Skip to main content
Certificate of Conformance #001. The first unit pulled back through Verify and Certify under the certification machinery, and the first to prove why re-certification exists: it opened red, and the defect was four months old.

​
The card

op-headless is a skill for headless credential-CLI operation; the credential specifics stay off this page by design (see Credentials and Security). A skill is prose consumed by whatever agent loads it, so it binds no model at runtime and lapses on a consuming-model swap.

​
Bill of materials

The BOM is content-hashed. A re-certification recomputes every hash against the file on disk; a single mismatch means the certified unit no longer exists, and the certificate fails closed.
  • skills/op-headless/SKILL.md, sha256 2b7512a5f11c6696a04aecfa1adff8b77a7fbdfeb6b1296aaa3b79f185542654 (the sole file constituent)
  • Part: none
  • Model: one binding in the certifying role; the exact model and version are recorded in the certificate on disk. No runtime model binding.

​
Evidence

  • Eval: the skill-validator gate battery, 5 of 5 required gates PASS (verdict PASS_WITH_WARNINGS, 2 advisories: environment assumptions and error handling). No dedicated skill-family eval suite exists yet; that gap is recorded as a finding.
  • Proof of work: the recorded gate output, all 5 required gates green. The pre-fix failing state is on record too (the code_syntax gate FAIL), so the defect and its fix both live in the evidence, not just the passing result.
  • Judge: none. The independent-judge harness is not built yet (a published gap). The probabilistic layer was manual: a second agent’s independent review of the delivery, one model reviewing another’s work.
  • Traces: the original production-use retro (issue #4) and the re-certification retro (WO-A7, issue #130).

​
Provenance

  • Work order: re-certification, WO-A7, issue #130.
  • Assembling agent: original assembly by a sub-agent; re-certification by a separate worker session.
  • Trial runs: none required. Supervised trial runs are the agent family’s gauge; a skill’s proof is its planted-defect fixture battery. Prior production use is recorded in the original retro.

​
Liability

  • Owner of record: the operator of record.
  • Escalation: the consuming agent or session opens an issue on the factory’s public repository, and the operator of record takes the call.
  • Recall: kill switch. Revert the skill file to the last certified revision, mark the certificate recalled, find root cause by Five Whys in a retro, fix the die (the skill or its validator) upstream, re-run the Verify gates, and re-issue.

​
Expiry

  • Re-certify by: 2026-08-10 (30 days from issue).
  • Triggers: a change to the skill validator (the gauge changed, so this certificate’s verdict no longer means the same thing); a dependency change in the underlying credential CLI or the private credential store’s auth model (see the private registry); a consuming-model swap; or 3 production failures.

​
The defect, published

Certificate #001 is not a clean-sheet story, and the page keeps it honest. The skill shipped on 2026-03-12, before the line had a syntax gate. It carried an unrunnable shell block: a conditional whose body was a comment only, which the shell rejects. It also referenced a command with no matching prerequisites documentation, so the command-reference gate failed on a clean continuous-integration machine where that command is absent. Both defects lived for four months. Worse, a 2026-03-19 test run saw the red gate and rationalized it as “expected, the snippet is intentionally incomplete” instead of filing a defect. A red gate that gets explained rather than fixed is a gate with a human bypass. Re-certification under WO-A7 caught both deterministically and fixed them: the shell block now emits a diagnostic, and a prerequisites section carries real install lines. The die fix is the machinery this certificate inaugurates. Expiry triggers, including “the gauge changed,” now force shipped units back through the gates so a red verdict can no longer be talked past. Published defects are the brand; this is the first one.

​
Self-audit