context

Context Management

Context is the agent's only perception interface. Everything the model knows about the current task comes from the context window.

Five Partitions

C = C_system ⊕ C_memory ⊕ C_skill ⊕ C_history ⊕ C_working

Partition	Purpose	Compressed?
C_system	Harness-injected instructions	Never
C_working	Dashboard, plan, events, scratchpad	Never
C_memory	AGENTS.md, persistent notes	Rarely
C_skill	Active skill/tool declarations	Swapped on demand
C_history	Execution history	Yes — first to go

Protection priority: C_system > C_working > C_memory > C_skill > C_history

C_working — The Decision Interface

C_working is not a log. It's the model's working memory:

C_working = {
    "dashboard": {
        "rho":                  # context pressure (0.0–1.0)
        "token_budget":         # remaining budget
        "goal_progress":        # model's self-assessment
        "error_count":          # recent errors
        "depth":                # sub-agent recursion depth
        "last_hb_ts":           # last heartbeat timestamp
        "interrupt_requested":  # H_b interrupt flag
    },
    "plan":            [...],   # current task plan
    "event_surface":   {...},   # pending events, active risks
    "knowledge_surface": {...}, # active questions, evidence packs
    "scratchpad":      "...",   # short-term reasoning
}

Five Compression Levels

Triggered by context pressure ρ, one level per turn:

Level	Trigger	Strategy
Snip	ρ > 0.7	Truncate messages over 2000 chars
Micro	ρ > 0.8	Deduplicate tool results by call ID
Collapse	ρ > 0.9	Extractive summary of middle messages
Auto	ρ > 0.95	Score-based keep: K·rel·e^(−λ·age)
Reactive	API 413	Emergency: auto + snip at 500 chars

Context Renewal (Disk Paging)

When ρ ≥ 1.0, Loom snapshots C_working to M_f (filesystem), compresses history, and rebuilds a fresh context window. The agent continues without losing its working state.

Code: loom/context/