context-budget-guard

name	context-budget-guard
description	Heuristic context budget checks using file-size and count proxies — integrates with context-cache-discipline; no tokenizer introspection claims.
version	1

Context budget guard

Use this skill before major reads, large merges, or multi-tool bursts when you need a cheap, model-agnostic signal that the session is approaching an unsafe context load.

Non-claims (read first)

• Agents do not have reliable access to true remaining context window or exact tokenizer token counts from the runtime.
• This skill uses proxies only (bytes, line counts, call counts). Treat outputs as rough estimates with up to ~2× error acceptable; thresholds include margin by design.

Integration

• Prefix stability, lean rules, Task payload shape: load and follow context-cache-discipline first. This skill does not duplicate that material.
• Delegation policy: when proxies say you are hot, mandatory-delegation (mandatory-delegation.md) still governs who may do work inline vs Task.

Proxy “units” (cumulative heuristic)

Maintain a running estimated_units for the current turn (reset per user turn unless your orchestrator policy says otherwise). Units are not tokens — they are comparable weights for relative overload only.

Signal	Add to estimated_units
File read (full file in context)	file_size_bytes / 4 (floor 1 per file)
Tool call with substantial returned body	500 × number of such calls (cap each call’s contribution at 8000 unless you have byte counts)
Sub-agent YAML envelope (in parent context)	200 per envelope merged
Plan or long rule file read	line_count * 8 (floor 80 for any file >10 lines)

Large-file shortcut: if a file is >500 lines OR >40 KiB, treat as high-risk for coordinator inline work — prefer Task to a scoped sub-agent even before summing units.

Threshold bands

Let U = estimated_units. Let W = a notional window proxy chosen conservatively:

• Default W = 1_000_000 units (~order-of-magnitude stand-in for “large context”; tune per platform using observability in later phases).
• 40% band: if U > 0.40 * W before the next large action → prefer delegation (Task a sub-agent with a tight brief) rather than pulling more bulk into the coordinator.
• 80% band: if U > 0.80 * W after an action → force-delegate the remainder: coordinator stops bulk work; Task fresh sub-agent(s) with shard ids + disjoint touches[].

If you cannot estimate W, use the large-file shortcut + 40%/80% as relative triggers: crossing 40% of a session-local baseline you set at turn start (e.g. sum of first three tool returns) means slow down and delegate; 2× that baseline ≈ 80% emergency path.

Coordinator checklist

1. After read-only discovery, estimate U for planned reads; if a single read would jump U over 40% of W (or 2× baseline), Task instead.
2. Before merging many sub-agent envelopes, ensure parent is not also holding full duplicates of child artifacts — cite paths, merge verbatim fields only (see subagent-response-protocol).
3. On 80% crossing, do not “finish inline” — split remaining work across new Tasks with smaller payloads.

Context demotion bands (policy-linked)

When estimated_units crosses bands defined in memory-demotion.yml → context_demotion (40% / 80% proxies aligned with this skill):

• >40% of notional W: reduce L2/L3 KB reads; prefer L0/L1 + delegation.
• >80%: L0/L1 only unless explicit escalation or fresh_eyes session flag.

Load ladder details from brain-memory-kb; do not invent parallel retention integers in this skill.

Calibration

Thresholds are initial. Prefer agent-observability + org observability phases to compare proxy U vs observed session behavior and tune W / multipliers — no false precision.