Prototype the Run Workbench as an agent screen

[Sandstorm831]

Question

What should the Run Workbench look like as a mirror of the agent, the way OpenCode's session screen is, rather than a literal event timeline? Build a rough rendering to react to that settles the layout; what is shown, collapsed, and hidden; truncation and how truncated content is reached; colour and emphasis; the working indicator; where Workflow progress such as the current Step sits; and how Human Gates, Harness Requests, agent questions, and the human's own Turns stay distinct. Settle whether the rule that durable, live, and preview content stay visibly distinct holds or is amended.

[Sandstorm831]

Starting context

Input	Why it matters	Authority
Report 5: Run workbench	What the screen must fix: a literal timeline, truncated questions, timestamps and ids, no emphasis, lingering sending state, no working cue, a replaced preview	Canonical (source report)
Charting record	The everyday-versus-expert rule, and the Workbench fixes already going to the first milestone, so the prototype builds on them	Canonical
Study how OpenCode's session screen presents an agent conversation	The default answer for layout, truncation, emphasis, the working indicator, prompt keys, and questions replacing the prompt	Research
Study how T3 Code turns Claude Code and Codex events into a conversation	How Harness events become rows, what counts as noise, and the seam gaps each visible item exposes	Research
Prototype Crucible's launch and Run information architecture	The timeline-first decision this supersedes, and the rule to keep: Human Gates, Harness Requests, agent questions, and the human's Turns stay distinct	Canonical decision
Define Crucible's TUI-facing command and projection interface and ADR 0024	The durable, live, and preview update split, and the rule that they stay visibly distinct	Canonical
ADR 0022	What may cross the Harness seam; private reasoning stays private	Canonical
docs/agents/tui-workbench.md, src/tui/AGENTS.md, src/tui/run-workbench.tsx, src/tui/run-timeline-rows.ts	Current key routing, modal stack, and rendering the prototype replaces or keeps	Canonical (current behaviour)
docs/agents/prototypes.md and slicing rule 7	How a prototype is isolated and thrown away, and the TUI checks every later slice must copy	Canonical

Required methods

• /prototype
• /grilling
• /domain-modeling

Reading rule

Read report 5, then both research Answers. Decide what the screen shows before what the Harness seam must carry, and keep the prototype throwaway.
Record every seam or Projection gap the chosen design needs; that list lets the map's "What the agent screen needs beneath the TUI" patch graduate
into tickets. Settle, or spin out as its own ticket, two recorded rules: whether ADR 0024's visibly-distinct rule holds as written, and whether a
provider-supplied reasoning summary counts as private reasoning under ADR 0022.

[Sandstorm831]

Starting context addendum

Decide how the human interrupts a Turn and sends a message while a Turn is working (ADR 0035) leaves presentation to this prototype:

• the Interrupt and Steer keys in both agent Step kinds (Steer is now offered during Interactive agent steps, and on Claude Code);
• the Run waiting on the human after an Interrupt;
• delivered and dropped Steers shown as the human's messages inside the Turn;
• a dropped Steer's text restored to the compose;
• a Steer rejected after the Turn ended, with its draft kept.

[Sandstorm831]

Resolution

The human chose layout B from three variants run in a real terminal. The Run Workbench is an agent screen: a transcript that only appends, a single bottom region, and OpenCode's working indicator, with Workflow progress in a sidebar. Recorded as ADR 0036 (PR #265), which supersedes Prototype Crucible's launch and Run information architecture's timeline-first Run view and keeps its stay-distinct rule.

Decided

1. Layout. No header. The transcript fills the screen. Above 120 columns a 42-column sidebar shows the Bundle, the Steps (done, current with its Iteration, still to come), the Harness, model, and effort with any pending Model choice, and the context used; below 120 columns that moves into the prompt's meta row. Earlier Steps stay in full behind thin Step rules, with no folding. No ids, processes, or timestamps; the Run id appears once the Run leaves an active state.
2. Shown, collapsed, hidden. The human's messages sit in agent-colour panels, and Steers are the human's messages inside the Turn, marked waiting, read, or not delivered. An Entry Turn's Bundle prompt is one muted "Secant started the Step with the … prompt" line. Assistant text streams as markdown. A reasoning summary is a collapsed Thought: <title> · <duration> row. Each tool call is one row that changes in place; shell output and diffs are panels. Agent calls show their reason and outcome. A settled Turn ends with ▣ Harness · model · duration (· interrupted when it was). Rate limits, token usage, account notices, and unknown methods never reach the transcript.
3. Truncation. Only shell output collapses, at 10 lines. Assistant text, questions, the human's messages, and diffs are never cut. ctrl+o or a click expands.
4. Colour. Vendored theme roles only; everforest is the default theme. The agent colour marks the human's messages, the prompt bar, the Turn line, and the scanner; muted marks settled work; warning marks reasoning and Harness Requests; error marks failures; success marks Agent calls and a finished Run; markdown carries bold.
5. Working indicator. OpenCode's block scanner and esc interrupt appear only while a Turn works. The prompt placeholder says whose move it is ("Message the agent — it reads this at its next step" or "Reply to the agent"). Sending clears the draft at once.
6. Distinct interactions (rule kept). Each replaces the prompt and names its kind in its header: "Permission required" (Harness Request, warning), "Workflow decision" (Human Gate), "Review checkpoint". Agent questions are plain, untruncated transcript text, and the human's Turns are the human's messages.
7. Interrupt and Steer (ADR 0035's presentation). After an Interrupt the Turn line reads · interrupted, a note says the agent is waiting on the human, and undelivered Steers are back in the draft. A Steer that arrives after its Turn ended gets a notice and keeps its draft.
8. Keys and mouse. No bare letter is a Workbench command, since the prompt is always editable. esc esc interrupts, enter sends or Steers, shift+enter or ctrl+j adds a newline, ctrl+e ends the Step, ctrl+n continues a Repeat, ctrl+o expands, ctrl+g opens the details panel (where resume, cancel, and delete keep their letter keys and confirmations), and ctrl+c clears the draft or quits when it is empty. The mouse is on (wheel scroll, click to expand), and every mouse action also has a key.
9. Transcript overlay retired. Captured command output and Run Artifacts are inspected from the details panel.
10. Model choice in the Workbench. The sidebar (or meta row) shows Harness · model effort with a pending note, such as → medium applies when you send for Codex; Claude Code shows the change applied once observed. The /model and /effort compose menus go with the typed-commands fog.

Recorded rules settled

• ADR 0024's visibly-distinct rule is amended. Durable, live, and preview updates stay distinct as data at the Projection Port. On screen, only the working indicator and spinners mark live work.
• ADR 0022: a provider-written reasoning summary is not private reasoning. It may cross the Seam and is shown collapsed. The raw chain of thought stays private. Reasoning summary joins the Run lifecycle glossary.

Seam and Projection gaps

These graduate into tickets:

• Establish whether Claude Code's direct stream-json transport emits reasoning summaries, and under which setting
• Decide the Harness Seam's tool, command, diff, and reasoning-summary events: call identity that pairs start and end; running, failed, and declined states with error text; typed rows; command output and exit code; diffs as data; a reasoning-summary event; noise kept out of activity; context from both Adapters.
• Decide how a Turn's history grows mid-Turn in the Run Projection and headless --json: message identity, durable rows mid-Turn, Steer state, Turn duration and the interrupted mark, Entry Turn authorship, Agent-call rows, and headless parity.

Assets

• Prototype, the primary source, on prototype/issue-247-run-workbench-agent-screen at 60d1b31. Run it with bun run prototype:workbench. It is throwaway and never merges.
• ADR 0036, the ADR 0024 amendment, the ADR 0022 note, and the glossary entry, all in PR docs(adr): record ADR 0036, the Run Workbench mirrors the agent (#247) #265.