[Feature Request] Context-optimized `sources` file

[johnpyp]

I like including sources.json in the @ in my claude.md, because it means the model implicitly knows which projects are available already, and it also gives it a boost in using opensrc proactively without me asking.

My only issue is the file is pretty context inefficient atm. Could be significantly improved with either a "derived" or just more compacted format. Example:

612 tokens in Claude Opus 4.6:

{
  "updatedAt": "2026-02-17T02:49:53.178Z",
  "packages": [
    {
      "name": "zod",
      "version": "4.3.5",
      "registry": "npm",
      "path": "repos/github.com/colinhacks/zod",
      "fetchedAt": "2026-01-12T01:38:55.980Z"
    },
    {
      "name": "release-it",
      "version": "19.2.4",
      "registry": "npm",
      "path": "repos/github.com/release-it/release-it",
      "fetchedAt": "2026-02-05T22:15:56.222Z"
    },
    {
      "name": "@release-it/bumper",
      "version": "7.0.5",
      "registry": "npm",
      "path": "repos/github.com/release-it/bumper",
      "fetchedAt": "2026-02-05T22:15:58.211Z"
    },
    {
      "name": "tsdown",
      "version": "0.20.3",
      "registry": "npm",
      "path": "repos/github.com/rolldown/tsdown",
      "fetchedAt": "2026-02-10T17:55:23.209Z"
    },
    {
      "name": "@opentelemetry/exporter-trace-otlp-http",
      "version": "0.211.0",
      "registry": "npm",
      "path": "repos/github.com/open-telemetry/opentelemetry-js",
      "fetchedAt": "2026-02-12T17:25:06.685Z"
    },
    {
      "name": "kysely",
      "version": "0.28.11",
      "registry": "npm",
      "path": "repos/github.com/kysely-org/kysely",
      "fetchedAt": "2026-02-13T20:40:59.992Z"
    },
    {
      "name": "tsx",
      "version": "4.21.0",
      "registry": "npm",
      "path": "repos/github.com/privatenumber/tsx",
      "fetchedAt": "2026-02-17T02:49:53.178Z"
    }
  ]
}

233 tokens sources.yaml, using yaml + purls:

updated: 2026-02-17
packages:
  pkg:npm/zod@4.3.5: repos/github.com/colinhacks/zod
  pkg:npm/release-it@19.2.4: repos/github.com/release-it/release-it
  "pkg:npm/@release-it/bumper@7.0.5": repos/github.com/release-it/bumper
  pkg:npm/tsdown@0.20.3: repos/github.com/rolldown/tsdown
  "pkg:npm/@opentelemetry/exporter-trace-otlp-http@0.211.0": repos/github.com/open-telemetry/opentelemetry-js
  pkg:npm/kysely@0.28.11: repos/github.com/kysely-org/kysely
  pkg:npm/tsx@4.21.0: repos/github.com/privatenumber/tsx

could even be a derived txt which is more loose in favor of conciseness (without losing info the agent cares about).

197 tokens, loose "purl-like" sources.txt:

npm/zod@4.3.5: repos/github.com/colinhacks/zod
npm/release-it@19.2.4: repos/github.com/release-it/release-it
npm/@release-it/bumper@7.0.5: repos/github.com/release-it/bumper
npm/tsdown@0.20.3: repos/github.com/rolldown/tsdown
npm/@opentelemetry/exporter-trace-otlp-http@0.211.0: repos/github.com/open-telemetry/opentelemetry-js
npm/kysely@0.28.11: repos/github.com/kysely-org/kysely
npm/tsx@4.21.0: repos/github.com/privatenumber/tsx

a bit inspired by the vercel blog https://vercel.com/blog/agents-md-outperforms-skills-in-our-agent-evals#addressing-the-context-bloat-concern :)