AIWG — September 2026

AIWG brings project guidance and reusable skills into the AI tools you use. Its setup puts those files where each tool can read them. September added tools for coding, saved decision records, and checks on civic sources.

TL;DR

September widened the choice of tools you can use with AIWG. Pi, Oh My Pi, Google Antigravity CLI, and DeepSeek Harness gained support. Grok Bot reached stable support, while Grok Build joined as an experimental coding provider.

The Decision Engine added an alpha contract that defines decision inputs, results, and saved run records. Civic Action added workflows for public sources, records, meetings, and corrections, with people still approving contact, submissions, and publication.

By the numbers

What's public	Value
Managed npm packages	`aiwg`, `@aiwg/cli`, `@aiwg/cockpit`
Stable provider milestone	Grok Bot in `2026.9.16`
New decision contract	`decision.aiwg.io/v1alpha1` and structured `v1alpha2` entries
Civic workflow kit	Opt-in `civic-action` addon
Docs	docs.aiwg.io
Public repository	github.com/jmagly/aiwg

Highlights

1. More tools can read the same project notes. What it is: Pi gained coding runs and session import; Grok Bot gained native skill setup. How you'd use it: choose Pi for a coding session, or deploy skills into Grok Bot using its own provider name. Why it helps: Grok Bot can load skills through its own setup, replacing the Cursor workaround.

2. Small decisions now have a shared contract. What it is: the opt-in Decision Engine defines inputs, rules, results, and saved run records across backends. How you'd use it: test a rule for routing a work item, then inspect the result and its record. Why it helps: you can switch decision backends while keeping the same inputs and result format.

3. Civic work has a source and review trail. What it is: Civic Action supplies structured review workflows for work in the public interest. How you'd use it: compare a meeting transcript with approved minutes and flag a vote that does not match. Why it helps: the evidence and unresolved questions stay attached to the review packet.

4. Data workflows have a clearer handoff. What it is: Dataset Intelligence defines sources, plans, run records, and checks. How you'd use it: preview a data import and inspect the record of what ran. Why it helps: checks connect each data plan to its run record and result.

5. Tests can be reviewed for what they prove. What it is: Testing Quality can collect test-run evidence and assess the checks inside a suite. How you'd use it: ask for a test inventory and a plan to repair weak checks. Why it helps: the review can include deliberate failures that a sound test should catch.

6. Session evidence can travel with its files. What it is: session export creates portable Fortemi Knowledge Shards from selected sessions. How you'd use it: review an export plan, build the archive, then verify and unpack it. Why it helps: supported attachments and registered outputs can travel with the session record.

Features shipped

Pi and Oh My Pi. Pi began with project notes, skills, and prompt files in 2026.9.1. Version 2026.9.2 added model discovery, headless runs, and session import. Session import preserves the session tree. Pi remained experimental, and its core had no MCP support.

Oh My Pi arrived in 2026.9.3 with its own experimental support. It supports native agents, model lookup, teams with set limits, and session import. Choose `omp` when that is the tool you run. Pi and Oh My Pi each need their own setup paths.

Google Antigravity CLI and DeepSeek Harness. Antigravity arrived in 2026.9.3 with project agents, skills, and a bounded headless adapter. Its checks used CLI 1.1.26 offline. Live model runs still lacked proof. DeepSeek Harness followed in 2026.9.5 with project notes, native skills, defaults that limit permissions, and session import. Both remained experimental.

Grok Bot reaches stable support. Version 2026.9.16 gives Grok Bot a stable provider name, skill setup, guides, and manual session-export import. Use `grokbot` when setting up the Bot. The migration guide covers projects that had used the Cursor workaround.

For user-scope skills, set an absolute `AIWG_GROKBOT_SKILLS_DIR`. AIWG stops when that root is missing rather than guessing where to write. After setup, start a new Bot chat or ask it to read the skills again. The stable status rests on Linux checks. The release waived checks on macOS and Windows/WSL.

The later `aiwg bot-handoff` command prepares a draft for a routine, teammate, connector, or memory reference. You can review that draft before creating anything in the provider. The extra Grok Bot native adapters still had blocked writers. Stable skill support did not let AIWG create those objects for you.

Grok Build gets its own path. Grok Build is the coding CLI, separate from Grok Bot and the web Build experience. AIWG added it by 2026.9.17, with support still marked experimental. Project skills use `.grok/skills`, and the context files point back to shared project notes.

Before a build, `aiwg build-verify --provider grok-build` checks that setup and tool lookup work. It also uses `grok inspect` to check what Grok Build sees. This helps catch missing context before a coding task starts. The check does not prove a model followed that context. Version 2026.9.21 added a test contract and a gate for stable status. The adapter still needed proof from the released CLI across the required platforms.

Decision Engine alpha. Version 2026.9.19 shipped one contract for Jev and ordinary LLM subagent backends. A definition states what to ask. A ruleset combines decisions. A binding selects the backend. A receipt is a saved record of a call. It lets you check retries and replay.

The contract includes instructions and criteria in a set format, pinned versions, and limits on retries and fallback paths. For a trial that routes work items, you can define the allowed result, try a backend, and inspect its record. Switching backends uses the same AIWG contract.

Version 2026.9.21 expanded the runtime with opt-in batching, human review, and an offline playground. It also requires a policy that selects which data may leave the process for a network call. The playground runs recorded examples through the real evaluator. The runtime stays off by default. Offline checks show how local code behaves. They do not prove a live model is right or grant the right to act.

Civic Action workflows. The September 1 release, 2026.9.0, shipped an opt-in kit for source review and civic writing. Start by asking the steward to set up Civic Action and check its tools. Then ask for a narrow artifact, such as a draft public-records request or a meeting review packet.

Each source record lists its origin, license and terms, how it was obtained, the legal region, and planned uses. Meeting work keeps the transcript, motions, votes, and approved minutes tied to sources. A missing vote stays unresolved. The workflows also cover public-technology review, local-resource profiles, corrections, and publication review.

Say a transcript and approved minutes disagree about a motion. The workflow cites the mismatch and sends a packet to a named human reviewer. It does not fill in the gap from a guess. A correction plan tracks changed claims and affected files. A reviewer can see what needs another check.

Local checks reject malformed records before checking policy rules. The rules block claims with no source, unresolved vote evidence, and publication without named approval of the exact file. People still decide whether to get source material, make contact, submit, approve a correction, or publish. A passing check only covers the declared fields. It does not verify every fact or settle legal questions. The kit did not ship adapters for local laws and rules.

Data, schemas, and portable sessions. Version 2026.9.1 added Dataset Intelligence and a catalog of data shapes with checks for changes that might break them. Later releases exposed the data runtime through `aiwg/dataset` and shipped the Fortemi capability contracts with it. Use these to check whether a planned data operation fits what the receiving tool says it can do.

Session exports use a reviewed plan tied to the selected records. The 2026.9.16 updates added available attachment and output bytes, plus checks against changes after planning. You can check an export before you restore it. An export will not write over an archive unless you choose to do so.

Writing controls and another provider. Version 2026.9.5 added opt-in author profiles, fact and audience briefs, protected text, and checks that edits keep the meaning. Use them to keep required terms and facts intact while reviewing edits. Choosing a profile alone does not rewrite every reply from your AI tool. Version 2026.9.21 also shipped experimental Muse Code support, including skills, a trusted context bridge, and export-first session import.

Fixes

A refresh for one tool leaves the other tools' files in place. A bundle install also leaves another bundle's agents, commands, and rules in place. These repairs landed in 2026.9.7.

Project skills remain discoverable when the project cache is missing, stale, or corrupt. Nested workspaces can also find checked shared bundles outside their own folder.

The lightweight CLI now stops an unsupported bundled setup early and points to the full package. Later doctor output names a stale install when it causes an unknown-provider error.

Session export rejects changed selections and files after planning. It also protects an existing archive from accidental overwrite.

Performance & reliability

The rule engine can keep compiled data checks in a cache with a size limit. It can reuse them on the next call.

The final September patches fixed CPU measurement and the test for the cap on active tasks. The CPU test gate now measures use by the thread under test. The dispatch-cap test holds work open while it checks a full queue. Discovery also warns about stale framework caches and gives steps to repair the affected cache.

Breaking changes & migrations

Code that uses the Decision Engine in 2026.9.21 must supply a policy that sets which data a network call can send. New batch, context, and cache result fields require readers that accept `v1alpha2`. Durable batch stores need keys to check records and encrypt stored results; old unkeyed stores need an approved move to the new format.

Version 2026.9.20 retired the `next` stable package alias. Use `latest` for the stable channel or choose an exact release version.

Releases

The shipped changes below use dates from the CHANGELOG.

• 2026.9.0 (Sep 1) — Civic Action source, meeting, correction, and publication review workflows.
• 2026.9.1 (Sep 4) — Dataset Intelligence, schema governance, and Pi resource deployment.
• 2026.9.2 (Sep 4) — Pi runtime, model discovery, and session integration.
• 2026.9.3 (Sep 5) — Oh My Pi and Antigravity support, stable Hermes, and project discovery fixes.
• 2026.9.4 (Sep 5) — clear recovery guidance for lightweight CLI setup.
• 2026.9.5 (Sep 7) — DeepSeek Harness, shared bundles, and author-controlled writing tools.
• 2026.9.6 (Sep 8) — test conformance review and reversible repair plans.
• 2026.9.7 (Sep 12) — scoped deployment cleanup and safer provider settings changes.
• 2026.9.9 (Sep 13) — clearer install checks, stale-file previews, and rule discovery.
• 2026.9.11 (Sep 14) — bounded waits for package proof and pinned test workflow actions.
• 2026.9.12 (Sep 15) — separate release workflow lanes and longer waits for package proof.
• 2026.9.15 (Sep 15) — a longer setup-test allowance to ship the session-export work described in 2026.9.14.
• 2026.9.16 (Sep 17) — stable Grok Bot, session attachment bytes, and export safeguards.
• 2026.9.17 (Sep 20) — packaged dataset runtime and contracts, with Grok Build in the provider inventory.
• 2026.9.18 (Sep 21) — executable dataset checks against prior release evidence.
• 2026.9.19 (Sep 21) — Decision Engine contracts, provider handoff drafts, and build verification.
• 2026.9.20 (Sep 21) — one stable package alias and release-source checks.
• 2026.9.21 (Sep 27) — expanded decision runtime, experimental effect ledger, and experimental Muse Code support.
• 2026.9.22 (Sep 27) — doctor guidance for stale installation identity.
• 2026.9.23 (Sep 27) — isolated CPU measurement for a decision test gate.
• 2026.9.24 (Sep 27) — dispatch-cap test repair and framework-cache diagnostics.

Dependencies & security

When it updates tool settings, AIWG keeps existing JSON and TOML values. It rejects input in the wrong shape. When a run saves receipts, the Decision Engine stops if the receipt key is missing or invalid. The decision runtime also rejects secret material in receipts and cache entries.

The experimental effect ledger records planned actions and what was seen after them. Its checks distinguish present, absent, and unknown results. That helps code keep an unknown result as unknown when it cannot prove the work is done.

Docs & DX

Tool guides cover setup, support limits, and how to remove managed files. Grok Bot has a migration guide; Grok Build has a guide to checks before a build. Civic Action's quickstart leads with tasks you can ask for and the review packet each task should return.

Session docs walk through import, choice of records, export, checks, and restore. Decision docs explain contracts, backend choices, replay, and how to move to new formats.

Tests & CI

Testing Quality gained reviews tied to source files, test-run records, planned failure cases, and repair plans you can undo. Checks covered real Vitest and pytest runs.

Data tests run against the packed package and the prior release contract. Decision tests include offline fixtures and a playground that uses the production evaluator. Checks with live services remain a separate step.

Cross-project impact

Fortemi: AIWG's session archives use Fortemi Knowledge Shards, and Dataset Intelligence carries Fortemi capability contracts. Live-server data checks were still pending in the 2026.9.18 evidence. These contracts do not prove that a server can keep or restore data.

Known issues & open threads

Pi, Oh My Pi, Antigravity, DeepSeek Harness, Grok Build, and Muse Code remained experimental in the September releases. Grok Bot's stable support includes skill setup and manual session import; native object creation remains blocked.

The Decision Engine still needs live checks for load, linked stages, and saved human review. Civic Action still needs people to check consent, how long to keep data, and who should review it.

What's next

Live tests must check decision load, linked stages, and saved human review before broader use. Civic work still needs source review and named people to approve contact, submissions, and publication. Grok Build needs checks of the released CLI on all four required platforms before it can reach stable status.

Appendix

• Managed packages: `aiwg`, `@aiwg/cli`, `@aiwg/cockpit`.
• Public repository: github.com/jmagly/aiwg.
• Release notes: GitHub releases.
• Public changelog: CHANGELOG.md.
• Docs: docs.aiwg.io.
• npm: aiwg, @aiwg/cli, @aiwg/cockpit.
• Window: September 1–30, 2026.