BudgetStatus is the client's latest snapshot of server accounting. Set is false until the first cloud response, so zero values never render. Remaining

Config is the on-disk schema at .codehamr/config.yaml. Strict decoding: unknown top-level keys fail Bootstrap so typos and stale schemas surface immed

ErrUnreachable wraps transport errors (refused, timeout, DNS miss) so the TUI can render a hint instead of a raw wrap.

Event is what the TUI consumes. One event per stream update.

PackResult records what Pack kept: the packed messages.

ProbeResult holds what Probe extracts from a one-shot hello request: the live context window and budget snapshot. The TUI shows the real window at act

Profile is one named model endpoint in config.yaml; `/models` switches between them. ContextSize is omitempty so server-managed cloud profiles omit it

argOption is one popover entry, used at command-level (one row per command) and argument-level (one row per accepted value for the active command).

chipContent is the payload behind a chip, kept in store by id so the visible label can collapse while the real text survives for Value() and replay.

chipSpan marks the rune range [start, end) in the value holding a chip's label. Reconciled after any key event that can shift the value.

command is one row in the popover, --help, and the dispatch table. args, if non-nil, supplies the argument-level popover entries.

fakeRelease serves a goreleaser-style manifest plus one binary asset; tests plug its URL into the package vars via withReleaseURLs.

goreleaserConfig is the slice of .goreleaser.yaml we assert against: the goos x goarch cross-product goreleaser builds a binary for.

headTailBuffer is an io.Writer keeping the first bashOutputHead and the last bashOutputTail bytes written, discarding the middle. The tail is a fixed

phase is the turn state machine and single source of truth: idle = no turn; thinking = awaiting the model, no tokens yet; streaming = content flowing;

pingMsg carries a backend-reachability result. baseURL is the URL probed; Update drops the message when it no longer matches the live client's URL, el

platformKey is a comparable os/arch pair for set comparison.

probeMsg carries the outcome of an activation-time Probe (hello-world chat): validates URL+model+key in one round trip and harvests the live context w

promptEntry is a frozen promptInput snapshot for history replay: displayed text plus the chip metadata needed to restore atomic-chip behaviour on ↑/↓.

promptInput wraps bubbles/textarea with an atomic-chip model. A large paste collapses into a single inline label [Pasted text +N lines] that acts as o

queuedPrompt is a prompt the user committed while a turn was running, held until the turn ends and then auto-submitted. send is the chip-expanded text

quitArmResetMsg fires ~3s after Ctrl+C arms the quit: if not already quit or re-armed, clear the hint from the status bar.

streamChunk is one OpenAI SSE frame. finish_reason is deliberately not decoded: readSSE dispatches accumulated tool calls at stream end, not on finish

streamEventMsg and streamClosedMsg tag their originating channel so the model can drop events from a stream the current turn no longer owns. After Ctr

streamOptions: without include_usage, OpenAI-compatible servers omit the usage block in the SSE tail chunk and the per-turn token counter sits at 0.

toolResultMsg carries one finished tool call back to Update, tagged with the turnCtx it was dispatched against. Update drops it when that ctx no longe

toolSlot accumulates one streamed tool call. OpenAI delivers `arguments` as JSON fragmented across chunks, each fragment invalid alone; we append raw

turnOutcome is how a finished turn ended, frozen into the status bar until the next submit. outcomeNone is the zero value (no turn has finished yet, o

wireMessage is the outbound OpenAI request shape; responses parse via streamChunk. Content has no omitempty: silent bash commands (e.g. heredoc write