ACP Integration
An ACP-first client boundary with custom presentation, explicit capabilities, and durable execution semantics.
The Grida Agent — sessions, tools, permissions, providers, hosts, and their client contracts.
View all tagsAn ACP-first client boundary with custom presentation, explicit capabilities, and durable execution semantics.
Grida Agent capabilities, expected behavior, and availability by client and connection.
Configurable lifecycle automation with explicit decision points, scoped execution, and recoverable outcomes.
Glossary and reference for how an agent handles binary attachments the model cannot natively read. Three resolution paths (provider-native multimodal, skill-per-format, shell-based conversion), a format matrix, the scratch-space pattern for archive extraction, and the boundary between protocol and implementor.
How a person's intent is represented and lowered. The structured parts of submitted input, references versus attachments and the routing policy between them, directory references, commands, mentions, editor context, attachment classes, and the rules that turn what a person composes into what the model receives.
Doctrine for running frontier models in an agent loop without paying for waste. The loop's cost model (every token is re-sent every step, at one of several price tiers), the recurring failure classes — missing prompt caching, tool results that echo their input, unbounded and stale tool output, window and compaction decided apart from pricing, rewrite-heavy revision, an unaudited baseline — and the quality-first order for addressing them.
This document decides where an agent host applies confinement. The boundary is
Session-scoped lifecycle facts, their delivery and ordering, and how a host uses them for background attention.
The agent server as a long-lived, discoverable local process — one server, many clients. The discovery contract (registration record, persistent credential, atomic publish, single-daemon convergence), the authenticated probe and protocol gate, connect-or-spawn, the browser exception, and the production shape.
MCP consumption and Grida capability export, with shared tool contracts, bounded discovery, credential handling, and explicit recovery.
How an agent host decides whether an action runs, asks, or is refused — effects declared by tools, a policy of sandbox level, approval mode and reviewer, named presets, defaults, and configuration as a file.
The Plan operating regime as a host-owned property of each accepted input. The regime pair, the transition invariants (the host owns the regime, the agent recommends but never effects a change, a person authorizes execution, the plan is reviewed before Act), capability-layer enforcement, and the narrow plan-artifact carve-out.
The three environments an agent can be hosted in — web (a browser page), cloud sandbox (an ephemeral container or virtual machine), and computer (the user's machine) — and which capabilities, enforcement, and persistence each can honestly offer.
Host-managed working storage for staged inputs, temporary outputs, and recoverable agent work, separate from authored project files.
The design discipline for shaping an agent tool before it is written — the doctrine behind the tool contract. A tool is a contract authored for a consumer that cannot be renegotiated with and cannot be migrated. Minimal surface, host config off the arguments, grounded and honest knobs, auto-resolved inputs, context-frugal and clearly-failing results, and when to reach for a tool versus a connector versus a skill.
Shared invocation semantics and product-owned tool profiles, authority, results, and presentation.
What may start or resume work besides a person submitting input — scheduled wake-ups today, and the requirements any external source (configured schedules, webhooks, programmatic calls, protocol notifications) must meet before it can. Origin provenance, admission through the ordinary queue, trust, headless approval, and lifetime bounds.
The host states which accepted input each turn serves and what became of it; a client renders those facts and never reconstructs them from its own pending view. The authority direction, the identities and transitions a client reads, and how optimistic presentation reconciles.
The read/view modality split. Why a text read and a visual view are separate tools, the perception-tool contract, the input matrix (bitmaps now, rendered sources later), how a tool result becomes a provider image block, and the retention policy that keeps perceived pixels from re-filling context every turn.
A resolved workspace binds named directory identities and a primary directory to agent work while keeping membership, authority, and scratch separate.