Agent-libOS

Register AgentImage packages, publish checkpoint-derived images, or replace the current process image. Use when creating reusable runtime configurations or deliberately changing process code, tools, prompt identity, memory, or authority.

Diagnose this process's Capability authority and request one exact Human policy decision when a protected operation lacks requestable authority. Do not use it for intent questions, validation errors, provider failures, or speculative privilege expansion.

Delegate one attenuated Capability to a known direct child or revoke one exact Capability. Do not create children, grant peers or grandchildren, amplify authority, or treat revocation as rollback of completed effects.

Capture, inspect, compare, restore, or fork durable process-subtree checkpoints. Use for recoverable internal-state milestones, isolated replay, or deliberate rollback of reconstructable Agent libOS state.

Spawn or fork direct AgentProcess children, coordinate their lifecycle and messages, then import selected results. Use for bounded parallel delegation, isolated exploration, or structured parent-child collaboration.

Run one policy-governed, non-interactive host command through the argv-only Shell boundary. Use for builds, tests, searches, or utilities lacking a narrower tool; not for general Git, TTY/streaming work, ad hoc remotes, or attempts to evade policy.

Manage validated local branches, tags, checkout targets, and structurally trusted Runtime-managed worktrees without accepting arbitrary host paths.

Stage, unstage, and commit reviewed changes in one managed Git worktree. Use only when the entire index is intended for the commit; stop if unrelated staged entries must remain staged.

Inspect the fixed Runtime Git repository, managed worktrees, status, diffs, history, refs, remotes, and blame before making or verifying a Git decision.

Restore selected paths, integrate or abort one history operation, manage stashes, reset state, or clean untracked content after exact Git inspection.

Create immutable, byte-safe Git patch Objects and apply them to an authorized worktree in the same repository. Use for bounded patch transfer, reviewable change handoff, or cross-process application through Object Memory; do not use rendered diffs as patch payloads.

Create, list, inspect, review, merge, or close repository-local simulated pull requests backed by Git snapshot refs and metadata. Use for Agent libOS's evidence-preserving local PR workflow; do not use it for GitHub, GitLab, or any external forge.

Fetch, pull, or push through an existing Host-configured Git remote. Use for remote synchronization when the exact local state, remote name, branch or full destination ref, integration strategy, and guarded destructive intent are known; never supply ad hoc URLs or credentials.

Ask one blocking Human intent/data question or send one one-way update/final report through the configured Human provider. Do not use conversation as Capability approval, trusted verification, event polling, or arbitrary transport.

Propose, validate, and register a process-local Deno/TypeScript tool when a reusable bounded operation is missing from the visible Agent libOS tools. Use for narrow computation or composite logic with an explicit JSON contract; never use JIT authoring to bypass Capability, approval, provider, filesystem, shell, or network controls.

Discover, inspect, and call a Host-registered plain JSON-RPC-over-HTTP integration by governed logical endpoint and method IDs. Use when the intended integration is JSON-RPC rather than MCP; never invent URLs, wire methods, headers, credentials, or transport commands.

Discover, inspect, refresh, and call Host-registered MCP tools through governed logical server and tool IDs. Use for MCP Tools over registered stdio or Streamable HTTP transports; never invent servers, URLs, commands, raw MCP names, credentials, or treat live metadata as authority.

Import governed workspace text into Object Memory or export Object text to a file. Use for content boundary transfer, not inspection, editing, binary data, or byte-exact copies.

Create, discover, read, and append governed runtime-local Object Memory state. Use for named JSON that must survive prompt or compaction changes within one live Runtime; use files or captured checkpoints for reopen durability.

Run one visible tool asynchronously in a Host-managed child bound to an owner Object, then inspect, wait, watch, or cancel it. Use for single background calls; use child-processes for multi-step Agent work.

Read wall-clock time, make one bounded timer delay, compact an enabled LLM-context Object, or finish the current AgentProcess. Never use timer sleep as an event wait or treat runtime-local Object payloads as reopen-durable storage.

Discover an applicable Agent Skill, activate its exact snapshot, read declared resources from that loaded snapshot, and unload it when its guidance or tools are no longer needed. Use when the required Skill is not already loaded or when a registered Skill must be found; never use Skill lifecycle operations to obtain Capability authority or bypass a denied tool.

Extract heuristic FAILED, error, and AssertionError lines from supplied pytest text. Use to triage an already captured bounded log; it does not read files, retain traceback context, run tests, return line indexes, or prove pass/fail.

Round-trip a small top-level JSON object through echo to isolate model/tool argument and result plumbing. Use for semantic transport diagnostics; not for domain validation, authorization, provider health, JIT registration, arbitrary persistence, or application work.

Create, replace, or delete workspace text files and directories with explicit mutation and verification semantics. Use after resolving cwd, target, and baseline; not for binary/Object transfer, Git metadata, range edits, or shell commands.

Inspect cwd-relative workspace directories and bounded text, or select this AgentProcess working directory. Use to establish paths and textual evidence before acting; not for writes, binary or range reads, Object transfer, or Git metadata under .git.

SWE-Agent inspired coding workflow for fixing, reviewing, and improving software repositories through a compact agent-computer interface.

Register a benchmark JIT tool that attempts a filesystem syscall.

Expose read_text_file as a benchmark visibility-only skill.