v0.91.8

🌟 Release Highlights

This release focuses on richer, more reliable unified agent sessions across engines, stronger sandbox and security hardening, and better diagnostics for model routing and workflow failures.

✨ What's New

• Unified session timelines: agent sessions now render as full elapsed-time timelines, with consistent trace normalization for Copilot, Codex, Claude, Gemini, Pi, Kiro, Aider and AGY, preserving native evidence and accounting.
• Built-in graders: new code-read-precision trajectory grader, builtin grader defaults loaded from setup-action JSON, and unified Go/JavaScript outcome semantics.
• Cloud Hypervisor primaries with static script enclaves are now supported.
• Memory validation: declarative JSON schema validation for memory, with repairable invalid repo-memory baselines during the agent turn.
• Model routing: Copilot wire APIs are selected from AWF metadata, model mismatches are surfaced, and new model-routing smoke workflows verify behavior. Firewall bumped to v0.28.51.
• GitHub MCP mapping now covers governance and Copilot Automation tools.

🐛 Bug Fixes & Improvements

• Closed a custom grader sandbox escape and hardened workflow YAML, scanner file handling and temp-file writes.
• Fixed Copilot SDK model routing and cross-family sub-agents, and GPT 6.1 Sol pricing propagation to the firewall.
• Repaired post-agent state push retries and now expose git errors.
• Restored safe-output server startup after a memory schema change.
• Emit AWF agentTimeout for literal timeouts; audit/logs show delivered steering notices.
• Failure reports gain root-cause diagnostics and clearer incomplete-workflow messaging; bounded MCP compile diagnostics keep errors actionable.
• Red-team findings now require evidence before being reported.
• New Go linters (unchecked deferred close, walk-dir ignored error) and race-detector coverage for concurrent packages.

📚 Documentation

• Clearer safe-output permissions and trigger guidance, documented Claude add-wizard authentication, and a leaner safe-rollout guide. See Safe Outputs and Engines.

Generated by 🚀 Release · copilot · auto · 16.3 AIC · ⊞ 11.7K

---

What's Changed

• Dispatch engine conformance through a priority-aware work queue by @pelikhan with @Copilot in #67222
• Extending work queue protocol to support checkpoints and maintenance jobs by @pelikhan with @Copilot in #67105
• Preserve model routing attribution in audit and logs by @SivaKesava1 with @Copilot in #67114
• Share scoped KVM permission setup between NVX and Cloud Hypervisor by @lpcox with @Copilot in #67234
• Use full page width for zoomed slideshow snippets by @pelikhan with @Copilot in #67239
• Remove broad DEBUG logging from maintenance forecasts by @pelikhan with @Copilot in #67243
• Exclude Cloud Hypervisor HOME state from Git staging and agent patches by @SivaKesava1 with @Copilot in #67020
• Use repository labels for work-queue Issue status by @pelikhan with @Copilot in #67223
• Clarify why the ESLint factory cannot dispatch workers by @pelikhan with @Copilot in #67217
• Restore classifier cost attribution in routed-run audits by @SivaKesava1 with @Copilot in #67092
• fix: register work queue compaction tests with Vitest by @pelikhan in #67283
• fix: embed missing work queue session parser dependencies by @pelikhan in #67284
• test(cli): update work queue compaction fixture for checkpoints by @pelikhan in #67287
• fix: restore atomic queue bootstrap and split stress CI assertions by @pelikhan in #67288
• fix: avoid Docker Hub outages in Alpine container builds and validation by @pelikhan in #67286
• Add a slideshow-only title slide to the docs landing page by @pelikhan with @Copilot in #67285
• Centralize discussion reports and DeepReport in the work queue by @pelikhan in #67289
• fix: defer work-queue bootstrap until checked publication by @pelikhan in #67291
• Expose Codex standalone web search for OpenAI Responses Lite models by @pelikhan with @Copilot in #67246
• Scope PR push file policy to agent commits by @pelikhan with @Copilot in #67241
• Keep threat detection gated in routed workflows by @SivaKesava1 with @Copilot in #67272
• Add cross-module model-routing contract coverage by @SivaKesava1 with @Copilot in #67254
• Preserve agent-selected close_issue state reasons by @pelikhan with @Copilot in #67245
• Harden work-queue initialization for empty repositories by @pelikhan with @Copilot in #67297
• Allow allow-listed GitHub Apps to check out same-repository PRs after comment triggers by @pelikhan with @Copilot in #67247
• Add fallbacks for downstream activation repository outputs by @pelikhan with @Copilot in #67244
• [docs] docs: unbloat measuring-impact.md by @github-actions[bot] in #67314
• Allow Copilot coding agents through the bot allowlist without collaborator lookup by @pelikhan with @Copilot in #67238
• fix: preserve Codex native session evidence by @pelikhan in #67360
• Preserve Gemini session observations with standard event vocabulary by @pelikhan in #67355
• fix(docs): use hero background on slideshow title slide by @pelikhan in #67362
• Preserve Goose tool diagnostics and unified session evidence by @pelikhan in #67349
• Normalize Claude session diagnostics and snapshot evidence by @pelikhan in #67351
• Preserve DeepSeek partial sessions and startup failures by @pelikhan in #67352
• fix: preserve OpenCode session evidence and snapshot accounting by @pelikhan in #67354
• Keep OTLP configuration out of AWF agent sandboxes by @pelikhan with @Copilot in #67242
• Export sparse model-routing attributes to OpenTelemetry by @SivaKesava1 with @Copilot in #67240
• Fix Copilot session evidence preservation and scoped accounting by @pelikhan in #67368
• Preserve AGY native session evidence and accounting by @pelikhan in #67396
• Preserve Kiro headless conversations in canonical session artifacts by @pelikhan in #67394
• Preserve session error status codes in unified artifacts by @pelikhan in #67402
• Clarify incomplete workflow failure reports by @pelikhan in #67403
• fix: scope Impeccable reviewer to UI paths and bound review exploration (#66819) by @pelikhan in #67412
• Harden temp-file writes in clustering, cadence, and sous-chef scripts by @pelikhan in #67408
• Make AW work queues auto-bootstrap and evolve workers durably by @pelikhan in #67322
• Fix prompt-exhaustion misclassification and restore Avenger's bash tool access by @pelikhan in #67411
• Preserve Aider native session evidence and accounting by @pelikhan in #67395
• Bound MCP compile failure diagnostics without losing actionable errors by @pelikhan in #67424
• Add least-privilege failure reporting and systematic scenario matrix to Windows CLI integration by @pelikhan in #67410
• Prefer compatible local actionlint before Docker by @pelikhan in #67425
• Fix GPT 6.1 Sol pricing propagation to the firewall by @pelikhan in #67413
• fix: use pinned Prettier for repository lint targets by @pelikhan in #67434
• Diagnose Copilot assignment credentials only after failures by @pelikhan in #67409
• Fix console formatter stream contracts and guard stderr usage by @pelikhan in #67423
• Fix Pi standard session mapping and shared evidence preservation by @pelikhan in #67393
• [spec-enforcer] Enforce specifications for modelsdev, parser, repoutil by @github-actions[bot] in #67407
• [dead-code] chore: remove dead functions — 4 functions removed by @github-actions[bot] in #67443
• Add tracker-ID payload regressions for PR comments and issues by @pelikhan with @Copilot in #67441
• [actions] Update GitHub Actions versions - 2026-10-10 by @github-actions[bot] in #67378
• Make invalid repo-memory baselines repairable during the agent turn by @pelikhan with @Copilot in #67400
• Add declarative JSON schema validation for memory by @pelikhan with @Copilot in #67439
• Strengthen workflow recommendation guidance by @pelikhan with @Copilot in #67468
• Keep harness routing outcomes advisory to model attribution by @SivaKesava1 with @Copilot in #67280
• Align workflow tool allowlists with required operations by @pelikhan with @Copilot in #67472
• Restore safe-output server startup after memory schema change by @pelikhan with @Copilot in #67506
• Preserve episode identity, caller lineage, and unified session metadata by @pelikhan in #67518
• Close the custom grader sandbox escape by @pelikhan with @Copilot in #67475
• Add root-cause diagnostics to generic failure issues by @pelikhan with @Copilot in #67471
• [linter-miner] Add unchecked-deferred-close linter by @github-actions[bot] in #67481
• Unify Go and JavaScript outcome semantics and evidence rules by @pelikhan in #67519
• Add golden coverage for routed model audit analysis by @SivaKesava1 with @Copilot in #67253
• Fix Copilot SDK model routing and cross-family sub-agents by @SivaKesava1 with @Copilot in #67509
• Add model and engine misconfiguration diagnostics to agentic-workflows by @SivaKesava1 with @Copilot in #67488
• Add model-routing smoke workflows with deterministic runner/proxy evidence checks by @SivaKesava1 with @Copilot in #67523
• Repair post-agent state push retries and expose git errors by @pelikhan with @Copilot in #67474
• Harden workflow YAML and scanner file handling by @pelikhan with @Copilot in #67473
• Deduplicate memory allowed-extension parsing by @pelikhan with @Copilot in #67543
• Require evidence before reporting red-team security findings by @pelikhan with @Copilot in #67542
• Close recurring Go linter analyzer blind spots by @pelikhan with @Copilot in #67544
• [avenger] Fix stale work-queue integration fixtures by @github-actions[bot] in #67585
• [docs] docs: unbloat safe-rollout.md by @github-actions[bot] in #67582
• Select Copilot wire APIs from AWF metadata and surface model mismatches by @SivaKesava1 with @Copilot in #67487
• docs: document Claude add-wizard authentication by @pelikhan in #67624
• fix(linters): reconcile deferred Close error guidance by @pelikhan in #67631
• fix: enforce consolestderr on wasm and guard native/wasm CI parity by @pelikhan in #67634
• refactor(cli): embed DomainBuckets in WorkflowDomainsDetail by @pelikhan in #67642
• refactor(workqueue): centralize assignment receipt projections by @pelikhan in #67649
• Share scanner finding conversion across integrations by @pelikhan in #67643
• Preserve unified trace ordering and import OTel observations by @pelikhan in #67635
• refactor(cli): extract shared WorkflowRunFlags by @pelikhan in #67644
• refactor(cli): embed shared MCP stats in gateway server metrics by @pelikhan in #67646
• Add race-detector coverage for concurrent Go packages by @pelikhan with @Copilot in #67623
• Normalize engine session traces into shared events by @pelikhan in #67632
• fix(linters): safely resolve reflect.DeepEqual function aliases by @pelikhan in #67636
• [instructions] Split experimental ledger docs out of memory.md by @github-actions[bot] in #67685
• [spec-extractor] Update package specifications for tty, types, typeutil, workflow by @github-actions[bot] in #67691
• [spec-enforcer] Enforce specifications for scanfindings, semverutil, setutil by @github-actions[bot] in #67708
• Fix deterministic work-queue portfolio submission by @pelikhan in #67648
• Fix CGO failures in Aider attribution and generated skill lists by @pelikhan in #67751
• Preserve grader and eval attribution in unified sessions by @pelikhan in #67748
• Emit AWF agentTimeout for literal timeouts and show delivered steering notices in audit/logs by @SivaKesava1 with @Copilot in #67597
• Fix routing smoke assertions: attribute agent traffic by model instead of nonexistent purpose values by @SivaKesava1 with @Copilot in #67595
• Clarify safe-output permissions and workflow trigger guidance by @pelikhan with @Copilot in #67737
• fix: import only current-run grader and eval results into unified sessions by @pelikhan in #67761
• Add built-in code-read-precision trajectory grader by @pelikhan in #67742
• Render unified agent sessions as full elapsed-time timelines by @pelikhan in #67759
• Add governance and Copilot Automation tools to GitHub MCP mapping by @pelikhan with @Copilot in #67738
• [linter-miner] feat: add walk-dir-ignored-error linter by @github-actions[bot] in #67757
• Bump gh-aw-firewall to v0.28.51 by @lpcox with @Copilot in #67758
• fix: normalize unified session evidence and upgrade agent CLIs by @pelikhan in #67741
• Load builtin grader defaults from setup-action JSON by @pelikhan with @Copilot in #67777
• Support Cloud Hypervisor primaries with static script enclaves by @lpcox with @Copilot in #67772

Full Changelog: v0.91.7...v0.91.8