aid-guide

AID Official Guide

Use this skill as the authoritative operating guide for the installed AID release. Prefer the installed binary's aid <command> --help when exact flags may differ from the reference.

What the dispatcher owns

AID's premise is that the caller dispatching work is the best-informed component in the system. AID therefore declines to guess what that caller already knows, which puts these responsibilities on the caller:

1. Declare the task profile. --difficulty --budget --urgency --rigor, and --kind when the category matters. Undeclared values are stored as null rather than inferred, and a null tells every downstream decision that nobody knows. Data locality is a separate declaration: --egress any|local|private-network (default any), decided by the provider endpoint — not by CLI identity. local admits loopback only (localhost, 127.0.0.0/8, ::1); private-network admits loopback or RFC1918/link-local IPs plus .local / .home.arpa hostnames and does not widen local.
2. Declare tools and skills. AID applies no skill unless one is declared (--skill, or a project default). Omitting --kind describes every resolved toolbox tool rather than hiding some behind a guessed category.
3. Route by provider, not by agent name. A route is <cli>/<provider>/<model>. One exhausted route says nothing about another provider that reaches a model of the same class. Never dispatch to a weaker model on the provider pool the caller is already running on: a different provider is delegation, the same pool for a worse model is waste.
4. Verify by running. An agent's own report of success is not evidence. --rigor states the proof owed — draft compiles, standard runs the changed path and captures real output, critical adds an independent audit. A task's judgment is its derived TaskOutcome, not its lifecycle TaskStatus: only Verified and Delivered are success. Unverified is inconclusive and Broken is a verification failure; neither is success. A Done task assessed hollow_output or missing_final_delivery is also not success — those mean no deliverable was observed anywhere, not merely empty stdout.
5. Read unknown as unknown. A model, provider, or cost AID could not establish is reported as unknown rather than filled in with a plausible value. Several CLIs never name the model they ran, so unknown is the honest and expected answer for many tasks.
6. Keep briefs short. State the goal and the red lines; the implementation path is the thing being delegated.
7. Do not edit a directory while an agent is working in it. AID snapshots the dirty paths once, at dispatch, and excludes them when it rescues an agent's uncommitted output — so edits made before dispatch are safe. Edits made during the run are not in that snapshot and are indistinguishable from the agent's, so they are swept into the same rescue commit. Dispatch with --worktree <branch> and the agent works somewhere else entirely. Every rescue prints the files it staged; read that line rather than discovering it later.