AIWorkHub

<div align="center"> <img src="docs/assets/aiworkhub-hero.svg" alt="AIWorkHub — repository-native AI orchestration" width="100%"> </div> <p align="center"> <strong>The open-source control plane for multi-model AI coding agents.</strong><br> Plan work, delegate it to coding models, preserve project context and accept changes only when the evidence passes. </p> <p align="center"> <a href="https://github.com/shrec/AIWorkHub/actions/workflows/ci.yml"><img alt="CI" src="https://github.com/shrec/AIWorkHub/actions/workflows/ci.yml/badge.svg"></a> <a href="https://shrec.github.io/AIWorkHub/"><img alt="Website" src="https://img.shields.io/badge/website-AIWorkHub-1687e8.svg"></a> <a href="https://github.com/shrec/AIWorkHub/releases"><img alt="Release" src="https://img.shields.io/github/v/release/shrec/AIWorkHub?include_prereleases&sort=semver"></a> <a href="LICENSE"><img alt="MIT License" src="https://img.shields.io/badge/license-MIT-2dd4bf.svg"></a> <img alt="Python 3.12+" src="https://img.shields.io/badge/Python-3.12%2B-38bdf8.svg"> <img alt="VS Code 1.93+" src="https://img.shields.io/badge/VS%20Code-1.93%2B-0ea5e9.svg"> </p>

AIWorkHub is an open-source, local-first multi-model AI coding agent orchestrator for VS Code and MCP. It coordinates Codex, Claude, Copilot, DeepSeek and GLM workers through dependency-aware task graphs, durable project memory, source-code intelligence and evidence-based review. Each Git repository remains an isolated AI engineering workspace; no AIWorkHub cloud account or HTTP service is required. Visit the AIWorkHub product site or follow the getting-started guide.

The whole system

<div align="center"> <a href="site/assets/aiworkhub-system-architecture.png"><img src="site/assets/aiworkhub-system-architecture.svg" alt="AIWorkHub system block diagram: seats and MCP surface over a five-pillar control plane, a durable runtime and repository-local .aiworkhub storage" width="100%"></a> </div>

AIWorkHub follows one durable loop: Observe → Decide → Delegate → Verify → Promote → Learn. The map separates live foundations from the current hardening wave and planned capabilities. Live foundations include NeedFix and Roadmap intake, the Task DAG, Source Graph and durable context authorities, outcome-aware workforce routing, isolated worktrees, semantic edits, validation, sealed review receipts, provider-free replay and manager-controlled promotion. The active closure wave covers reviewer MCP delivery, exact liveness, finalizer leases, validation isolation, cross-platform parity, Source Graph refresh truth, storage retention and NeedFix lifecycle reconciliation. Planned capabilities include automatic NeedFix closure and TTL cleanup, provider controls, CPU-aware parallelism, richer evidence visualization and quality-calibrated release automation.

Continuous Audit as a Service (CAAS)

CAAS stands for Continuous Audit as a Service — AIWorkHub audits its own repositories continuously, as a service of the normal lifecycle rather than a step someone remembers to run. The canonical protocol is docs/CAAS_PROTOCOL.md. aiworkhub.caas_enforcement checks the protocol's automatically-enforceable properties on every guarded lifecycle transition, so a repository cannot drift out of compliance silently; aiworkhub.audit_system runs read-only, narrow-scope review passes that emit structured findings into NeedFix with provenance.

Naming correction. The expansion "Continuous Automated Assurance System" is wrong; the owner-canonical expansion is Continuous Audit as a Service. That incorrect wording lives in the separate UltrafastSecp256k1 README and cannot be fixed from this repository — it must be corrected in that repository.

What's new in 0.12.58

• The dashboard loads quickly again after a VS Code reload (the Claude Code usage scan is cached on disk and the snapshot reads its task cards in one pass). A rework after a "no change" attempt keeps credit for the files it inherited, and extension tests run even when the repository scratch folder cannot be created.

What's new in 0.12.57

• After this update, installing a new version switches the running runtime without a VS Code window reload. Manager tools answer much smaller by default (cost ledger, workforce catalog, blocked-rework recovery), review chains are skipped when a worker changed nothing, skill matching, Source Graph text search and roadmap counting are more accurate, and the manager chat panel renders long transcripts safely.

What's new in 0.12.56

• A rejected or blocked rework whose correct code was stranded (a false "no change" refusal after a failed validation, or a release bump landing in between) can now be relaunched or recovered; a provider API outage is treated as a temporary failure, not bad code.

What's new in 0.12.55

• GLM through VS Code LM is no longer rejected as a stale host while VS Code is open; OpenCode recovers its model list when a probe briefly returns nothing; a rework blocked only by an environment or provider failure can be relaunched with the same correct code, and a rework rebased onto newer canonical code no longer fails as "unchanged".

What's new in 0.12.54

• GLM through VS Code LM edits by first/last-line anchors and no longer finishes a task while an edit recovery is pending; every OpenCode Go model (for example Gemini Argon 4) is available and a slow opencode models probe no longer hides them; Source Graph bodygrep searches every indexed file; a new manager chat no longer receives a flood of old task wakes.

What's new in 0.12.53

• Semantic edit can bind a range by its exact first and last line and relocates a slightly shifted range instead of editing the wrong lines; it refuses replacements that drop a range edge or silently change backslash escaping. Routes enabled in models.json (GLM 5.3 through VS Code LM, OmniRoute through OpenCode) are launchable without a catalog row.

What's new in 0.12.52

• The Manager chat no longer fails with mcp_initialize_failed:mcp_request_timeout on the first start after an update; OpenCode 2.x workers on Linux find their Go/OmniRoute credentials in opencode.db; semantic edit refuses replacements that break a file's syntax; sandbox relay errors and validation-lane refusals are reported under their real cause.

What's new in 0.12.51

• OpenCode workers in the Windows sandbox can now use OmniRoute, or any owner-configured provider on 127.0.0.1, through a per-worker relay that serves only that worker and dials only the configured address.

What's new in 0.12.50

• Every Claude seat now runs the 5.5 family (Opus 5.5, Sonnet 5.5, Haiku 5.5) instead of older models behind the bare aliases; OpenCode workers can use an owner-configured custom provider such as OmniRoute; aged scratch storage is cleaned up; several task-recovery and accept-receipt fixes.

What's new in 0.12.49

• OpenCode runs that recover from a provider error are no longer failed; sandbox Y:/Z: drives are cleaned up after each run; a recovered task no longer relaunches by itself and now receives the manager's recovery instruction; accepting a review with a manager correction can no longer lose the workspace to background cleanup.

What's new in 0.12.48

• First published build with the 0.12.47 changes (Restart, Stop, Copy, Manager Chat no longer self-starts, OpenCode key and model fixes); fixes the CI blocker that held 0.12.47 back (NF-2026-01492).

What's new in 0.12.47

• Restart any pending, blocked or processing task from the dashboard; the server picks the right recovery path (NF-2026-01481, NF-2026-01483).
• Manager Chat gains Stop (switch model mid-run, NF-2026-01479) and Copy (NF-2026-01478), and never starts turns on its own after reopening (NF-2026-01475).
• OpenCode workers get their API key and only route-eligible models (NF-2026-01474, NF-2026-01477); the AppContainer Claude auth retry works (NF-2026-01484).

What's new in 0.12.46

• OpenCode models join the workforce chain and the Manager Chat picker (NF-2026-01454).
• Manager tool replies are compact by default, so each manager turn reads less tool output (NF-2026-01447).
• Claude workers follow Claude Code's own family aliases and run the current Claude 5.5 models; the picker shows the resolved id per alias.
• Manager Chat restores its session on open and gains Rotate and Close; the MCP preflight skips a redundant health probe (NF-2026-01435).

What's new in 0.12.45

• A lost Kilo or OpenCode worker prompt now fails closed (worker_prompt_not_delivered) instead of launching the CLI on an empty stdin (NF-2026-01423).
• Linux sandbox verified on WSL2 (Landlock+seccomp): the OpenCode chmod no-op answers success without changing modes while chown stays denied.
• New opencode_go_api_key_env auth helper reads the projected OpenCode auth file fail-closed.

What's new in 0.12.44

• Kilo/Grok workers on Linux run again: the chmod change from 0.12.43 now applies only to OpenCode workers (NF-2026-00153).

What's new in 0.12.43

• OpenCode workers start in the Linux sandbox: a chmod inside the worker reports success without changing anything, instead of killing OpenCode's server (NF-2026-00153).

What's new in 0.12.42

• OpenCode workers on Windows finish their work: the sandboxed CLI no longer hangs, and it reaches the AIWorkHub worker tools through the host bridge (NF-2026-01454).
• Live output shows OpenCode, DeepSeek and GLM worker steps as readable rows (reasoning, tools, progress).
• Non-ASCII provider output no longer shows up as unparsed lines.
• Kilo cannot run in the Windows sandbox; use Grok through OpenCode.

What's new in 0.12.41

• Rank and preflight agree on which workers can run a card (NF-2026-00731).
• A correct candidate whose tests only fail inside the sandbox can be re-measured on the host instead of being stranded (NF-2026-01030).
• Source Graph skips .gitignored files (NF-2026-01393).
• A rework that reproduces a sandbox-failed attempt with passing tests is accepted instead of refused (NF-2026-01405).
• ctest with no tests found fails validation (NF-2026-01338).
• One long lesson no longer costs a whole mined skill proposal (NF-2026-01432).
• Task-creation refusals say which template was closest and what differs (NF-2026-01441).
• Deleting blank lines or imports no longer blocks review with an unresolved-deletion unknown (NF-2026-01437).
• Hitting Claude's output-token cap retries automatically (NF-2026-01439).
• Transient-retry cards relaunch on their own when their wait expires (NF-2026-01438).

What's new in 0.12.40

• A sent-back card whose first attempt started from promoted-but-uncommitted files relaunches cleanly instead of failing with a false rework_base_drift (NF-2026-01431).

What's new in 0.12.39

• The dashboard Tool Recipes window shows each recipe: where it came from, what it does, how often it ran, how it exited and when it was last used (NF-2026-01424).
• Tool recipes grow per project: accepted learning commits mine recurring validation commands into proposed recipes automatically (NF-2026-01425).

What's new in 0.12.38

• The skill system runs end to end on its own: accepted learning commits mine recurring lessons into skill proposals (NF-2026-01412), evidence reaches every stage a card went through, and a skill backed by two independent accepted actors activates automatically (NF-2026-01411).
• One oversized skill no longer blanks the whole skills packet; it is excluded with a reason and the rest are still injected (NF-2026-01416).
• The dashboard Skills window shows each skill: what it does, its lifecycle, how often and when it was used (NF-2026-01399).
• Grok (Kilo) and OpenCode workers get their prompt on stdin, so normal cards no longer fail at spawn on Windows with command_line_too_long (NF-2026-01417).

What's new in 0.12.37

• Skills act