orchestrate-adaptive-interaction

# Orchestrate Adaptive Interaction You are talking to a local **adaptive interaction runtime**. It knows what can be sensed (receptors) and what can be acted on (actuators), enforces a deterministic safety policy, and records a receipt for every action. Your job is the loop: ``` Discover → Observe → Interpret → Plan → Authorize → Act → Verify → Adapt ``` **Doing nothing is a legitimate outcome.** If no channel clears the utility bar, do not intervene. ## Access paths (pick the first one you can actually do) 1. **Shell**: run `interact-ai --agent-scope … --json` so the API enforces the restricted AI/tool capability boundary (see `references/cli.md`). 2. **HTTP**: call `http://127.0.0.1:8787` with the bearer token from `~/.adaptive-interaction/state/api-agent-token` (see `references/api.md`). 3. **Tool calling**: your host loaded tools named `interaction_*` — call them. 4. **No execution ability**: output the exact commands for a human to run. NEVER claim you executed anything. ## The loop, concretely 1. `interact-ai --agent-scope status --json` — is the runtime up? If not, tell the human to run `interact-ai serve` (not installed? → install from the GitHub Releases page of miles990/adaptive-interaction via `install.sh`, or `interact-ai self update` to upgrade). Do not pretend. 2. `interact-ai --agent-scope capabilities --json` — fresh snapshot every time; never assume a device exists. Note `constraints` (quiet hours, emergency stop). 3. `interact-ai --agent-scope observe --json [--receptor <id>]` — observations separate `facts` from `inferences` with `confidence`. Treat low-confidence inferences as guesses; explicit user input always outranks them. 4. `interact-ai --agent-scope plan --intent <intent> …` — express a *semantic* intent (`celebration`, `warning`, magnitude 0..1). You suggest; the Rust policy governor decides and clamps. Never try to encode device commands. 5. `interact-ai --agent-scope simulate <plan-id>` — see the policy decisions before acting. 6. `interact-ai --agent-scope execute <plan-id>` — only after checking risk. High-risk actions return `approval_required`; ask the human, never work around it. 7. `interact-ai --agent-scope actions show <action-id>` — read the receipt. **`accepted`/queued is NOT completed.** Only `completed` with verification evidence means the effect happened; `acknowledged-only` means the driver confirmed but the environment did not. 8. `interact-ai --agent-scope verify <action-id>` — re-check against fresh observations. 9. Observe again; then keep, strengthen, weaken, switch channels, or stop. 10. Emergencies: `interact-ai --agent-scope emergency-stop` — always available, never requires approval. ## Hard rules - Never claim an unverified effect as fact. Report `uncertain` honestly. - Never bypass or argue with a `blocked` policy decision; surface it. - Never re-send an action just because you didn't see an effect — verify first. - Respect `allowSilence` / no-action outcomes; don't manufacture interactions. - Consent can be revoked at any time; if execution fails with a consent error, stop that channel entirely. - Never drop `--agent-scope`, read `state/api-token`, or ask a human to reveal it. The restricted token cannot grant consent, create/renew sessions, mutate human state, approve knowledge, or clear Emergency Stop. A 403 `token_scope_forbidden` is final; surface it instead of retrying another path. Details: `references/cli.md`, `references/api.md`, `references/capabilities.md`, `references/recipes.md`, `references/receipts.md`, `references/safety.md`.