Harnesses

Harnesses are the differentiator. Instead of asking one agent to do a whole task in one context, a harness spawns focused sub-agents and adds adversarial verification so its output is trustworthy. They are Claude Code skills that live in .claude/skills/ of your scaffolded workspace.

Every harness is a template (tuned per workspace via args), read-only (it returns data; the caller writes), and demonstrated against the synthetic examples/sample-engagement/, so you can run every one with zero real data.

Shipped harnesses

These three ship today and are the heart of the system.

`/decision-audit`

Lints the decision log against the workspace governance rules and returns only verified findings.

Pattern: one verifier per rule + an adversarial verify pass.
Run it inside Claude Code, in your scaffolded workspace:
```
/decision-audit
```

`/scope-creep`

Flags deliverables that drift from the scope baseline/ledger, with a refuter pass that downgrades false positives.

Pattern: per-deliverable classify + severity-downgrade refuter.
Run it:
```
/scope-creep
```

`/transcript-decisions`

Turns meeting transcripts into novel, grounded decision-log rows. Pairs well with the Granola integration — export transcripts into 1-inbox/transcripts/ first.

Pattern: fan-out extract + dedup + adversarial grounding.
Run it:
```
/transcript-decisions
```

Workflow skills

These skills ship in the scaffold alongside the harnesses above.

`/aios-sync`

Review what’s staged to push, decide what to promote, then run aios push — the curated path from “I did work” to “the team can see it”.

/aios-sync

`/okf-traverse`

Answer a question by traversing the local OKF (Open Knowledge Framework) link graph. Works offline — no brain required. Pull the graph first with aios pull-bundle.

/okf-traverse

`/weekly-synthesis`

A rubric-gated weekly digest that revises until the rubric passes or the budget is spent.

/weekly-synthesis

CLI-driven loops

Two aios commands drive multi-agent loops from the terminal rather than as in-repo skills.

`aios analyze`

Builds an Agentic Engineering Maturity (AEM) report from your local agent-session logs (Claude, Codex, Cursor). It computes structural signals — delegation, verification, cost, and more — and places you on the AEM Spine. Runs offline; raw session text never leaves your machine.

aios analyze                       # last 7 days, printed
aios analyze --since 30d --report  # longer window + deep report
aios analyze --push                # send the daily aggregate (ratios + counts only) to the brain

`aios relay`

Runs an automated Opus 4.8 ↔ Cursor plan/review loop over a task: Opus plans, Cursor executes, Opus reviews, repeating for N rounds (default 3).

aios relay "add rate-limit tests"          # 3 rounds, /review-plan
aios relay "refactor parser" --rounds 2     # custom round count
aios relay "tidy imports" --dry-run         # plan/review without writing

Roadmap

More harnesses are deliberately deferred and make good contribution targets: a weekly-synthesis harness with a fidelity verifier, a ticket-hygiene harness, and a classifier-router that picks single-pass vs harness by input size. See the issue tracker for the current list.