Codex Operator Sprint Back to offer

Code Agent Proof Sample

Example format only, using a synthetic public-safe repo scenario. The deliverable is not "the agent says it worked"; it is the trace of context, proposed change, review gate, evidence, and what is still unsafe to claim.

Scenario

Allowed Surface Public or synthetic repo, local branch, package scripts, browser preview, and generated proof notes. No private source, secrets, customer data, production credentials, or hidden logs.

Buyer Question Can an AI code agent complete one narrow product task while leaving enough evidence for a human reviewer to trust, reject, or safely scope the next run?

Operator Trace

01. Context Map Inspect README, package scripts, app routes, target components, recent diffs, and active blockers. Write down the exact task boundary before editing.

02. Agent Work Make the smallest working change that satisfies the task. Keep unrelated refactors out. Preserve user edits and local conventions.

03. Review Gate Stop before public release, account actions, sensitive data use, payments, claims about performance, or anything requiring production access.

04. Evidence Pack Show changed files, command output, rendered browser proof, screenshots if useful, unresolved risks, and the next exact action.

Evidence Ledger

Check	Concrete Evidence	Status
Scope fit	Task is one repo question, change-plan, review pass, or small safe edit. Anything broader becomes a separate sprint.	Pass
Context retrieval	Relevant files and scripts are named in the handoff. Unknowns are listed instead of papered over.	Pass
Implementation	Diff is limited to the named task. No unrelated cleanup, config churn, or private data movement.	Human review
Verification	Closest real test, build, lint, typecheck, browser smoke, or screenshot is run. If tooling fails, the exact failure is captured.	Pass or named blocker
Claim boundary	No benchmark, security, compliance, production-readiness, revenue, partnership, or endorsement claim without direct proof.	Stop if unproven

Sample Handoff

Result: The agent completed the narrow code task on a public-safe branch and produced a reviewable diff. Verification covered the relevant local test and browser smoke check. No production deploy, customer data, benchmark claim, or public post happened.

Reviewer packet: changed files, command output, screenshot or rendered URL, unresolved risks, and the exact next action.

Remaining caveat: this proves workflow discipline and reviewability for the sampled task. It does not prove general autonomy, production safety, benchmark superiority, or buyer ROI.

Next action: buyer reviews the packet in writing and either accepts, rejects, or narrows the next run.

Want this kind of proof packet built for your repo? Start with a written mini-audit or support the sample if it saved you time.

Buy mini audit Buy written audit Run code-agent proof loop: $39 OS Optional $5 receipt