v0.1.4

Measure the context your AI coding agent loads before you type a single word — and fail the build when it grows.

agentdiet is a single static binary that walks a repository, finds every file your agent runtime injects into the prompt — AGENTS.md, CLAUDE.md, .cursorrules, copilot-instructions.md, skills, MCP declarations — and reports what that costs: always-on tokens paid on every request, the on-demand surface, cost per 1,000 sessions, and an A–F grade against a budget you can enforce in CI.

It does not lint your code. It lints the instructions: the one part of the repository that every agent pays for on every turn, and the part nobody measures.

ALWAYS-ON   paid on every request, before you type
     442   0.22%    41 L  .claude/CLAUDE.md

TOTAL      442 always-on tokens  (0.22% of a 200,000-token window)
GRADE      A  (budget 1,500)   on-demand surface: 76,533 tokens
COST       $1.33 per 1,000 sessions at $3.00 per 1M input tokens

Use it as a step

- uses: indraoverflow/agentdiet@v0.1.4
  with:
    max-tokens: 4000   # fail the job above this always-on budget; 0 disables the gate
    window: 200000     # context window for the share column

input	default	meaning
path	.	repository path to audit
max-tokens	4000	fail when always-on tokens exceed this; 0 disables the gate
window	200000	context window used for the share column
version	latest	agentdiet version to install

output	meaning
always-on-tokens	tokens injected into every request
grade	A–F against the budget

Deterministic and offline: no API key, no network at audit time. The estimator (prose chars/4.5 + fenced-code chars/3.4, rounded up per segment) is calibrated against cl100k_base at a measured MAE of 8%.

Why this exists

Every agent runtime has a folder of files it loads unconditionally. They grow by accretion — a rule per incident, a table per architecture decision — until the agent spends 12k tokens of attention on stale instructions before it has read the question. agentdiet turns that into a number in the CI log and a budget that fails the pull request that would have made it worse.

MIT licensed. Source, docs, and real sample output: https://github.com/indraoverflow/agentdiet