Skip to content

title: page: "Chad Autonomy Loops" nav: "Chad Autonomy" description: main: "The six self-driving loops Chad runs from cron, the wrappers that close them, the maintenance crons behind them, and the sub-agent orchestration contract." agent: "Reference for the autonomy surface area: which loops exist, where their wrappers live, what they read and write, how sub-agents are spawned/reconciled, and the safety boundaries." keywords: ["chad autonomy", "self-improve", "feedback proposals", "cron telemetry", "gbrain dream", "skill discovery"] topics: ["operations", "autonomy"] tags: ["openclaw", "openshell", "nemoclaw", "chad", "operations"] content: type: reference difficulty: technical_intermediate audience: ["developer", "engineer", "operator"] status: published

Chad Autonomy Loops

Chad's "always-on" claim cashes out as six distinct cron-driven loops that read state, propose action, and apply or surface that proposal. This page enumerates them and the wrappers that close each one.

The six loops

# Loop Wrapper(s) Schedule Reads Writes
1 Self-improvement chad-self-improve v2 (detached signal collection + ONE single-turn LLM call → machine-readable proposals) → chad-proposal-apply (applies safe-list) Sun 03:00 UTC + daily 04:30 UTC last 7 days of spawn-ledger failures, sub-agent results, cron run logs (cron/runs/*.jsonl, read directly — job UUIDs mapped to names), auto-action-log.jsonl, feedback memory titles, outstanding proposals prepends ## Self-improvement run + ### Proposals (machine-readable) block to feedback-proposals.md; safe-list applies via openclaw cron edit; ## Applied block back into proposals
2 Bug → fix → ship chad-issue-triage-cron → chad-drafter → optional chad-spawn --kind researcher/coder daily 10:00 UTC top-N open issues in ${CHAD_BUG_REPO} drafts under ### Issue triage drafts in today's memory; never auto-comments
3 Cron telemetry / budget audit chad-budget-audit (emits prose + structured JSON) → chad-proposal-apply (consumes JSON) Mon 04:00 UTC + daily 04:30 UTC last-50 openclaw cron runs per task, /tmp/chad-premium.jsonl, task-profiles.json recommendation table + ### Proposals (machine-readable) JSON block to feedback-proposals.md; bounded openclaw cron edit calls; ## Applied block
4 Dreaming (gbrain consolidation) chad-gbrain-dream + dream-digest step nightly 03:30 UTC today's memory/<date>.md, events-<date>.jsonl, workspace doc set, gbrain stale chunks, gbrain doctor gbrain pages upserted; memory/dream-digest-<date>.md (24h delta + doctor); memory/feedback_brain_health_<date>.md if doctor reports anomalies
5 Skill discovery chad-skill-watch (daily diff) + chad-setup.sh skill-sync (host-driven) daily 09:00 UTC + on-demand from host openclaw skills list --json vs snapshot at /sandbox/.openclaw-data/state/skills-snapshot.json; ~/.claude/skills/gstack/.openclaw/skills/ snapshot updated; ## Skill catalog diff block in today's memory; signal-detector skill picks it up next reasoning cycle
6 Autonomous experiment lifecycle chad-experiment-cron (deterministic observe → evaluate → design driver) + chad-experiment CLI nightly 02:00 UTC active experiment records, recent memory digest, archived hypotheses heartbeat observations + LLM evaluation verdicts (promote/retire/extend) in state/experiments/ledger.jsonl; new whitelisted OpenWebUI artifacts (notes / automations / memories) via chad-webui; one-line summary in today's memory. Detailed in chad-experiments.md.

Every wrapper stays inside the wrapper-only invariant documented in chad-readme.md §7.1: cron messages are one-line invocations, slow work is nohup'd inside the wrapper. Inputs/outputs are markdown + JSONL files under /sandbox/.openclaw/workspace/memory/ and /sandbox/.openclaw-data/, never live agent state.

Closing the loops

Both write-only gaps are now closed by three new wrappers + one extension. The autonomy loops are no longer one-way:

Loop Closing wrapper Trigger What it does
Loop 1 / 3 chad-proposal-apply daily 04:30 UTC cron Reads the latest ### Proposals (machine-readable) JSON block from feedback-proposals.md. Per-kind handlers: cron_edit (default) applies via openclaw cron edit if (field ∈ {timeoutSeconds, maxOutputTokens}, value ≤ absolute cap 7200/32768, change ≥ 1.1×, last run ok, chad-action-gate chad_self_modify_cron returns auto); memory appends content via chad_self_modify_memory (auto by default); automation/note/calendar_event are recognised but parked for v3. Entries failing the gate (decision=draft) or above the cap route to a ## Pending operator review table so they don't disappear. ## Applied + ## Pending blocks older than 30 days are pruned.
Loop 1 (signals) chad-self-improve v2 weekly Sun 03:00 UTC cron Full rewrite (2026-06-10): the cron agent only runs chad-self-improve --detach and acks the detach line. The wrapper collects signal deterministically (failed spawns, sub-agent results, cron failures read straight from cron/runs/*.jsonl, auto-action errors, feedback memory, outstanding proposals — capped at 7000 prompt chars), makes ONE single-turn no-tools LLM call, validates the response, and prepends prose + machine-readable proposals to feedback-proposals.md. Replaces the v1 multi-turn researcher spawn that hit LLM idle timeouts and duplicate-spawned 4× in 5 minutes. flock + per-day dedupe make reruns safe.
Loop 4 chad-dream-digest (appended to chad-gbrain-dream) nightly 03:30 UTC After the dream cycle, writes memory/dream-digest-<date>.md with new pages since yesterday + gbrain doctor tail + stats. If doctor reports [WARN]/[FAIL]/[ERROR], mirrors them into memory/feedback_brain_health_<date>.md so the next self-improve sees them as a recurring rule.
Loop 5 chad-skill-watch daily 09:00 UTC cron Diffs openclaw skills list --json against /sandbox/.openclaw-data/state/skills-snapshot.json and surfaces added / removed / description-drifted skills under ## Skill catalog diff in today's memory. The signal-detector skill picks them up on chad's next reasoning cycle.

Safety properties

  • No code or policy modification. Every wrapper stays inside the draft-only boundary documented in project_chad_autonomy_roadmap. Code changes still round-trip through bug-intake → researcher → coder → human PR review.
  • Bounded knob tweaks. chad-proposal-apply only ever moves timeoutSeconds/maxOutputTokens for an existing cron, up to absolute caps (7200 / 32768) — values above the cap route to operator review instead of being silently rejected. Replaces the earlier ±2× bound that blocked evidence-based proposals from chad-budget-audit.
  • Per-action gating. Every applied tweak passes chad-action-gate check chad_self_modify_cron <name> first; the global kill-switch file (/sandbox/.openclaw/workspace/.auto-disabled) shorts the whole loop on demand.
  • Idempotency. Re-runs against the same proposals block are no-ops — entries are matched against the audit log (/sandbox/.openclaw-data/state/auto-action-log.jsonl) by (target, field, new) within 24h.
  • Fully audited. Every apply emits one line to the audit log and one row to the ## Applied table in feedback-proposals.md.
  • No daily noise. Pending entries are deduped against blocks already in the file (same target/field/old/new), and a fully no-op applier run appends nothing — the daily cadence against weekly proposals doesn't grow the file.

Sub-agent orchestration contract

Chad routes big tasks to sub-agents rather than absorbing them into the live session. The contract lives in the pod's AGENTS.md (backed up via chad-workspace-backup) and the chad-orchestrator skill:

  • Detect: multi-step research, content drafting from sources, code or repo analysis, anything over ~2 minutes of focused tool work, or an explicit "in the background" → spawn; small lookups stay inline.
  • Route: check queue/tasks.jsonl for a running duplicate first, then chad-intake --from chat --task-file <file> (auto-routes the kind) or chad-spawn --kind <kind> --task-file <file> directly.
  • Report — always: every spawn (or refusal) is reported to the user: what was kicked off (kind + task_id), where results land (subagents/<task_id>/ + today's memory), when to expect them, or why the task was not spawned. Big tasks are never silently absorbed.
  • Reconcile: chad-spawn-poll (every 5 min, host launchd) transitions async gha entries; the weekly spawn-gc cron terminal-izes entries stuck queued/running >24h (orphaned by a killed parent) and prunes spawn branches.

Maintenance crons behind the loops

Cron Schedule What it protects
spawn-gc Mon 02:30 UTC Ledger hygiene (orphan reconcile) + chad-spawn/* branch retention (done=7d, failed=30d)
gbrain-prune Sun 02:00 UTC gbrain page retention (memory/events >365d, chat >180d) + stale workspace digests (>30d)
memory-curator Sat 04:00 UTC Memory consolidation proposals (never auto-applied)
workspace-backup every 6h Pod state survives sandbox resets

Always finish with text after tool calls

Chad's runtime expectation, enforced via AGENTS.md on the chad pod (at /sandbox/.openclaw/workspace/AGENTS.md): every agent turn must end with a text response, even when the substantive work was done via tool calls. The pattern breaks when a turn ends with a read / exec / gbrain_search / webui__* tool call but no closing text — the morning-summary automation hit this on 2026-05-17, producing empty replies in OpenWebUI chats despite the underlying tools succeeding.

Source-of-truth note: AGENTS.md is not git-tracked in NemoClaw because the canonical copy lives on the (ephemeral) chad pod and gets backed up to CHAD_STATE_REPO (tantodefi/chad-state by default) via the chad-workspace-backup cron. The "always finish with text" rule was added to the pod copy on 2026-05-17. When the next major refactor pulls workspace identity files into source (planned), this rule moves into the source template.

Cross-reference