@truefoundry@polaris_code How Feature Flag Reaper runs on TrueForge:
- reaper core = deterministic MCP tools: inventory → trace → classify → plan → apply + test → PR
- The agent follows a Git-backed skill.
- open_pr pauses for human approval
- The agent's token cannot merge
#agentsthatact
Zombie feature flags: dead code nobody can prove is dead.
Our agent proves it: real Unleash + git → verdicts w/ evidence → sandbox-tested removals → PRs you approve. Asks on every unknown. Never merges.
Built on @truefoundry TrueForge @polaris_code#agentsthatact
@truefoundry@polaris_code Filed my first upstream bug too. 👇
While running the agent, the TrueForge harness crashed when a registered MCP server went down - an unhandled rejection in the reconnection path.
Isolated it, reported it: https://t.co/AODgum8rPx
#agentsthatact
@truefoundry@polaris_code Update: (#agentsthatact, built with @truefoundry):
Every cleanup goes through:
- real Unleash traffic signals
- ripgrep tracing with file:line evidence
- a classifier that never guesses
- pytest before any patch ships
Then it stops at the merge. By design.
The reaper's rule: act where safety is provable; stop where it isn't. Dynamically constructed flag keys get a human question in chat - it cannot proceed without an answer. And it can't merge: the token has no merge scope. Every action lands in an audit log. #agentsthatact
The reaper's rule: act where safety is provable; stop where it isn't. Dynamically constructed flag keys get a human question in chat - it cannot proceed without an answer. And it can't merge: the token has no merge scope. Every action lands in an audit log. #agentsthatact