 — check cadence (default 10 min)
• Both parse from yaml inline or via idle_prompt_file (same pattern as initial_prompt)

main.py — new _run_idle_loop() asyncio task alongside the existing _send_initial_prompt:

1. Wait min(interval, 60) for server + initial prompt to settle
2. Loop:
   a. Sleep idle_interval_seconds
   b. Local check: if heartbeat.active_tasks > 0: continue (no LLM, no HTTP)
   c. Self-POST idle_prompt via existing /workspaces/{id}/a2a proxy
3. Cancel cleanly in the server's finally: block, matching initial_prompt_task cleanup

Gated on config.idle_prompt being non-empty. Default "" means existing workspaces upgrade to this binary as complete no-ops. Zero behavior change until someone explicitly opts in.

What this PR does NOT do

• No Go code — zero platform changes
• No new DB columns / tables / migrations
• No new API endpoints
• No org-template changes — org.yaml untouched
• No scheduler changes — idle loop is a workspace concern, not a scheduler concern
• No default enable — opt-in per workspace via idle_prompt: field

Cost model

• while(true) sleep-loop pattern: ~$93/day/org (12 agents × 12 thinks/hour × $0.027). Unshippable.
• Hermes reflection-on-completion: ~$0.45/day/org. Cost ∝ useful work.
• This PR at 10-min cadence per role: upper bound ~$5/day/role IF genuinely idle every check. Real-world much less — busy periods skip the LLM call entirely via the local active_tasks==0 gate.

Rollout plan (per research-report recommendation)

1. Land this PR — mechanism shipped, zero behavior change
2. Opt in Technical Researcher first (via a tiny follow-up PR or direct DB/template edit) — ~24h measurement window
3. Measure: count activity_logs delta vs baseline, count idle-loop fires vs skipped, confirm cost stays within model
4. Roll to all 12 via a template-level defaults: idle_prompt: | block if the measurement looks good
5. Tune per-role — leaders get shorter intervals (5 min), deep-work roles get longer (30 min), workers probably no idle loop (they stay reactive)

Test plan

• Python syntax check (ast.parse) on main.py + config.py
• CI will run unit tests on the new self-hosted runner
• Dogfood on Technical Researcher for 24h, measure, tune

Related

• chore(template): orchestrator/worker split — leaders pulse every 5min, workers stay reactive (supersedes #158) #159 orchestrator/worker split (complementary — leaders push work via orchestrator pulses, workers pull work via idle loop)
• docs/ecosystem-watch.md → ### Hermes Agent entry
• Today's research report (full framework survey)
• The queued Hermes multi-provider work benefits from this — a Technical Researcher with an idle_prompt pointing at "pull from research backlog" keeps researchers working on the Hermes Phase 0 investigation without needing cron fires