context-loading-protocol

Context Loading Protocol

Token-budget reference (CLAUDE.md baseline, full-load ceiling, per-agent and per-skill costs) is the Baseline Budget section of CLAUDE.md. This skill is the runtime procedure; don't duplicate the table here — it goes stale.

Constraints

• Never load all agents upfront; load only the primary agent for each phase.
• Keep total context below 40% of the model's window at all times.
• Load agents on demand when their phase begins, not speculatively.
• Use tool-based file reads (Read); do not paste file contents into the prompt.

Enforcement

This protocol is backed by a PreToolUse hook — hooks/context_ceiling_guard.py (registered on Agent and Skill). Before a capability-loading call it measures utilization = (input + cache_read + cache_creation) / model_context_window from the transcript's latest assistant-message usage against the model's context window, which the hook auto-detects from the session's most recent message.model by family/version substring: Haiku family -> 200K; current 1M-window models -> 1M (Fable, Mythos, Opus 4.6/4.7/4.8, Sonnet 5, Sonnet 4.6); unrecognized model, or a same-family model outside those pinned versions -> 200K conservative fallback (window is a fixed per-model property, so an unrecognized model is never assumed large — over-nudging is a minor false alarm, under-nudging risks running well past the real ceiling). Set DEV_TEAM_CONTEXT_WINDOW to override detection explicitly.

The effective ceiling is min(ceiling_pct% of window, 150K tokens) — an absolute-token cap (DEV_TEAM_CONTEXT_ABS_CEILING, default 150000, matching Anthropic's server-side compaction default) that keeps large windows from pushing the trigger point far past where compaction already kicks in; it's a no-op on the 200K base window (40% = 80K, already under the cap). The warning names which bound is binding — percentage or absolute, never both — and the window's provenance (override, detected, or default).

As occupancy climbs past the ceiling, the hook escalates through three Handoff action bands keyed to multiples of the effective ceiling — 1x nudge, 1.25x run /handoff now, 1.5x full summary + fresh conversation (see Handoff → When to Summarize) — before nudging (warn, default) or, at/above the ceiling under DEV_TEAM_CONTEXT_STRICT=on, blocking the load. Recovery skills (/handoff, /context-loading-protocol, /continue, /review-summary, /session-review) are never gated — blocking the path back under budget would deadlock the session.

Knobs: DEV_TEAM_CONTEXT_CEILING_PCT (default 40), DEV_TEAM_CONTEXT_ABS_CEILING (default 150000), DEV_TEAM_CONTEXT_WINDOW (overrides auto-detection), DEV_TEAM_CONTEXT_CEILING=off (disables entirely). The hook is a backstop measured from real usage; the budget estimate below is still the planning tool you apply before loading.

Why 40%

The 40% ceiling is a conservative planning target, not a claimed accuracy cliff. Chroma's Context Rot study found degradation across 18 models (including Claude 4) is gradual, not a sharp drop at any single percentage. Needle-in-a-haystack benchmarks like RULER and NoLiMa show a model's effective context is often only about half its advertised window, with sharp accuracy drops on non-lexical retrieval well before the window limit. Anthropic's effective context engineering guidance recommends proactive compaction well ahead of the limit — the Claude API's own compaction default is 150K absolute tokens even on 1M-window models. Given that evidence, budgeting to 40% of the window (capped at 150K absolute) leaves headroom before quality degrades, rather than chasing a precise threshold that doesn't exist.

Full guide — warning-line field reference, concrete band fire-points per window size, knob table, troubleshooting: Context Management.

Loading Decision Procedure

Step 0: Confirm there is a task

Before loading anything or reading files, confirm an actionable instruction exists. If the user has not yet said what they want, wait — do not speculatively read files, verify code, or load agents. Premature investigation before a task is given wastes context and is a common interrupt trigger. Once a task exists, proceed to Step 1.

Step 1: Classify the task

Step 2: Select agents

Load the minimum set:

1. Identify the primary agent (owns the deliverable).
2. Identify supporting agents (input or review).
3. Do NOT load agents for downstream validation yet — load them when their phase begins.

Order: primary first, then supporting agents one at a time as their phase begins.

Step 3: Select skills

For each loaded agent, check its ## Skills section:

• Only load skills relevant to the current task — not all skills the agent references.
• Skills shared by multiple loaded agents only need to be loaded once.

Step 4: Calculate token budget

Total = CLAUDE.md baseline
      + conversation history (estimate)
      + agent files (sum selected)
      + skill files (sum selected)
      + expected output (estimate)

Target: total < 40% of the model's context window, capped at 150K absolute tokens. For Claude with a 200K window, that's < 80K tokens; on a 1M-window model the cap (150K) binds before the percentage would. See Why 40% for the rationale. The config files are a small fraction; the real budget concern is conversation history + output accumulation over multi-turn tasks.

Step 5: Load via tool-based file reads

Read agents/software-engineer.md
Read skills/hexagonal-architecture/SKILL.md

Do NOT copy file contents into the system prompt or conversation.

Loading Profiles

Pre-computed loading sets for common task types.

Code Implementation

• Load: Software Engineer + relevant skill(s)
• Defer: QA (load after implementation), Architect (load only if design questions arise)

Architecture Design

• Load: Architect + relevant architecture skill(s)
• Defer: Software Engineer (load at implementation), QA (load at validation)

Bug Fix

• Load: Software Engineer only
• Defer: QA (load if regression test needed)

New Feature (full lifecycle)

Three phases, each in a fresh context window with a human review gate between. Each phase's output is a structured progress file in memory/ that onboards the next phase.

Key rules:

• Each phase starts with a fresh context window, loading only the previous phase's progress file.
• Human reviews and approves the progress file before the next phase begins.
• Sub-agents primarily provide context isolation — they search, read, and return concise findings.
• If implementation is large, compact mid-phase: update the plan progress file with completed steps and continue in a fresh context.

Unloading

Since tokens can't be literally removed from context:

1. Phase transitions — summarize completed phase output into memory/ and start a new conversation for the next phase.
2. Within a conversation — stop referencing the agent/skill; the orchestrator mentally notes it's no longer active. Use the Handoff skill (continue mode) to compress stale content.
3. Multi-turn accumulation — when conversation history crosses 30% utilization, trigger summarization before loading additional agents.

Anti-patterns

• Loading all agents upfront — wastes tokens before any work begins. Load only the primary agent.
• Loading all of an agent's skills — most are irrelevant to the specific request.
• Never unloading — context grows monotonically until hallucination risk. Summarize and phase-transition.
• Loading agents "just in case" — adds cost without value. Load on demand when the phase begins.

Output

Loading plan as one table: selected agents + skills, token costs, estimated total, and utilization percentage against the 40% ceiling. No narration.