── messages ────────────────────────────────────────────────────────────── AppendMsg appends text to the scrollback. Sent by the stdout-pipe forwarder

Agent owns one conversation: a provider, a tool registry, a compaction strategy, and a slice of messages. The same struct serves the root REPL and any

AgentRunner is what the program calls when the user submits a line. It covers both slash commands and normal agent turns — the harness wires it up in

AnthropicProvider is the reference Provider implementation. It is the only file in the harness that imports the Anthropic SDK.

ApprovalRequest is sent (via program.Send) when the agent needs y/n from the user. The program shows the prompt and writes the answer to Reply. Detai

Block is one piece of message content. Fields are interpreted based on Type.

Client wraps an MCP session and the transport it was opened on. One Client corresponds to one connected server; Close terminates it.

Config is the top-level JSON shape.

DebugRefreshMsg tells the TUI to re-read the debug ring and redraw the panel. Sent from debug.SetSink whenever a new event is recorded.

DebugToggleMsg tells the TUI that debug.Enabled() may have changed, so it should recompute viewport heights. Sent by the /debug slash command.

DelegateTool exposes one Subagent to the model as a callable tool. The model sees "delegate_<subagent name>" with the subagent's description and a sin

Entry is one piece of memory. Kind tags the role of the content so implementations can route or group it — e.g., SessionFiles uses KindSessionSummary

Event is one entry in the debug log. Message is the one-line summary shown in the panel; Payload is the optional long-form content (request JSON, full

Level classifies an event's severity. Used by the UI to color-code lines.

LoggingStrategy wraps another strategy and appends each non-trivial compaction (before / after) to a file. Use it to inspect strategies and compare th

MCPStatusMsg reports progress while MCP servers are being connected in the background at startup. The TUI shows a "loading…" line above the input box

MCPTool wraps one remote tool so the agent's local Registry can dispatch to it. The agent loop sees a Tool like any other.

MockProvider is a Provider implementation that returns canned responses. It's the recommended way to test the agent loop, compaction strategies, and a

NoCompaction is the default — never modifies messages.

NoMemory is the do-nothing fallback. Every method succeeds with empty output. Used as the initial value of Default and as the test seam for "what if m

OpenAIProvider implements Provider on OpenAI's Chat Completions API. Works for any chat-capable model: gpt-4o, gpt-5, gpt-5-codex, o3-mini, etc. This

ProgressFunc reports background progress while Register is iterating the server list. server is the current server's name (empty on Begin/Done); total

ProgressStatus is one phase of MCP server bring-up. A ProgressFunc gets called once per phase per server, plus ProgressBegin/ProgressDone bracketing t

RecallTool surfaces past memory entries (typically prior session summaries) that match a query. Match semantics depend on the Store — SessionFiles (th

RememberTool writes a piece of memory through the active memory.Default store. Persistence semantics depend on the Store implementation — see follow_a

Research is a focused subagent for read-only investigation. It gets the file-reading tools but not bash, and runs with a tight system prompt so it sta

ServerConfig describes one MCP server to connect to. Either Command (with optional Args) for stdio, or URL (with optional Headers) for HTTP. The loade

SessionFiles is the default Store: one markdown file per session under <root>/sessions/, with <root>/index.json as a fast-lookup layer over them. Both

SessionRecord is the metadata stored in index.json — one entry per session file. Path is relative to the store's root directory.

SlidingWindow keeps the last KeepLast messages and drops everything older. It snaps to a safe boundary so it never separates a tool_use from its tool_

Store is the persistence layer. Three operations cover the natural rhythm of agent memory: write (Save), search (Recall), and the always-loaded prefac

Subagent is one focused agent type. Run takes a task description and returns the final text — same shape as a tool but with its own context window und

Summarize calls the provider to summarize old turns once the history exceeds Threshold, replacing them with a single synthetic user message.

Usage reports the token accounting for one API call. Providers fill it in; the harness accumulates totals for the session.

UsageFunc returns the session's cumulative token usage and estimated cost in USD. Cost is -1 if unknown (e.g. unrecognized model). The TUI renders thi

internal messages

chapter is what the templates see for a single rendered chapter.

chapterData is everything the chapter template needs.

── main message input ────────────────────────────────────────────────────

── confirm prompt ────────────────────────────────────────────────────────

The agent loop prints `[tool] …` lines and assistant text to stdout by design (chapter 01). Tests don't silence them — running `go test -v` will inter

exercise is an extension prompt: a self-contained task that builds on the harness. Has a numeric slug prefix (01-, 02-, …) and a difficulty tag pulled

exerciseData is everything the exercise template needs.

fakeTool is a Tool used by tests that need a registered tool but don't care about its behavior — Execute just returns a canned string.

group is a section in the landing-page curriculum.

guide is a how-to recipe: shorter than a chapter, no curriculum number, linked from its own section on the landing page.

guideData is everything the guide template needs. Guides don't get prev/next nav (there are only a handful) — just a back-to-index link.

headerRoundTripper injects a fixed set of headers into every request. Used for auth tokens on HTTP MCP transports.

landingData is everything the index template needs.

── model ─────────────────────────────────────────────────────────────────

pricing is per-million-token rates in USD. Update from https://www.anthropic.com/pricing when rates change.

shineTickMsg drives the startup banner shine animation. Self-perpetuates until the shine peak passes off the right edge of the banner.

tokenReporter is what cmdTokens type-asserts to. We didn't widen the Provider interface because not every backend has a "tokens" concept; any provider

Active tracks subagents currently running. Used by the UI to show "what's executing right now."