Skip to content
agentctlv1

Live framework verification

Date: 2026-07-27, Asia/Kolkata.

Status: passed.

The packaged macOS arm64 CLI and the production Linux arm64 OCI image executed the bounded public OpenAI matrix with model gpt-5.6. Every run used a clean temporary database and workspace. Raw provider responses and databases remain only in ignored temporary or evidence paths. No credential value or resolved secret was written to committed content.

Retained successful matrix

Section titled “Retained successful matrix”
Scenario Requests Tool calls Run Result
Stateless tool workflow 2 1 run-019fa39d-de49-7410-b973-d50fed711010 exact verdict and artifact
Basic agent 1 0 run-019fa39d-e8ca-7f73-b218-f105e4c16b92 succeeded
Environment secret reference 1 0 run-019fa39d-eeb6-79a1-9f65-879196c4193d succeeded without persisted secret
Provider portability 1 0 run-019fa39d-f6b2-7a21-946a-81d4b6d5269f succeeded
Framework live composite 11 0 run-019fa39d-feea-7ad2-9ff7-38e7bc005655 all bounded constructs succeeded
Deliberate agent failure source 3 2 run-019fa39e-3567-7bb3-a47f-496162a23d12 failed at the designed task boundary
Terminal retry of that failure 1 1 retry-019fa39e-4ce3-7773-a3d0-dcd18d46b85b upstream agent reused; failed task reattempted once
Selective repair 2 1 repair-019fa39e-55f9-7142-b6ae-06f21a1ead8e upstream reused; artifact CAS reference retained
Linux arm64 OCI failure source 3 2 run-019fa3a6-2fc6-7733-b201-eb806fd70425 expected failure retained in mounted state
Linux arm64 OCI selective repair 2 1 repair-019fa3a6-5b2e-7851-9a44-e3ea3e019ce0 upstream reused and artifact exported

The retained matrix used 27 Responses API requests, 8 tool calls, 3,939 input tokens, 560 output tokens, 20 reasoning tokens, and no reported cache tokens. The provider did not return billing metadata, so no billed amount is fabricated. The gate’s deliberately conservative token calculation is below USD 0.07 for this retained matrix and its hard ceiling is USD 10.

Composite feature proof

Section titled “Composite feature proof”

The public live-composite.yaml compiled to 21 tasks and made 11 bounded provider requests:

  • two independent agent branches executed under maxConcurrency: 4;
  • a two-child agent matrix retained stable child IDs and ordered aggregation;
  • a structured agent result selected execute, while the default route was durably skipped;
  • two agent loop iterations committed and the third bounded child was skipped;
  • one agent ran inside a versioned typed sub-workflow;
  • collector and reviewer roles exchanged one typed durable handoff;
  • one OpenAI streaming task persisted stream events and produced the exact final marker;
  • final deterministic verification succeeded.

Replay replay-019fa39e-34a3-71a1-a1d2-e7fcefbb66d4 reproduced the composite output without the credential and recorded zero fresh effects, tool calls, or provider sessions.

Recovery, CAS, and container proof

Section titled “Recovery, CAS, and container proof”

The selective-repair source deliberately gives publish only one turn. It therefore persists its tool call and fails before a final structured response. The identical-workflow terminal retry reused analyze, made exactly one fresh model attempt for publish, and failed again as designed. The repaired workflow raised the bounded turn allowance. Repair then reused analyze, executed only the repair closure, wrote the expected artifact, and succeeded.

Replay replay-019fa39e-6273-7212-a60e-5519e2d88a25 reproduced the repaired local result with zero fresh effects. The production image repeated the source and repair path as UID/GID 65532 with read-only configuration and workspace mounts plus writable state and artifact mounts. Container replay replay-019fa3a6-7187-7481-bb8f-abed11943457 ran with no credential and --network none, producing zero fresh effects, tool calls, or provider sessions.

Resource-budget proof

Section titled “Resource-budget proof”

Run run-019fa2f4-4252-7aa3-8aa8-72c31cb5dfcf allowed exactly one provider request. The first task succeeded with 18 input and 5 output tokens. The second reservation was denied before dispatch, and inspection recorded providerRequests with limit 1 and attempted value 2.

Request accounting

Section titled “Request accounting”

The successful retained framework matrix used 27 requests. Two fail-closed gate attempts were also made while implementing the evidence assertions:

  • the first stopped after 20 requests because it counted the model and tool effects together for the terminal retry;
  • the second completed all 22 local requests, then exposed that the live OCI block copied the credential-free repair fixture instead of the public OpenAI workflow;
  • the focused OCI continuation used the remaining 5 requests and passed.

Including the separate one-request resource-budget proof, this limitation burn-down used 48 live requests, below the authorized maximum of 80. The first failed attempt’s temporary database was deleted by the fail-closed harness, so its token counts are not reconstructed. No blind retry loop or live fuzzing was used.

Sanitized machine-readable evidence is retained at the ignored path .release-evidence/selective-repair/live-summary.json.

Canonical source: docs/execution/LIVE_FRAMEWORK_VERIFICATION.md. Verified against agentctl commit 2aeaa88fba71162206b5f08f5bda4f0150247e4f.