internal/

Package acp is the agent side of the Agent Client Protocol (https://agentclientprotocol.com), protocol version 1: newline-delimited JSON-RPC 2.0 over stdio.

Package agent runs the step loop: query the model, execute tool calls, repeat.

Package attempts runs one task up to N times (best-of-N), each attempt as a child tinyagent run in its own Git worktree, grades every finished attempt with user-configured check commands, and picks a verified winner: the lowest-numbered attempt whose run ended Submitted or Completed and whose checks all exited 0.

Package config loads agent settings from an optional YAML file layered over defaults.

Package eval runs a suite of coding tasks against tinyagent and reports how often each model/profile configuration passes the task's own checks.

Package eventlog fans run events out to sinks: the --json JSONL stream and the opt-in SQLite event log.

Package gitwt holds the Git plumbing for worker subagents: snapshotting a working tree into a commit without touching the index, creating a linked worktree on its own branch, committing a worker's changes, and applying them back to the parent's working tree atomically.

Package harness runs a workflow of minimal markdown skills: a router picks the skills for a task, each skill is one narrow prompt with a success check, and a loop runs them in order, feeding each stage's output to the next and stopping at the first stage that fails its check after retries.

Package hooks runs user-configured commands before and after tool calls, and stop hooks that verify a run before it ends successfully.

Package lockdown holds the run-level taint state of the opt-in untrusted-content lockdown (untrusted_content.enabled / --lockdown, issue #171).

Package lsp is a small Language Server Protocol client: it starts the configured servers lazily over stdio, keeps the files it has shown them in sync with disk, and collects diagnostics, definitions and references.

Package mcptools connects explicitly configured local MCP servers.

Package memory is tinyagent's cross-run project memory store: short, redacted notes that a run saves about a project and later runs load into their system prompt.

Package model talks to LLM providers using a single OpenAI-shaped message format.

Package outschema loads a user-supplied JSON Schema for --output-schema and validates a run's structured result against it.

Package permissions implements the opt-in tool permission policy: a set of allow/ask/deny rules matched on a tool call's name and one primary argument (a shell command, a path, or a URL host).

Package plugins loads Agent Plugins 1.0.0 packages (https://agent-plugins.org).

Package sandboxnet implements the sandbox network allowlist: entry parsing and matching, and the filtering HTTP proxy that sandboxed commands use.

Package skills discovers Agent Skills (folders holding a SKILL.md) in the project's and the user's .agents/skills directories, renders the catalog added to the system prompt, and loads a skill's instructions or one of its bundled files for the skill tool.

Package snapshot captures and restores the files under a working directory for opt-in workspace snapshots (--snapshots) and --rewind.

Package telemetry exports runs as OpenTelemetry traces over OTLP/HTTP.

Package tinygo manages the TinyGo compiler TinyAgent uses to build Go plugin source into WebAssembly.

Package tools implements the agent's file and shell tools, sandboxed to a root directory.

Package trajectory records an agent run as JSON.

Package transcript renders a saved trajectory as a Markdown transcript for people: the task, each step's assistant text, tool calls and results, the run's plan, compactions, model switches and resume boundaries, and the outcome.

Package update replaces the running binary with the latest GitHub release.