Transparency
All work must be transparent. The system never operates in the dark.
Every agent action is surfaced on the dashboard in real time. Every tool call, every file read, every command executed — logged and explainable. If an agent skips a step, it says why. If a stage takes an hour, you see it take an hour. Failures are never hidden behind success messages.
- // Show what agents are doing in real time
- // Log every tool call, every file read, every command executed
- // Never hide failures behind success messages
- // If an agent skips a step, explain why
Fail honestly
If work will fail because the system can't support that action, articulate it directly and instantly.
Never attempt work the system knows it cannot complete. A model that can't tool-call doesn't silently degrade to text generation — it stops and says so. A repo that hasn't cloned doesn't produce "hypothetical code" — it blocks until it can do real work. Every failure answers three questions: what failed, why, and what the user can do about it.
- // No degraded fallback — stop and say so
- // No hypothetical work on unavailable inputs
- // No "tests pass" without actually running tests
- // Every failure: what, why, what to do
Trust through honesty
A user cannot trust a system that has lied to them. Every claim must be grounded.
Every output is grounded in real data. Research findings report what was actually found in the code — never invented patterns. Commit claims reflect commits that actually made it to disk. Metrics reflect measurements that actually ran. If the LLM's output can't be verified, the system flags it as unverified rather than pretending it's ground truth.
- // No synthetic research or invented patterns
- // No claims of writes that didn't hit disk
- // No "tests passed" without executed tests
- // Unverifiable outputs marked as unverified
- // "I don't know" is always acceptable