title: Agent Workbench v0.5.0 release notes
doc_type: release-notes
status: published
owner: platform
last_reviewed: 2026-07-06

Agent Workbench v0.5.0

Added trust calibration to public Workbench MCP responses so agents can tell
what each result is safe to use for, what it is not safe to use for, and what
must be verified before making stronger claims.
Added the per-repo MCP daemon and shared cache path so parallel agents can
reuse one repository runtime instead of each cold-starting independent
indexing and SQLite state.
Added evidence-backed release-note generation and tag-driven GitHub release
automation for repository releases.
Improved packaged Agent Workbench guidance for Codex, Claude Code, and Kiro,
including release-note skills and Windows MCP launch-shim fixes.

Optional meta.trust response metadata with shared
safe_to_use_for, not_safe_to_use_for, and must_verify_by vocabulary.
Shared trust policy derivation for repository resources, docs tools,
context routing, graph queries, diagnostics, validation planning, workspace
edit preview/apply, integration health, and structured error envelopes.
Golden MCP trust tests covering routing-only evidence, direct reads, static
diagnostics, planned validation, edit previews, applied edits, stale or
degraded states, and provider failures.
Per-repo daemon entrypoint, socket transport, shared runtime startup, daemon
health metadata, and integration tests for daemon launch and stdio proxying.
awb release notes evidence generation, release-note refinement skills, and
a tag-triggered GitHub release workflow.

MCP presenters now derive public response envelopes through the shared trust
boundary after warnings, errors, and sanitized response data are known.
Static Codex integration profile metadata is treated separately from live
integration-health evidence.
Docs and integration-health provider failures now return structured
provider_unavailable envelopes instead of bypassing the standard response
shape.
Markdown duplicate-heading checks are scoped by parent section, reducing
false positives for repeated headings in different sections.
Packaged Agent Workbench skills and hooks now avoid external
spec-lifecycle script references and use the packaged runtime guidance.

Fixed direct-read trust so precise direct-read claims require returned
direct-read evidence instead of relying on surface type alone.
Fixed Windows MCP launch-shim behavior for packaged plugin startup.
Fixed Spec 032 and Spec 035 closure metadata so archive and closure records
point at the correct implementation and cleanup commits.
Fixed backlog routing so completed trust-calibration work no longer appears
as the next active implementation spec.

Added durable runtime-contract and MCP-surface documentation for trust
calibration, failure-state semantics, public-surface coverage, and additive
compatibility.
Added and closed lifecycle records for the per-repo daemon, release-note
generation, workspace watcher cleanup, and trust-calibration work.
Updated release-process, plugin, Codex runbook, agent-readable changelog,
backlog, archive-index, and closure-log documentation.

Added contract, MCP, docs, daemon, release CLI, hook, package, and golden
tests for the new runtime, release, and trust-calibration behavior.
Recorded successful validation for Spec 035: pnpm typecheck, full
pnpm test, focused docs/contract/MCP tests, docs metadata tests,
lifecycle lint, git diff --check, live MCP smoke on this repo, and
cross-repo local smoke across 10 repositories.
Recorded successful closure validation for Spec 035: lifecycle close check,
lifecycle scan with zero active specs, docs metadata tests, Workbench
diagnostics for changed closure docs, and git diff --check.

GitHub PR metadata enrichment is not part of release-note generation yet.
Cross-repo MCP tool sweeps can still report blocked or degraded results for
harness verification limits and workspace-write safety skips; those are not
treated as trust-calibration regressions.
Broader turnkey native dependency and marketplace distribution improvements
remain routed through the backlog.