Skip to content

Latest commit

 

History

History
99 lines (70 loc) · 4.01 KB

AUDIT.md

File metadata and controls

99 lines (70 loc) · 4.01 KB

Is your agent setup actually verifiable?

A free diagnostic. It takes about thirty seconds and it writes nothing.

git clone https://github.com/faresrafat3/agent-handoff
cd agent-handoff && ./install.sh
python3 bin/agent-handoff adopt          # read-only

adopt scans the markdown notes you already keep and reports four things it can determine without a model:

Finding What it means
hollow-section A heading like Current Status or Next Steps with nothing under it. Looks complete; carries nothing.
vague-next-action continue the work, finish it, next steps — no path, no command, no id, nothing runnable.
unknown-commit The record cites a commit that is not in your repository. It is describing a world that has moved on.
secret-pattern A credential in plain text in a note — which lands in git, in every clone, and in every context pack built from it.

It prints a count per kind and a fix for each finding. It creates, moves, and deletes nothing, and it never calls a model.


Read this part before you decide anything

If adopt reports zero findings, you do not need anything I sell. That outcome is common, and it is the correct answer for a lot of teams. Close this tab.

If it reports findings, you now have a number instead of a promise. That number is the whole conversation.


What the number does not tell you

It cannot tell you whether your agent is producing the right answer. Nothing mechanical can, and any vendor who says otherwise is selling a guarantee they do not have. That limit is documented as gate D3 in acceptance-matrix.md rather than hidden, because a tool that implies more than it can do is worse than no tool.

What a finding does mean is narrow and checkable: something you are relying on is not true right now, and nothing in your current loop would have told you.


If the number is not zero

That is a solvable engineering problem, not a rewrite. In rough order of what usually moves the number most:

  1. The next actions. Most of the count is usually this. It is also the cheapest: name a path, a command, or the check that closes the work.
  2. The checkpoints. Re-record from git rev-parse HEAD rather than from memory. Minutes of work, and it stops the record from describing a state that has moved on.
  3. The secrets. Rotate, then remove. A redaction at output time is not a fix — the value is still in the file, in git, and in every clone.
  4. The hollow headings. Either fill them or delete them. A heading with nothing under it is worse than no heading, because it reads as coverage.

Most teams that run this find that steps 1 and 2 clear most of it in an afternoon, without adopting any tool at all. That is a perfectly good outcome and I would rather it happened than not.


If you want it handled rather than diagnosed

I do a fixed-scope, one-week Agent Reliability Audit. Same analysis, then the fixes, wired in.

  • $1,800 — one agent pipeline
  • $4,500 — up to three pipelines, or one plus the regression harness in CI

You get a failure-mode map, a working regression suite in your repo, a prioritized plan, and a recorded walkthrough. 50% up front.

About a third of audits end with me telling you the real problem is somewhere else, and referring you on. That is a real outcome, not a sales tactic.

Start here: [email protected] with the repo URL and a 30-minute slot. I will tell you within two days whether this is the right thing or whether you do not need it at all.


Why the free part is genuinely free

The tool is MIT, single-file, dependency-free, and vendored into your repo. It has no account, no telemetry, and no network call in any gate.

I would rather a hundred teams run the diagnostic and fifty of them conclude they are fine than have one team pay for a fix they do not need. The diagnostic is also the first day of the audit, so the paid work starts from a number rather than from a pitch.