I builty an AI observability tool with Codev! #1226
mohidmakhdoomi
started this conversation in
I built X with Codev!
Replies: 1 comment 1 reply
|
This is awesome! Great work. |
1 reply
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I Built a Filesystem Observer for AI Agent Sessions with Codev!
I just shipped
ai-observev0.1.0 — a Linux-first tool that wraps any command (especially an AI coding agent) and reports the net files it created, modified, or deleted under your watched roots, with a local browser viewer.What I Built
ai-observe --session demo -- bash -c 'echo hi > note.txt' ai-observe-viewer .codev/observe/demo.jsonlWrap anything —
bash, or your actual agent (claude,codex,gemini) — then open the viewer. It answers: "what did this thing actually touch on my disk?" It layers two sources of evidence — livestrace(direct) and start/end snapshots (a backstop) — into one event stream where every event carriessource+confidence, so directly observed changes stay distinct from inferred ones.The Interesting Bits
strace/snapshot sit behind a small protocol, with room for fanotify/eBPF later.127.0.0.1and strips raw syscalls/argv/PIDs before rendering; the named shims are deliberately not installed. (On-disk artifacts are unredacted — treat them as sensitive.)Built with Codev
Orchestrated through Agent Farm (
afx) + porch — autonomous builders in isolated worktrees, gated specs/plans, and 3-way review on the risky PRs.The part I'm proudest of: two rounds of live-agent EXPERIMENTs (Claude Code, agy, codex) that diffed what the agent actually did against what ai-observe reported. They surfaced three real bugs the unit tests missed — a codex sandbox path-filtering miss, dropped
unlinkatdeletions, and a sidecar mislabeling a snapshot-only session as authoritative — each fixed in follow-up work after v0.1.0. "It compiled" is not "it works."All reactions