Agent Workflow 0.3.6

The dashboard now provides a compact Current run panel and searchable Change run dialog. Select a row once to make it current; controls require fresh matching identity evidence. Worker menus open failure reasons and read-only logs, desktop panes resize, phone layouts stay usable, and theme/polling preferences persist with data-only refresh.

New runs can use bounded JSON argument files. Native inbound Codex frames now default to 16 MiB; answer, preview, evaluator and artifact limits remain separate. Existing saved runs keep their recorded runtime and limits.

Validation: 544/544 tests passed on Windows/Node 24.18.0, with no skips. Package check, smoke, isolated clean install and production audit passed; zero vulnerabilities. Desktop and 390px Chromium picker checks passed. Exact local verification receipt · Scope and browser limitations.

The owner approved a local verification exception for this release while GitHub Actions is paused. No hosted CI pass is claimed. Required branch protection was restored after squash-merging PR #11; Actions remains disabled.

Live append, new worker deadline controls, durable command facts and consumer observer receipts remain in separate development.

Agent Workflow 0.3.5

Agent Workflow 0.3.5 delivers the selected Map + inspector dashboard: compact blue header, dominant map, independently scrolling desktop inspector, readable phone cards and a closable sheet. Full workers, secondary controls and raw evidence remain accessible on demand.

Dismissed messages stay in local history. Polling preserves keyboard focus, Escape closes the foreground panel, stale full-worker reads cannot replace newer attempt evidence, and new run failures remain visible. Light-mode status and dependency contrast improved.

Cleanup receipts retain bounded helper exit and acknowledgement diagnostics. Strict drain requirements are unchanged. The original unexplained cleanup failure remains open and its output remains unaccepted.

The unpublished 0.3.4 tag passed 505/506 tests but timed out in an existing PowerShell/COM alias fixture before catalog assertions. That tag remains unchanged. Version 0.3.5 uses CMD's built-in alias expansion in that fixture, preserving its deadline, filesystem identity and junction checks; production policy is unchanged.

Validation: 506/506 local tests; package check, process smoke, clean production-only native installation and zero production dependency vulnerabilities; independent review; 81 painted browser checks across five viewport sizes in both themes; installed command-only parent-policy doctor with zero inference requests and positive owned-job cleanup. Required Windows PR and tag verification passed before publication.

Attached acceptance records bind the release and installed runtime to exact source, Codex and dependency identities. Earlier runtimes and saved data remain preserved.

Agent Workflow 0.3.3 — parent worker access and verified skills

Agent Workflow 0.3.3 delivers the worker access and failure reporting fixes found while preparing PlanetForge workflows.

• New runs capture explicit parent filesystem, network and approval authority, and automatically provision original enabled native plugin/skill roots. Explicit restricted mode remains available.
• Selected skills require observed native initial main-resource reads. A post-model check rejects removed/changed selected sources; saved catalog identity blocks resume/restart with changed capabilities.
• Windows uses the packaged Microsoft Public Trust signed job helper, with source/binary, approved-profile and timestamp verification. The 0.3.2 Session 0 behavior is preserved.
• Capture and retained diff bounds rise to 512 MiB. Metadata/previews remain bounded, with integrity verification and reviewed replay.
• Startup failures retain diagnostic and dashboard evidence. Workflow authoring explicitly covers skill selection, tool prerequisites, live observation and recovery boundaries.

Validation: 485 local tests, package check, smoke, clean production-only installation and zero production dependency vulnerabilities. The installed final package passed command-only parent-policy doctor with zero inference requests and confirmed owned-job cleanup. Native acceptance observed all eight required initial skill reads; signed artifact identities were independently reviewed. Required Windows PR and tag verification passed before publication.

Windows CI exposed two additional fixes: verified short-name aliases are accepted by filesystem identity, and the signature verifier binds its required Windows built-in PowerShell modules. Verifier command failures are distinguished from returned signature rejections. Trust policy and the 15-second verification deadline are unchanged.

Parent authority is an explicit coordinator assertion. Full-access worktrees organize review; host-only connections are not cloned. Source checks detect drift rather than prevent cache deletion. Full skill-body pagination and instruction compliance remain unverified, and live skill grants are unsupported. Paired authoring drafts failed planning acceptance and are not used for delivery. PlanetForge's nine objectives remain unaccepted; this release delivers the workflow prerequisites.

The attached acceptance receipt records source/runtime identities and scoped evidence. Earlier failures are preserved. The test fixture's outer runner budget changed from 10 to 60 seconds; production subprocess deadlines did not change.

Agent Workflow 0.3.2 — Windows Session 0 compatibility

Explicit Windows MXC Session 0 compatibility for OpenSSH-launched workflows:

• doctor and new run accept --session0-desktop shared; default desktop inheritance remains unchanged.
• Native session enforcement and fixed same-session desktop selection preserve existing Codex, filesystem/network policy, tokens, ACLs and process ownership.
• Saved runs retain the selection across resume/restart. Doctor and failure receipts distinguish requested/applied compatibility and preserve cleanup evidence.
• Setup/run/mode skills explain existing authorization and measured shared UI exposure. No Codex fork is required.

Validation: 422 local tests, check, smoke, clean package installation, zero production vulnerabilities, and successful Windows PR/tag CI. Exact final package passed elevated HAL9000 OpenSSH Session 0 doctor, real read command, detached isolated write, attached live observation and cached resume. Ordinary desktop defaults passed; shared mode outside Session 0 was rejected before launch. Attached acceptance records the exact runtime/source/dependency identities.

Shared UI visibility is an accepted compatibility tradeoff, not equivalent UI isolation. Non-admin tokens, other Windows builds/accounts and arbitrary descendants outside owned jobs remain unverified.

Agent Workflow 0.3.1

Workers can now activate explicitly selected skills through genuine native Codex skills.list and skills.read tools. Detached workflows also have a durable, session-bound JSONL watcher, so an attached coordinator can receive worker/run terminal records, resume from a saved cursor, and notice missing command or file progress.

The missed layout failure came from detached execution without a standing observer. This release adds durable observation, explicit observer ownership in the workflow skills, and early artifact checkpoints. The original layout worker remains an unknown timeout with an empty capture; its internal stall cause is unconfirmed. See the root-cause report.

• Immutable startup capability bundles select original skill resources and required managed tools. Tool-dependent workers check actual command readiness before inference.
• The workflow authoring skill explicitly maps every worker to selected skills and required tools before startup, including an explicit none where appropriate.
• Public doctor diagnostics report bounded admission stages and cleanup evidence. Failed and interrupted turns retain available command-status evidence without treating partial history as complete.
• Isolation, capture/replay integrity, owned-process cleanup, and unknown outcomes remain fail-closed. Existing runtimes and saved runs are preserved.

Validation: 409 local tests passed; protected Windows PR and release-tag verification passed test/check/smoke, clean-package installation and dependency audit. Marketplace and the installed package select this exact release. Installed command-only doctor passed with zero inference requests; a fresh three-worker native acceptance run completed with positive owned-job/native-inventory cleanup and an attached watcher through terminal cursor 4. Browser checks passed across eight phone/desktop/theme/zoom-equivalent views with no page errors. Sanitized installed verification reports are attached.

One paired read-only planning comparison produced correct routes in both variants. Selected guidance added concrete handoff and recovery details, while taking 2.93× the time and 2.31× the reported total tokens (including cached input); no general improvement claim is made.

Live skill grants are unsupported by pinned native Codex 0.162.1 and are outside this release, per owner scope. The watcher requires an attached host consumer; it does not create automatic native parent-chat notifications. This release does not claim fresh sprite-export, WAV-render/listening acceptance, or a fix for the separate host's unconfirmed Windows command-admission failure.

The three new dashboard layout proposals are available on Sites. They are review options, not a selected production dashboard redesign.

Agent Workflow 0.3.0 — session delivery and safe lifecycle controls

Agent Workflow0.3.0 turns the Tidepool case study improvements into a complete session workflow delivery surface.

• Workflows and dashboards remain available through legitimate work until completion or manual Stop/Close. Explicit deadlines retain bounded worker, CPU and memory protections.
• New decomposition authoring guidance and recipe carry scout evidence into disjoint writers, reviewed integration and fresh verification.
• Immutable capture policies protect assigned source outputs while disclosing excluded dependency/build metadata through bounded CLI and dashboard pages.
• Session dashboards add blue responsive navigation, phase grouping links, separate execution/task/health evidence, reversible visibility and reviewed safe retirement with durable receipts.
• Native output retention, final-answer selection, prerequisite handling and artifact failure diagnostics preserve honest failures and uncertain outcomes.

Validation:246 tests, package checks, detached process smoke, clean native plugin installation and zero-vulnerability production audit. Required Windows PR and tag CI passed. Rendered checks cover eight viewport/theme combinations, phase connectors, session navigation and zero page errors; fresh managed-ChatGPT source verification completed with clean job/native-terminal shutdown. Independent review findings and the Windows CI path-alias regression were fixed before release.

Boundaries: task reports remain unverified without authoritative acceptance receipts. No real session data was pruned, no historical failed attempt was relabelled or replayed, and old runtime directories remain immutable. SSH and broader native-host parity remain follow-up work.

Published installation verified: the enabled marketplace plugin matches all79 tagged runtime files byte-for-byte. Explicit setup and zero-inference doctor passed; a fresh installed-runtime native run completed1/1 with clean drain. Eight rendered viewport/theme combinations passed against the installed dashboard. Detailed installed verification and browser reports are attached.

agent-workflow 0.2.1

Display the plugin as agent-workflow in Codex and local marketplace metadata. This metadata-only patch leaves the 0.2.0 worker runtime unchanged. All 166 tests, package and smoke checks, clean native installation and production audit passed locally, on the PR and on the release tag.

Agent Workflow 0.2.0

Agent Workflow 0.2.0 replaces the old runtime with native Codex workers using managed ChatGPT subscription authentication. No custom MCP or API-key billing fallback is required.

• Validated structured results, model/reasoning settings, live tool activity, individual steering/cancellation and explicitly reconciled restart.
• Isolated writable Git workspaces with retained diffs for review.
• Authenticated graphical monitor with dependencies, status and discovered progress; System/Light/Dark themes behind the gear button.
• Scoped saved workflows, reusable commands and opt-in session/project mode.
• Fingerprinted managed installations that retain each run's original runtime/session and preserve prior 0.1.x files.

Validation includes 166 Windows tests, clean native plugin installation/upgrade, dependency audit, paired Yoda skill checks and real subscription-backed worker/control/writer evidence. The installed candidate completed a saved native workflow, verified owned-job cleanup and replayed without another model turn.

Windows is verified; SSH is deferred. Broader Ultracode parity boundaries remain documented in docs/codex-parity.md. Existing old runs are preserved with their old runtime, not automatically migrated.

After marketplace refresh, install agent-workflow@hartye-plugins and run the workflow-setup skill. Open the full dashboard URL returned by the CLI to authenticate.

Agent Workflow 0.1.1

The monitor now follows actual phase/worker dispatch with an Agent Marketing inspired navy/amber design. Routine dashboard and CLI observations use bounded summaries; full retained evidence is available through deliberate inspection.

Recovered tool failures remain visible, full observations carry timestamps, truncated approval requests require inspection, and resume shows the run's immutable permissions and limits. Journal paging preserves every range after large live jumps. Existing full CLI/API contracts and matching old runtimes are preserved.

Verified on Windows: 139 tests, package checks, detached-process smoke, clean native cold/warm installation, production dependency audit, independent review and release-tag CI. The three skills retain the original Yoda audit; the changed CLI reference passed inspection and a matched native planning pair. Rendered layout/accessibility is unverified without an attached browser. list --compact summarizes every run; it does not impose a total list-size or pagination cap.

The real PlanetForge audit hit its evaluator deadline after two inspections. Verification and other slices remain incomplete; progress-journal throughput is the next runtime fix. This release does not claim completion of PlanetForge's roadmap.

Agent Workflow 0.1.0

Agent Workflow 0.1.0 is a regular Codex plugin for durable scripted work with parallel workers, item-local pipelines, bounded loops, one-level composition, saved workflows and ordered replay. It includes recovery/control tools, native Codex app-server and exec adapters, retained writer worktrees, six recipes and an authenticated local dashboard.

Install from the tagged repository or attached source package, add its local marketplace, and run the workflow-setup skill. Node 24+ with npm and an authenticated Codex account are required for actual model workers. Managed setup installs pinned dependencies outside the plugin cache. Hook trust is a separate explicit host action; manual workflow tools work without it.

All three skills received a Yoda audit, independent inspection and matched actual-Codex planning comparisons. Recovery guidance, upgrade ordering, exact installed data paths and PowerShell quoting were corrected. All 11 routing cases matched the expected boundaries. See docs/skill-audit.md for evidence and limits.

Verified with 123 offline tests, package/schema checks, real process smoke, isolated native cold/warm setup and license copying, official skill validation, zero npm audit vulnerabilities, authenticated worker execution/edited replay and independent reviews. GitHub's pinned Windows CI runs the offline/release checks without a model turn.

This release owns a local runtime/dashboard. Native composer/task-panel UI, shared hard token pools and live parent grants/history remain host gaps. Windows is verified; Linux/Darwin recovery and native hook trust activation were not exercised. Unknown effects require explicit review and verified drain before reconciliation. This is a GitHub/manual-install release, not a public Plugins Directory submission.

Original code is MIT. The vendored Agent Plugins schemas retain Apache 2.0 licensing and attribution.

Agent-tests audit and strategy are complete: owner-approved production risk map, exact control forwarding assertions, isolated Windows fixtures, five dashboard regressions and mutation rejection, explicit subprocess subjects and merge lanes. Required Windows PR verification uses job-internal documentation filtering, superseded-run cancellation, nightly and tag gates. See docs/testing.md for retained warnings and measurement limits.