One task page for every task: a run's task opens the real task room

[santoshkumarradha]

One task page for every task

In a chat, a task on the run engine opened a page of its own. #1420 restyled it to look like the task room, but it was still a second page: no transcript or work tab, no steer box, no step rows, and no tokens or model. Now every task opens the real task room, on either engine. There is one page type.

What a run's task shows in the room

The room reads the run's task through the session layer:

• Transcript tab: the work order, then the trajectory's steps drawn as the room's own shell calls, with their output. The step in flight is the newest call. A refused action is one dim refused line. A call that did not run for any other reason has no row, but it still counts in the steps. Notes you left are drawn where corrections are drawn. Notes from a worker or the run name no author. The result comes last once the task has ended.
• Work tab: the run's working copy, diffed against the commit it was cut from, plus the files it added. The files codeaf keeps there are left out. A copy that is gone from disk is read off its branch. This uses the new session.PlanTaskWork, built on runCopyOf and the run's notice copy.
• Box: a message becomes a note on the store task (PlanNote). Once the store has it, the room says the worker reads a note at its next step. A refused note shows the store's own sentence on the line you typed. On a task that has ended, the room refuses the note before sending it, in the same words its box and foot already show (this task has finished), and keeps the text in the box.
• x: raises the Stop this task? card and then stops the task through the plan's door. For the run's root, that is the run's stop.
• Head:
  • the store's status, elapsed time and spend;
  • the model and tokens, read from the spend ledger. PlanTaskRow gains Model and Tokens, and the model shown is the one the task spent the most through.
  • Anything the ledger does not record is dropped, never shown as blank or zero.
• Child tasks: the waits and the tasks under it are drawn under the transcript with tui3: a run's rows and its page look like the old task rows #1420's row shape. A press on any of those rows opens that task's room.
• Following: the room follows a running task on its own 3-second beat, never on the paint clock. A task that has ended is read once.

The tasks place, the side list and the run's tab all open this room. The separate plan page is removed, along with its p key and its direct part cancel.

Tabs on every room

No room had tabs before, on either engine, so this PR adds them to every task room (not to the adaptive run's graph page):

• transcript and work are named on the trail row beside the way back.
• tab over an empty box switches between them, and a click on a name does the same.
• An older-engine task's work tab lists the files it changed once it lands.

Apart from the tabs, older-engine tasks are unchanged.

Kept from #1420

• There are still no ◑ $ rows.
• The live step still needs Live.Step > 0 and a command.
• A frame never reads the store.
• Keys typed while a room opens land in its box, unsent.

Tests

• Red first (9c0da30):
  • a new-engine task and an older-engine task open the same room type with the same tabs;
  • a message typed into a new-engine task's room reaches its store as a note.
• internal/tui3/planroom_test.go:
  • steps become calls;
  • no live row without a step or command;
  • the refused line;
  • head model and tokens, known and unknown;
  • the work tab diff;
  • stop through the card;
  • store refusal of a note;
  • keys typed in the gap;
  • follow on the beat;
  • the live step leaving at the end;
  • a part's note reaching the part;
  • the pickup sentence;
  • a part row press.
• Session:
  • the ledger's model and tokens on PlanTaskPage;
  • readPlanWork against a real git copy;
  • a harness file dropped whole from the patch.
• The old page's tests are ported to the room. The tagged e2e scenario now waits for the room's tabs, only logs the note box (an ended task's box names another door), and reopens the tasks place after esc.

Tagged e2e TaskOnTheRunEngine

The room assertions pass. The scenario still fails on this branch, but every remaining failure fails the same way on clean santos/dev2:

• the tasks place fold word → what ran under it never appears;
• the row is already done before running can be read;
• the landing names main instead of the run's branch.

On clean santos/dev2 the scenario also fails at the old page's own steps · $ wait.

Live check (tmux, isolated home, telemetry off, deepseek-v4-flash)

A /task on the run engine, opened from the tasks place:

• The room shows the transcript · work tabs.
• The head reads done 9s · 3 steps with deepseek/deepseek-v4-pro · 22.2k tok · $0.02.
• The transcript has the brief, the steps and the report.
• tab shows the three new files as a diff.
• A note on the ended task is refused in the room's own words and kept in the box.

On clean santos/dev2 the same press opens the old page, which has no tabs, no model and no tokens.

Manual

Updated reading-a-task-page.md (new section on the two tabs), worker-harness.md and task-controls.md.

[santoshkumarradha]

@AbirAbbas ready for review and merge. It's third in line, after #1429 and #1494.

• What it does: every task opens the one task room, on either engine. The separate plan page is gone. The room has transcript and work tabs, a header with state, time, cost, model and tokens, and a steer box that writes a note, and refuses one on a finished task. The work tab and the model also reach the screen through the engine daemon.
• Base: dev was brought in with a real merge commit (5988260df). Since santos/dev2 lands on dev: the worker harness as default, runs as work, skills from other harnesses, the chat as manager #1410 was squash-merged, the branch's own diff was replayed onto dev's tree: the same 48 files and line counts as before. No rebase and no force-push.
• Checks:
  • CI is all green.
  • Full make check on Spark (key unset, see tui3 credit tests read the real OPENROUTER_API_KEY and fail when it is set #1489) had one failure, TestTheVisionGateReadsTheDiskCacheWhileTheCatalogIsWarming, which passed 10 of 10 alone on this branch and on dev (added to Flaky under load: run note tests, session temp-dir races, workspace and provider timing #1441).
  • A live check through the engine daemon showed the work-tab diff and done · 17s · $0.0024 · z-ai/glm-5.3-flash · 42.6k tok.
• Squash message: commit 08a0b55d7 carries a Co-authored-by: Cursor trailer. Please leave it out of the squash message. Cursor attribution is now off.

[AbirAbbas]

Reviewed this and checked it by hand in the real binary (engine road plus one --no-host pass, every seat on deepseek-v4-flash). All three doors open the same room. The transcript, a note reaching the store, the ended-task refusal, x → Stop this task?, head model and tokens, tabs via tab and click, child rooms, and the older engine gaining tabs all hold. Git reads on the work tab stay on the 3 s beat.

I'm pushing three commits on top of 5988260df (plain fast-forward, no rebase):

• 3c478a59b session: read a given-back run copy off its branch in the work tab. Every finished or stopped run's work tab said the run's working copy is not here any more, because an ended run leaves its copy as a plain folder git no longer answers for. readPlanWork returned before the branch read. There's a test that failed first, and it's re-checked by hand: the ended and stopped runs now show their diffs.
• 76f5dd2b6 tui3: the work tab draws a run's difference through drawableLine. File bytes (\r, title escapes, screen clears) reached the frame raw.
• ac47bda31 manual: no plan page left in the run-task passages. tasks.md and worker-harness.md still described the removed plan page and a card-less x on a part.

Should-fixes left for you, not blocking:

• The diff path is taken from the last b/ in the header, so x b/plandb.db is hidden and non-ASCII names show as quoted octal.
• the worker reads a note at its next step is spelled in two constants.
• Older-engine rooms also gained the model on the title row.
• The tagged e2e harness billed claude-fable-5.1 despite --one-model, so not every seat is pinned there.

Merge order is #1429 → this → #1494. Once #1429 lands I'll merge dev into this branch (it conflicts mechanically with #1429 in app.go/input.go/render.go/view.go) and merge after CI.

[AbirAbbas]

Verified by hand in the real binary and on top of dev at 5d8324a (#1429); CI green on 67cd345. Merging.