Module update 

update — bare-naked streaming handler.

Refreshes every shipped binary from the latest GitHub release by downloading a single per-platform zip (objectiveai-<version>-<os>-<arch>.zip) and replacing the contents of the machine-wide bin/ directory wholesale. Emits one ResponseItem per stage as the run progresses.

Layout on disk (resolved via ctx.filesystem.bin_dir() — every binary is machine-wide, shared across states):

<bin_dir>/objectiveai{.exe}                       ← cli
<bin_dir>/objectiveai-api{.exe}
<bin_dir>/objectiveai-viewer{.exe}
<bin_dir>/objectiveai-mcp{.exe}
<bin_dir>/objectiveai-db{.exe}
<bin_dir>/objectiveai-claude-agent-sdk-runner{.exe}
<bin_dir>/objectiveai-codex-sdk-runner{.exe}
<bin_dir>/objectiveai-mcp-laboratory   (always musl-linux; no .exe)

Flow:

1. Resolve the single zip asset for this (os, arch). A missing zip emits ResponseSkipReason::IncompleteRelease.
2. Download the zip to a temp path (outside bin/, so the wipe below can’t clobber it).
3. Kill the running servers in-process: the machine-wide api (its lock lives at <bin_dir>/locks) and the per-state db / viewer across every state.
4. Rename the running updater aside (Windows can’t overwrite a running .exe; renaming frees the name — the process keeps running from the renamed file).
5. Wipe bin/ keeping only plugins/, tools/, and config.json (best-effort — a still-running server/.old that won’t delete is skipped). pg-bin/ is wiped; postgres re-extracts on next db spawn.
6. Unzip the download into bin/. Each binary is written via the same rename-aside swap so a still-locked straggler doesn’t fail the extraction.

Caveat: db/viewer are killed across every state, but a server that comes back (or a binary still locked at unzip time) is left as a .old; the cli’s own .old is swept on the next invocation.

Modules

request_schema

response_schema

Functions

execute