Skip to content

Unattended Claude Code Harness

Let the Claude Code CLI itself run unattended — driving the graph-os Loop engine while you sleep — behind a governance-derived permission fence. Pillar 5 (Agent OS) · Concepts OS-5.40, OS-5.41, ECO-4.47, SAFE-1.8.

The common "run Claude overnight" recipe is a hand-written settings.json permission fence. We derive the fence from the same ActionPolicy (orchestration/action_policy, OS-5.24) that gates fleet mutations, add a dynamic PreToolUse gate that consults governance at decision time, and point the unattended session at the existing LoopController (KG-2.78) — the feature-extraction + innovation-distillation cycle — so the overnight worker is genuinely productive, not just a faster autocomplete.

Three layers

Layer Concept Module Role
Static fence OS-5.40 claude_harness/claude_fence.py Generate settings.json (allow/ask/deny + defaultMode=acceptEdits) + .claudeignore from ActionPolicy + secret patterns. Self-updating: a new forbidden rule propagates on the next run.
Dynamic gate OS-5.41 claude_harness/pretooluse_gate.py The PreToolUse hook body. Static secret/irreversible deny first (daemon-independent), then ActionPolicy.classify() for the governed verdict. Fail-closed.
Loop driver ECO-4.47 / SAFE-1.8 claude_harness/overnight_runner.py Drive LoopController cycles, commit per productive cycle, halt on ask, write the MEMORY.md morning summary.

deny > allow > ask (Claude's own precedence); we additionally de-dup so allow/ask never shadow a deny. defaultMode is hard-pinned to acceptEdits — bypassPermissions is never emitted.

Surfaces

  • CLI — setup-config harness-fence, agent-utilities harness-gate (the hook command), agent-utilities sleep-run.
  • MCP — graph_configure(action="harness_fence").
  • Skill — unattended-loop-driver (universal-skills).
  • Install — scripts/install.sh step 4.5 + genesis.yaml ide_targets.harness_fence.

Flow

flowchart TD
    subgraph Setup["1 · Draw the fence (OS-5.40)"]
        AP["ActionPolicy<br/>forbidden / approval / auto"]
        SEC["Secret patterns<br/>_SECRET_SUFFIXES + globs"]
        AP --> FENCE["claude_fence.write_fence"]
        SEC --> FENCE
        FENCE --> SETTINGS["~/.claude/settings.json<br/>deny gt allow gt ask<br/>defaultMode=acceptEdits"]
        FENCE --> IGNORE[".claudeignore"]
    end

    subgraph Run["2 · Unattended session"]
        CC["Claude Code CLI"] -->|each tool call| GATE
        GATE{"PreToolUse gate<br/>OS-5.41"}
        GATE -->|secret / irreversible| DENY["deny"]
        GATE -->|ActionPolicy.classify| TIER
        TIER -->|forbidden| DENY
        TIER -->|approval| ASK["ask: halt + queue<br/>SAFE-1.8"]
        TIER -->|auto| ALLOW["allow"]
        CC -->|graph_loops action=run| LOOP["LoopController.run_one_cycle<br/>feature extraction + distillation"]
        LOOP --> COMMIT["commit per productive cycle"]
    end

    subgraph Review["3 · Morning"]
        COMMIT --> SUMMARY["overnight_runner.write_morning_summary"]
        SUMMARY --> MEM["MEMORY.md"]
        MEM -->|inject_project_context KG-2.1| NEXT["next SessionStart"]
    end

    SETTINGS -.governs.-> GATE
    IGNORE -.hides secrets.-> CC

Guarantees

  • Fail-closed — any error, unparseable hook input, or engine-down denies; the static secret/irreversible floor works even with the graph-os daemon down.
  • ask = halt, never auto-approve — an unattended session has no human, so an ask verdict stops the action and queues it for review (SAFE-1.8).
  • Propose-only loop — LoopController writes proposals (specs, skill candidates, teams); it never auto-merges or executes high-stakes actions.
  • Idempotent — write_fence deep-merges and is a no-op on an unchanged policy, preserving operator hand-edits in allow/ask.