docs: record live sandbox deck proof

Author: Gini

AGENTS.md

@@ -50,3 +50,5 @@ Artifact tooling required by an acceptance gate must be image/bootstrap managed
 For legacy persistence migrations, assume identifiers were only unique inside one task directory. Derive globally unique durable IDs from conversation identity plus the original local ID, keep the transformation deterministic, and retain duplicate rejection inside a single conversation.
 
 For binary acceptance artifacts, prefer a versioned server-controlled renderer executed inside the sandbox. Let the agent author bounded structured content, but do not depend on it to install libraries, invent a renderer, or self-report binary validity. Record the renderer identity and verify signatures before extraction.
+
+Keep capability gates separable in live diagnostics. A backend/artifact proof may run with visual capture disabled when the provider route is unavailable, but it must be labeled partial and must not close the combined E2E ticket; preserve the exact provider failure and require a later run with visual evidence enabled.

docs/IMPLEMENTATION-LOG.md

@@ -95,3 +95,4 @@
 - Studied the AgentCore Claude/Codex harness's AWS chain and documented the accepted production design in `AGENTCORE-AWS-RUNTIME.md`: a sandbox-scoped AWS container credential-provider endpoint backed by short-lived STS sessions, never mounted profiles or copied static keys. Added `LIVE-E2E-ENGINEERING-LOG.md` so failed provider experiments and their fixes remain durable engineering evidence.
 - Prepared deterministic sandbox deck generation by adding managed `pptxgenjs`/`pdf-lib` bootstrap verification, a fixed six-file Slide deliverable contract, and a true mode-specific Claude tool availability list. Bash is omitted from ordinary jobs and enabled only for Slide mode; live Azure deck and negative capability tests remain required before closing the gate.
 - Closed two additional E2E design gaps found live: sandbox relay evidence now records LiteLLM from the actual server-injected route, and Slide artifacts use a server-controlled renderer executed inside the conversation sandbox after Claude edits a bounded structured outline. The managed renderer, rather than the model, owns PPTX/PDF construction and signature checks.
+- Proved the corrected managed Slide path live on Azure: a real LiteLLM-routed task returned signature-valid PPTX/PDF plus the complete six-file deliverable contract and validation evidence, then released its retained sandbox. The combined E2E remains open on Azure's missing visual-runtime route and multi-conversation continuation/isolation checks.

docs/LIVE-E2E-ENGINEERING-LOG.md

@@ -59,3 +59,7 @@ A fresh committed-code run proved the Slide-only tool list and managed Node modu
 Deck generation is therefore no longer entrusted to model-authored binary tooling. ONEVibe materializes an eight-slide structured seed and a versioned server-controlled renderer into the task workspace; after Claude finishes, ONEVibe executes that renderer inside the same retained sandbox, signature-checks PPTX/PDF there, and only then extracts artifacts. Claude may improve `outline.json`, but cannot redefine the binary renderer. This preserves sandbox-origin bytes while removing live installs and model-compliance variance.
 
 The first managed-renderer run produced all six required files and independently showed an eight-slide PPTX with `PK` magic plus an eight-page `%PDF-1.7` export. The post-agent enforcement invocation nevertheless failed because provider exec shells do not inherit the managed Node PATH. The renderer now uses the absolute `/opt/node22/bin/node` path; the failed task was retained only long enough to capture this evidence and must be explicitly released before rerun.
+
+Task `task_8d95ae8dc37b4e` then completed the corrected sandbox deck gate through the server-controlled LiteLLM route. It extracted `deck.pptx` (105,879 bytes), `deck.pdf` (5,327 bytes), `outline.json`, `speaker-notes.md`, `index.html`, `README.md`, and a passing validation report; the timeline includes the managed-renderer receipt. Its retained sandbox was explicitly released after inspection.
+
+This is not yet the whole ONE-221 gate. The full harness intentionally failed on missing X11 evidence because live visual capture was disabled after Azure returned HTTP 404 for `POST /v1/sandboxes/:id/visual/start`. Same-conversation continuation, distinct second-conversation allocation, restart verification, and the visual route still require one combined passing run.