v0.5.2

---

title: Agent Workbench v0.5.2 release notes
doc_type: release-notes
status: published
owner: platform
last_reviewed: 2026-07-12

Agent Workbench v0.5.2

Highlights

• Adds repo:///orientation, a compact first-read receipt that directs coding
  agents to detailed status, scope, and overview resources only when needed.
• Makes context_for_task continuations executable and selective, with public
  arguments, decision reasons, expected evidence, intent-aware validation, and
  stateless suppression of repeated guidance.
• Improves first-read reliability and trust calibration so slow or unavailable
  work is reported as bounded skipped, stale, degraded, or blocked evidence.
• Extends workspace edit previews and applies to support safe creation of new
  files as well as replacement of existing files.

Added

• Added the public repo:///orientation MCP resource with snapshot-aware reuse
  guidance and links to repo:///status, repo:///scope, and
  repo:///overview.
• Added callable context_for_task continuation metadata with reason and
  expected_evidence fields.
• Added optional intent, changed_files, and satisfied_actions inputs so
  callers can make task ownership explicit and omit unchanged recommendations
  they already completed in the current task phase.
• Added cross-client activation and compatibility coverage for Codex, Claude
  Code, and Kiro guidance.
• Added new-file support to the preview-token and workspace-edit flow while
  retaining path, hash, token-expiry, and workspace-safety checks.

Changed

• Claude Code activation now uses concise conditional skill guidance without
  automatic SessionStart instruction injection.
• Task context recommends graph navigation and validation only when explicit
  task intent and available evidence make the action decision-relevant.
• First-read resources bound hidden work and expose missing provider or graph
  evidence instead of implying complete analysis.
• First-read metadata now preserves truncation, row budgets, skipped roots, and
  trust restrictions consistently across status and routing surfaces.
• Package manifests, plugin manifests, install URLs, and runbooks now target
  version 0.5.2.

Fixed

• Prevented generated continuations from exposing server-owned arguments or
  tools that are unavailable through normal client schemas.
• Prevented unrelated ranked symbols from causing references or impact
  recommendations.
• Preserved explicit validation as the primary recommendation when several
  action candidates compete for the response budget.
• Kept negated, conflicting, review-only, and otherwise ambiguous edit intent
  neutral instead of promoting mutation or validation guidance.
• Prevented unrelated SAM templates from influencing validation plans for
  tasks outside the SAM files or project shape.
• Reported changed-file diagnostics provider failures as limited evidence that
  still needs verification instead of presenting them as a clean no-op.

Documentation

• Promoted the first-read reliability and adoption-flow contracts into the MCP
  surface, runtime operations, graph store, coding-agent integration,
  edit-and-validation-loop, and runtime contract documentation.
• Closed Specs 036-038 and reconciled their archive, closure, and backlog
  records without retaining active implementation packages.

Validation

• awb release preflight --with-integration --allow-dirty
• pnpm typecheck
• pnpm validate:skills
• pnpm test:devcli (26 tests)
• pnpm test (79 files, 597 tests)
• git diff --check

Known Issues

• Agent Workbench is not invoked automatically at session start; agents must
  follow repository or provider guidance and use the packaged skill when the
  task warrants repository analysis.
• Persisted graph completion beyond the bounded first-pass warmup remains
  future work under backlog item EB014.
• Provider-history samples remain observational and too small to attribute
  session-quality improvements solely to Agent Workbench adoption.
• A combined navigation surface remains evidence-gated backlog work; it is not
  included without a controlled comparison showing lower interaction cost and
  equivalent evidence quality.

v0.5.1

---

title: Agent Workbench v0.5.1 release notes
doc_type: release-notes
status: published
owner: platform
last_reviewed: 2026-07-07

Agent Workbench v0.5.1

Highlights

• Repackages the Claude Code plugin guidance so Claude receives a concrete
  CLAUDE.md instruction file instead of relying only on SessionStart hook
  prose.
• Corrects stale Agent Workbench tool guidance in packaged hooks, skills, MCP
  instructions, and install docs so clients use tool discovery instead of
  hardcoded client-specific wrapper names.
• Refreshes package, plugin, and install metadata for the 0.5.1 release
  tarball.

Added

• Added repository-level CLAUDE.md guidance that points Claude Code at
  AGENTS.md and requires Agent Workbench-first repository orientation.
• Added packaged Claude plugin CLAUDE.md guidance for Workbench-first tool
  use, deferred schema discovery, and project-level Claude reload behavior.
• Added packaging and validation coverage so the root and plugin Claude
  guidance files are included in the release payload.

Changed

• Updated SessionStart hook output to use client-neutral Agent Workbench tool
  discovery wording instead of stale mcp__agent_workbench references.
• Updated Agent Workbench skills for Codex, Claude Code, and Kiro to avoid
  hardcoded MCP wrapper names.
• Updated MCP server instructions and Codex plugin documentation to describe
  deferred tool discovery without assuming a specific client namespace.
• Includes the committed docs-first warmup metadata and overview/session-start
  tightening work since v0.5.0.
• Updated package manifests, plugin manifests, install URLs, and runbooks to
  version 0.5.1.

Fixed

• Fixed Claude Code packaging so the directive Claude treats as project
  guidance is present in the plugin payload.
• Fixed packaged guidance that could point agents at non-existent or stale
  Agent Workbench tool names after plugin reload.
• Fixed validator and integration tests so missing Claude guidance files or
  stale hook wording fail before release.
• Includes the committed MCP daemon recovery fix for runtime upgrades since
  v0.5.0.

Documentation

• Documented Claude plugin guidance behavior and why hook prose is orientation
  rather than the only workflow authority.
• Updated Codex plugin runbook wording to avoid client-specific MCP tool names.

Validation

• pnpm validate:plugin
• pnpm validate:skills
• pnpm exec vitest run tests/integration/claude-plugin.test.ts tests/integration/hook-cli-compatibility.test.ts tests/integration/codex-hooks-install.test.ts tests/integration/codex-integration-profile.test.ts tests/integration/kiro-power.test.ts tests/mcp/stdio-entrypoint.test.ts
• pnpm typecheck
• git diff --check
• pnpm pack:dry-run
• node scripts/ci/hook-smoke.mjs

Known Issues

• Claude Code still requires plugin reload or restart behavior from the client
  after installing a refreshed package.

v0.5.0

---

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

Highlights

• 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.

Added

• 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.

Changed

• 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

• 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.

Documentation

• 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.

Validation

• 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.

Known Issues

• 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.

v0.4.0

Full Changelog: v0.3.0...v0.4.0

Agent Workbench 0.3.0

Spec 033 (cross-platform packaging). Agent Workbench ships as a normal npm package distributed through GitHub Releases (not the npm registry). Plugins register clone-free from the installed package — no repo clone, no --plugin-dir.

Install (Windows / macOS / Linux)

Prerequisites

• Node.js 22 recommended (Node 24 needs a C++20 toolchain for the native build).
• Python 3 + a C/C++ toolchain to compile native modules (Windows: VS Build Tools "Desktop development with C++"; macOS: xcode-select --install; Linux: build-essential).

1. Install the runtime (builds native modules + records the runtime pointer):

npm install -g https://github.com/Auriora/agent-workbench/releases/download/v0.3.0/auriora-agent-workbench-0.3.0.tgz

2a. Register the Claude Code plugin:

PKG="$(npm root -g)/@auriora/agent-workbench"          # PowerShell: $pkg = "$(npm root -g)\@auriora\agent-workbench"
claude plugin marketplace add "$PKG/plugins/agent-workbench"
claude plugin install agent-workbench@agent-workbench-local --scope user
claude plugin list   # -> agent-workbench@agent-workbench-local  v0.3.0  enabled

2b. Register the Codex plugin:

codex plugin marketplace add "$PKG/plugins/agent-workbench"
codex plugin add agent-workbench@agent-workbench-local

Restart the editor so the skill, hooks, and MCP server are discovered. Full guide: docs/runbooks/install-agent-workbench.md.

What's new in 0.3.0

• Normal npm package model. Removed the copy-to-prefix installer and archive install.sh/install.ps1; npm install builds natives in place and a postinstall pointer drives the in-place node shim launch.
• Clone-free plugin registration, verified end-to-end from the packed tarball for Claude and Codex (package-scoped agent-workbench-local marketplaces).
• CI green on ubuntu, macOS, and windows-2022 — native build + MCP initialize launch + hook smokes on all three.
• Distributed via GitHub Releases; not published to the npm registry.

Known gaps

• Kiro launcher not yet wired to the shim (docs/backlog/033-kiro-shell-free-launcher.md).
• windows-latest (VS 18) not yet supported by node-gyp; CI pins windows-2022.

Agent Workbench 0.2.0-rc.1 (cross-platform packaging prerelease)

Prerelease of the spec 033 cross-platform packaging work (shell-free install, MCP launch, hooks). Installs Agent Workbench as a properly registered Claude Code plugin — no repo clone, no --plugin-dir.

Install (Windows / macOS / Linux)

Prerequisites

• Node.js 22+ — https://nodejs.org
• Claude Code CLI, installed and logged in — https://docs.claude.com/claude-code
• Python 3 and a C++ toolchain (one-time, to compile native modules):
  • Windows: "Desktop development with C++" (MSVC Build Tools) + Python 3
  • macOS: xcode-select --install + Python 3
  • Linux: build-essential + Python 3

Windows

1. Download agent-workbench-0.2.0-rc.1.zip (below) and extract it to a permanent folder (e.g. C:\Tools\agent-workbench-0.2.0-rc.1).
2. Open PowerShell in that folder and run:

   .\install.ps1

3. Restart Claude Code.

macOS / Linux

1. Download agent-workbench-0.2.0-rc.1.tar.bz2 and extract it.
2. In a terminal in that folder:

   ./install.sh

3. Restart Claude Code.

The script builds the runtime, installs it to a stable location, and registers the agent-workbench plugin (MCP server + skill + lifecycle hooks) via claude plugin install. Verify with claude plugin list.

Notes

• Verified end-to-end on Linux (Node 24): build-from-source + MCP launch + hooks. macOS/Windows not yet run on a CI runner.
• The attached *.tgz is the npm package contract artifact; npm install of it is not supported yet (tracked follow-up) — use the archive + script above.
• Codex and Claude are the supported targets. Kiro launch is a deferred follow-up.

Agent Workbench 0.1.0

Packaged npm installer tarball for @auriora/agent-workbench 0.1.0.\n\nInstall from the downloaded tarball with:\n\nbash\nnpm install -g ./auriora-agent-workbench-0.1.0.tgz\nagent-workbench install\n\n\nThe package is not published to the public npm registry.