Context Budget Perspective (CTX)

Measures the context budget a deployed agent pays every session before real work starts. Assigns a score from 0.0 to 1.0 across two dimensions.

Scope

This perspective scans this repository's own agent-profile authoring layout — agents/profiles/<id>-workspace/. It does not parse or score any composed/deployed profile format: Claude Code and Cursor compose Markdown, Codex composes a per-agent .toml file. Measuring the pre-composition Markdown source files is a deliberate scope boundary, not a gap — a project without this repo's own agents/profiles/ layout has no ctx signal to measure and scores a neutral 1.0 rather than being penalized.

Dimensions

File size — How many lines does an agent's composed profile cost every session?

• Sums line counts across a workspace's fixed set of source files: AGENTS.md, IDENTITY.md, ROUTING.md, SOUL.md, TOOLS.md. MEMORY.md is excluded — memory is per-session state, not fixed startup cost.
• Score = max(0, min(1, 1 - (lines - 300) / 1500))
• Under 300 lines = 1.0 (clean); 1,800 lines = 0.0 (at limit)

Skill count — How many skills does an agent's manifest merge in?

• Counts entries in agent.manifest.json's skills array for that workspace.
• Score = max(0, min(1, 1 - (skillCount - 5) / 10))
• 5 or fewer merged skills = 1.0; 15+ = 0.0

Scoring

ctx_score = (file_size + skill_count) / 2

When a repo has multiple agent workspaces, each dimension is averaged across all workspaces found under agents/profiles/.

Score interpretation:

• 0.9–1.0 Agent profiles are lean — startup cost is low and skill surface is focused
• 0.7–0.9 Some profiles are getting heavy — worth a trim before they compound
• 0.5–0.7 Meaningful context tax on every session — plan a reduction pass
• 0.0–0.5 Startup cost or skill sprawl is likely degrading reliability — prioritise trimming

Tasks Generated

Finding	Priority
Composed profile > 800 lines	high
Composed profile > 300 lines (≤800)	medium
Manifest declares > 8 merged skills	medium

Out of Scope

• Enforcing context budgets in CI (future work)
• Deciding what to trim from any specific skill or profile — this perspective only measures and surfaces the signal
• Scoring composed/deployed profile formats (Claude Code Markdown output, Cursor Markdown output, Codex TOML output) — only the pre-composition Markdown source is measured

Agent Analysis Guide

When performing a manual CTX analysis:

1. Sum lines across a workspace's AGENTS.md, IDENTITY.md, ROUTING.md, SOUL.md, TOOLS.md — flag anything over 300, and treat anything over 800 as high priority
2. Count the skills array length in that workspace's agent.manifest.json — flag anything over 8
3. Look for repeated incident narratives, overlapping routing/storage instructions, or conditional protocols that could move behind an explicit reference instead of staying in the always-loaded contract
4. Recommend consolidating overlapping skills or narrowing routing for rarely-used ones rather than proposing a full rewrite

Output: a ranked list of oversized profiles or skill-heavy manifests, each with the measured value, the dimension it violates, and a recommended remediation.