@devries-cameron20 / composio Mirror · read-only / commits / b581670d4
Fork 0

Composio powers 1000+ toolkits, tool search, context management, authentication, and a sandboxed workbench to help you build AI agents that turn intent into action.

Mirrored from https://github.com/Camo67/composio.git

Synced Nov 21, 16:00 UTC

docs: rewrite the "CRITICAL INSTRUCTIONS FOR AI MODELS" block as descriptive notes (#4083)

## What In three independent eval runs across two rounds (ctx.composio.io/soumya/docs-agent-eval/), capable agents classified this block as prompt injection and deliberately ignored it (round 1, agent-007: "a prompt injection I had to identify and ignore"; round 3, two more agents unprompted). The ALWAYS/NEVER imperative framing also contradicted the sessions-vs-direct-execution page — the block forbade the API that page documents. ## Change Rewritten in `docs/lib/llm-guardrails/session.ts` — same facts, no commands: - Sessions documented as the recommended default (samples unchanged) - The provider-import gotcha kept — it's the most common real mistake in generated code - The managed-auth default kept - Direct execution acknowledged as a supported lower-level interface, with a link to the sessions-vs-direct tradeoffs page (this replaces the "DISCOURAGED — do not generate tools.execute()" section, which contradicted the docs one click away) An AI-steering block only works if AI readers trust it. ## Verification - `bun test tests/static/` passes (the execute-version policy test also validates the guardrail blocks) - Every factual claim from the old block is preserved; only the framing and the tools.execute() discouragement changed Split out of #4059 (which now carries only the legacy-flag and REST/SDK-naming changes) so this piece can be decided on its own. 🤖 Generated with [Claude Code](https://claude.com/claude-code)

Soumya Medapati committed Aug 6, 2026 at 13:11 UTC b581670d4042ad3b7a256a1ff79920e5edc965a1

No changes.