Agent capability profiles (Standard / Full Access)¶
Terminal fallback update (2026-10-12): ADR 0017 replaces command-form and network approval heuristics with canonical launch-directory policy. Terminal classifications and reason codes below are historical.
Current decision: ADR 0011 defines the canonical
plan/standard/full_accessmatrix, explicit canonical-identity trust, and native multi-call behavior; ADR 0021 moves the structured protected-path denial toplanandstandardonly. The sections below preserve this ADR's superseded rationale and history.
Ru exposes one generation mode: Agent. The Worker v2 settings currently advertise the same ten file/terminal bindings in Standard and Full Access. Standard applies the stricter Broker analyzer/rule/approval decisions; Full Access broadens policy fallbacks for well-formed calls. Workspace trust, hard invariants, protected paths, and extended timeout thresholds remain authoritative in both presets.
Grants remain session-scoped. Every launch starts in Agent Standard, Full
Access requires one explicit confirmation per run, and no grant survives
session replacement or restart. There is no Chat compatibility transport or
read-only generation profile. Elevation, admin, registry, and service
installation stay out of scope with no adapters. The picker was
retired in 2026-09-04; the current pending-approval model exposes
external_directory_required, terminal_delete_requires_approval,
terminal_external_mutation_requires_approval,
terminal_timeout_requires_approval, web_approval_required, and
terminal_unknown_executable as the ask reasons.
Recorded limitations and decisions¶
- Network isolation is intent-level, not OS-enforced: the model-declared
networkflag is checked against the session grant'sterminal.networkscope before a process starts, but nothing at the operating-system layer prevents a spawned process from reaching the network. OS-level enforcement is deferred to the separate sandbox project. - The bounded model selection loop allows at most 128 model steps and 256
sequential tool calls per budget window (
AgentBudgetPolicy); exhausting the window moves the run toWaitingUserand an explicit Continue opens a fresh window. Approvals still gate mutations. The former Worker v1 conversational generation branch and the device/runtime v1 adapters it later carried are both removed; the managed Worker serves/worker/v2only.
Amended: superseded in part by ADR-0012 (the worker process was merged into the host on 2026-09-30).
Amended 2026-10-31: a root run now opens up to eight budget windows by itself before it parks for an explicit Continue; internal child runs keep their bound. See runtime behavior.
Amended 2026-11-08: the window cap is gone. A root run keeps opening windows and ends only when the model finishes, the user stops it or a failure does, so the budget bounds one window's counters and never the length of a run. Internal child runs still keep their own bound. See runtime behavior.
Superseded history¶
The Standard / Full Access picker was a UI affordance that mapped to a reusable rule. Session-scoped agent permissions removed the picker; the chat composer now starts a run with the broker's effective Standard profile and lets the Broker decide each request from the active session grant plus terminal risk classification. Full Access is no longer a user-selectable preset. See the 2026-09-04 implementation plan for the new contract.