Repository navigation
Description
ADR-0018 runs the Harness headless over ACP and has ai-buddy draw the Chat surface. The open chat issues — #474 (which mind answers), #476 (history the session no longer has), #483 (no thinking shown), #459 (connect flow), #437's held permission prompt, #348 (switcher) — are each a piece of the Harness's own TUI being rebuilt in src/chat.js. The TUI solves all of them today.
The load-bearing reason
ADR-0018 rests on ADR-0008: one Harness session is Director and chat, and a TUI is a second process, so on "four of five Harnesses" a second conversation. The other reasons in ADR-0010 are weaker: the turn-finished signal is needed by the Director, not by chat; "terminal log" is already an approved Chat UI (#339); the zero-dependency stance was stated as a cost, not a principle.
The Director keeps ACP regardless — it needs a structured prompt in and one parseable line out. So the question is narrower than the ADR frames it: must chat share the Director's session, and can a TUI attach to it?
Probes
- Resume the ACP session from the CLI.
harness-session.jsonrecords the id the Claude adapter returned. Doesclaude --resume <that id>from the same working directory open that conversation? ADR-0010 asserted only opencode can attach; it never probed Claude.hermes acpadvertisessessionCapabilities: fork, list, resume(ADR-0017) and nothing reads it. - PTY spike.
portable-pty+ xterm.js in a Tauri window,claudeat the data folder with our MCP config, wrapped in a header/footer carrying the status line ADR-0010 wanted. Judged by Summon → ask →speaklands on screen.
What decides
- Probe 1 passes → supersede ADR-0018: ACP for the Director, the Harness's TUI in a PTY for chat, attached to the same session where the Harness allows it. feat(chat): Say which mind is answering, Harness or HTTP endpoint, in the Chat window #474, The Chat window shows history the session behind it no longer has #476, The Chat window never shows a Harness's thinking, and nothing decided that it should not #483, Give the user a chat UI switcher over the three designs #348 and the fix(harness): Loose ends from the first ACP client #437 forwarding become obsolete.
- Probe 1 fails → the amendment is two sessions sharing Memory plus one
recent_actionsMCP tool, and ADR-0008 weakens to one Harness, one account, one Memory. Still a net deletion. - Either way the costs to weigh: a second agent process per Summon, xterm.js selection and IME, ConPTY on Windows (Windows Shell furniture for tray, settings, and opener #197).
— Claude, in a session with @omesser
Metadata
Metadata
Assignees
Labels
Projects
- StatusShow more project fieldsIce Box
Activity
omesser commentedon Sep 7, 2026
Probe 1: resume works, from any cwd
harness-session.jsonholds8e3d40f9-…. That exact file exists at~/.claude/projects/-Users-oded-Library-Application-Support-ai-buddy/8e3d40f9-….jsonl, and every record in it carries thatsessionId: the id the ACP adapter returns is the Claude Code session id, with no aliasing. The file holds the Director's turns (55 user, 19 assistant; the last prompt endsstanding on: the top of the Dock).The same command from an empty scratch directory gave the same answer and appended to the same file, so cwd only stamps the turn; it is not needed to find the session.
So ADR-0010's "only opencode can attach" does not hold for the default Harness. The TUI can open the Director's own conversation. What this did not test: both processes attached at once. The ACP child was not running during the probe. Claude Code has no visible lock on the transcript, so a PTY-hosted TUI and the ACP Director appending to one jsonl is last-writer-wins until proven otherwise, and the spike has to check whether a resumed TUI sees turns the ACP child appends after the TUI loaded.
Side finding: the last two assistant records before the probe are
API Error: 400 Claude Code 2.1.185 does not support this model; version 2.1.251 or newer is required.@agentclientprotocol/claude-agent-acpbundles SDK 2.1.185, so the Director's turns were failing against the configured model. Separate bug.— Claude, in a session with @omesser