Source analysis summary

This page summarizes runtime pressure from implementation families. It is not a port of any single project; it extracts portable semantics for the Agent Runtime contract.

Terminal coding runtime pressure

A terminal runtime shows that tool calls are not simple request/response records. The standard needs tool schema, safe args, progress, partial and final results, concurrent read-only batches, serialized writes, hooks, permission modes, rule sources, classifiers, sandbox profiles, output spill, telemetry-safe errors, local and remote subagents, and compaction boundaries.

It also shows that task state cannot be only a model-facing checklist. Foreground tasks, local shell tasks, remote agent tasks, in-process teammate tasks, scheduled tasks, backgrounded work, and generated output files all need task ids, output offsets, terminal-state eviction, stop controls, and notifications.

Scheduler and gateway runtime pressure

Gateway and scheduler runtimes show that background work needs durable storage, secure output refs, origin/delivery context, per-job model/tool overrides, pre-run scripts, prompt-injection gates, inactivity timeouts, delivery failure state, run outputs, and checkpoint/resume scanning. A missed scheduled task or failed delivery is a task fact, not just a log line.

Typed systems runtime pressure

A typed systems runtime shows that SDKs need stable item lifecycle: thread started, turn started, item started/updated/completed, turn completed, turn failed. Command execution, file changes, tool calls, web search, image generation, todo lists, reasoning, approval policy, sandbox policy, hook schemas, process lifecycle, durable state, goals, jobs, remote control, and rollout reconstruction all need explicit facts.

It also shows why task and run must be separate. Thread goals, plan items, todo lists, job tables, job item assignments, spawn edges, turn status, approval wait guards, and thread read models each cover part of the work. The standard needs task ids, run ids, attempt history, dependency edges, worker assignment, and lost-state handling to join them.

Desktop agent runtime pressure

A desktop runtime shows that product execution is not a single model request. Submit, queue, resume, interrupt, compact, read models, tool inventory, evidence, replay, review export, turn input snapshots, request telemetry correlation, task profile, candidate sets, routing decisions, cost state, limit state, and quota/rate-limit events need one fact chain.

It also shows that foreground turns, subagent turns, automation jobs, execution summaries, task files, artifacts, and timeline UI projections should converge on the same runtime facts. Scheduler ticks and UI cards should not become separate task authorities.

Gap matrix

Gap	Standard surface
Approval exists only as a client dialog	permission.* events and action boundary.
Sandbox exists only as config	sandbox_profile and sandbox.* events.
Hooks are not portable	hook.* events and hook input/output contracts.
Commands are just tool text	process.* lifecycle and output refs.
Model choice is not explainable	task_profile, candidate_model_set, routing_decision.
Cost and limits live only in logs	cost.*, rate_limit.*, quota.*, read models.
Tasks are only todos or UI cards	task.* lifecycle, attempts, graph edges, progress, delivery state.
Subagents are just messages	Durable parent-child graph and job/item model.
Remote execution cannot recover	Channel identity, resume cursor, permission bridge.
Compaction breaks audit	Context boundaries, rollback, reconstruction.
Large output pollutes streams	Output refs, spill, truncation, redaction.