nulone

Nice. Does this work across multiple machines / moved repos (same project, different local path), or is it strictly path-based?

The ASCII output is the missing piece for AI-assisted coding workflows. LLMs can spit out Mermaid, but you can't see the diagram inline in a terminal/code-review context. This fixes that.

How did you catch it — scanner, review, or just noticed manually? I treat agent-generated diffs as untrusted by default now.

Makes sense — execution boundary is the right place.

I'd prioritize audit logs + correlation IDs, and short-lived creds per tool call. Do you expose tool capabilities to the planner without exposing creds?

How do you handle side-effectful tools and secrets? Least-privilege and auditing seem hard once any Python function can become a tool (scoped creds, revocation, logs).

Orchestration buys parallelism, not coherence. More agents means more drift between assumptions. Past a point you're just generating merge conflicts with extra steps.

NVLink postRxDetLinkMask errors show up right before the hang. Has anyone captured a bug report or stack trace while nvidia-smi is stuck to see what it's blocking on?

Looks like an airplane, fails under real inspection. Same smell in diffs: bugfix turns into refactor. Review eats the "velocity." Revert rate keeps you honest.

Merge cost kills this. Does the harness enforce file/ownership boundaries per worker, and run tests before folding changes back into the lead context?

The “craftsman to Ikea factory manager” line from the interview is the real headline here. AI does the fun creative stuff, you get stuck reviewing 2000 lines you didn’t write. Revert rate tells you more than any “10x” claim.

Triage time is the real cost. Even obvious junk takes a couple minutes to read. Surprised bounty platforms dont have a "prove you ran it" step -- version, command, actual output.

Cognitive overhead is real. Spent the first few weeks fixing agent mess more than actually shipping. One thing that helped: force the agent to explain confidence before anything irreversible. Deleting a file? Tell me why you're sure. Pushing code? Show me the reasoning. Just a speedbump but it catches a lot. Still don't buy the full issue→PR dream though. Too many failure modes.

I can never remember jq syntax.

Whenever I need to transform JSON, I spend 20 minutes guessing filters until something works.

So I built a CLI tool: give it input JSON and desired output, it generates the jq filter.

Example:

  Input:
  [{"name": "Alice", "email": "alice@example.com"},
   {"name": "Bob"},
   {"name": "Charlie", "email": "charlie@example.com"}]

  Wanted:
  ["alice@example.com", "charlie@example.com"]

  Generated:
  [.[] | select(.email != null) | .email]

1. Takes your input/output examples

2. Generates a filter, runs jq, verifies the output matches

3. If wrong, retries automatically

Works with local models (Ollama) or cloud (OpenAI/Anthropic).

~450 tests, MIT licensed.

Curious what edge cases break it.

Solid primitive. Two questions:

1. Crash edge case: If an agent executes a side-effect and dies before signing the receipt, is that action orphaned? Any WAL-style intent/completion model?

2. Multi-step workflows: Do receipts chain natively (parent pointers/Merkle) or via external linking? (I see storage/ledgers are out of scope, but curious about the linkage design.)

The negative proof angle (proving AI didn't touch prod) is compelling for compliance.