Skip to content

The Bot as orchestrator: Stop and follow-ups, agents as a list, Worked for and task details, finished tasks off the board - #1710

Open
davidmckayv wants to merge 38 commits into
mainfrom
orch/fused
Open

davidmckayv wants to merge 38 commits into
mainfrom
orch/fused

Conversation

@davidmckayv

@davidmckayv davidmckayv commented Nov 16, 2026 •

Copy link
Copy Markdown
Contributor

What

The Bot as orchestrator, slices 3, 4 and 5 and the board change, fused on main as one PR. Every branch's commits are kept as they were built; the spec is bundled and the clients regenerated once over the fused spec.

Branch What it adds
fixpass/done-tasks-leave-board A finished delegated task leaves the Needs you board and is listed under Done on the Bot's Activity page; the Bot page's Activity row counts as that page lists; a narrow Needs you row stacks and folds its least actions into its menu.
orch/stop-on-card stopTask in the API; Stop on the task card, its Activity row and the Needs you list; the Bot's follow_up and stop drawn as the task's card; delegating guidance for both.
s5/worked-for listTaskTurns in the API; a task turn's steps fold into one "Worked for 52s" row; a Task details pane beside the thread (Ran by, Outputs, Sources, Computer).
slice4/bot-agents Agents in the API (api/openapi/agents.yaml: an organization's agents, each Bot's delegation, callback tokens) and an optional agent on Task.
slice4/bot-agents-screens The Agents page, a Bot's Agents page with the Hand work to agents switch, no model picker on a Bot that hands work out, the agent named on a task's card and in its thread.

Chosen in the build

  • A finished delegated task leaves the Needs you board (decided 2026-11-16); its result is already in the conversation that asked, and it is listed under Done.
  • Slice 3: a follow-up sent while a task's turn runs waits for that turn to end, as a typed steer does. The Bot's stop passes the person's own words for the task, and only a task those words clearly fit is stopped; otherwise nothing is stopped and the open tasks are listed so the Bot asks.
  • Slice 4: admins add an organization's agents, anyone may add a private one for their own Bots; a Bot switches between working on its own agent (every existing Bot) and handing work to agents; an orchestrating Bot holds no work tools on the person's turn but still reads their memory; runs nobody waits on go to the Bot's default agent; the model picker is hidden on an orchestrating Bot; the computer stays a Bot setting; no A2A agents in this slice; deleting an agent is refused while it is any Bot's only agent; callback tokens per agent; agents are public or private like Bots, and another person's private agent is named "Private agent". API path bot-agents. On one person's OpenBot, added agents are public so its Bot can use them.
  • Slice 5: "Worked for" is the sum of the turn's run time, never the wait on the person; model, agent, sources and outputs show in a Details pane; no separate tasks side panel, since the Bot's Activity page is the one list.

Fusing

  • The narrow Needs you row: main already stacked a row's actions under its words; this takes the board change's container query and folding menu over it, and keeps main's whole question line on a waiting task (readDetail) and Mark done only where it does something.
  • The task card test file holds both Stop and the agent line's tests.
  • On main with 1709: the board row keeps main's Couldn't finish wording and Try again only where it can change the outcome.

Found on the fused build and fixed here

  • A task's waiting approval folded into its Worked for row, so the thread showed nothing to answer (the approvals journeys). A run with a step held for the person's approval, or unanswered once the turn is over, stays where it is, card and all; the fold reads the approval inbox the conversation already keeps.
  • Agents in OpenDots' rail squeezed its buttons in a short window and the Bots badge rode over the button above (opendots-rail-badge.test.tsx). The buttons keep their size and the rail scrolls.
  • follow_up and stop had no words for a task's line where no card is drawn.
  • The Task details pane said the Bot ran a task its agent did; it now reads Handed to under Ran by.

Tests

task-worked-for.test.tsx holds the approval fold (red before the fix), task-details.test.tsx the agent under Ran by (red before). New in the slices: task-thread.test.tsx (Stop on the card through stopTask; the agent named on the card and the brief), task-worked-for.test.tsx, task-details.test.tsx, agents-page.test.tsx, needs-you.test.tsx and bot-activity-groups.test.tsx (finished tasks off the board, under Done), needs-you-board-layout.test.tsx (real browser at 390, 650 and 768), beside main's board-row-layout.test.tsx.

davidmckayv and others added 23 commits November 16, 2026 03:09
A task a Bot delegated that has finished waits on nobody: its result line is
already in the conversation that asked for it, and its Bot's Activity page
lists it under Done. The board still listed it, as a Done row, untinted and
uncounted. It now leaves the board on the Bots page and on the Bot's page,
OpenDuet's Tasks and OpenFuse's bell, all by one rule (onTheBoard). A task
waiting on the person stays on each.

OpenFuse's Activity draws the board above the Bot's Activity groups and left
every task out of those groups, so a finished task would have been on
neither. The groups under the board now list the tasks the board leaves out,
under Done.
…lists them

The Bot page's row to its Activity page said "No tasks, routines or goals yet"
while the page it opens listed a finished task under Done: the row counted only
the Bot's own activity, and the Activity page lists its tasks beside it. Both
now read one set of groups (useActivityGroups): the row's in progress is what
the page lists under Needs you and Working, and it says nothing is running or
scheduled whenever the page lists anything at all.
…into its menu

At a phone's width a task waiting on a question, answered from its row, drew
Answer, Dismiss, Mark done and Open beside its words, and the words had what
was left: the title read "P.." and its status ran down the row a word to a
line. Where the row itself is narrow (a container query, so a phone, a chat's
width and a window with the sidebar beside it alike) it now stacks: the title
in full and the status across the row, and the actions on a line of their own
under them. Mark done and Open stay buttons while that line has room, and fold
into the row's menu when it does not; a task waiting on an approval puts them in
the menu it already has, so a row never has two. Wide rows are unchanged.

needs-you-board-layout.test.tsx draws the Bots page and a Bot's page in a real
browser at 390, 650 and 768 and holds every row to it.
POST /api/v2/organizations/{organizationId}/tasks/{taskId}/stop stops a working or waiting task the way Stop in its conversation does. Steps it finished stay done. Spec and regenerated clients.
…llow_up and stop

A task that is working or waiting on the person has Stop on its card in the conversation that asked for it and on its row on the Bot's Activity page, through stopTask. The card reads Stopped as soon as the answer comes back. The Bot's follow_up and stop calls draw the card of the task they reached, with what it was told on the line under it.
The two renderers were registered inside the reaction's component instead of OrchestrationTools, so the calls folded into a plain step line. Stop on a task's card is tested through the API's stopTask, red against the card without it.
… stop

The delegating section names the asks that are about a task already started (tell it to also..., stop the Figma one), so a Bot sends them to that task instead of starting another.
… and which model answered

A task's thread folds each turn's steps into one "Worked for" row. Nothing in the API said how long a turn took: the history keeps no times, and a task's workingSince covers only the turn running now. listTaskTurns answers each finished turn by the message it answered, with the time its runs took added up (a wait on the person is not counted) and the model its last run used.
In a task's thread, each turn's steps and the lines its Bot said on the way fold into one row that opens to them, as dots draws a task: "Worked for 52s", timed by listTaskTurns (the turn's run time, not the wait on the person). A turn still running reads "Working for 12s" and shimmers; one the server has not listed says how many steps it took. Anything drawn, and anything that asks the person, stays in place.
… made

A task's conversation gets a Task details pane, as dots lays out a task: Ran by (the Bot, what runs it, and the model its latest turn used, from listTaskTurns), Outputs (what it made or changed in an app, with the link the app gave), Sources (pages and apps it read) and Computer when it used one. Sources and outputs are read from the thread's own steps, by the line the server puts on each app call's answer. The Worked for row now also folds steps the app draws as cards of their own, such as a command it ran, and keeps out anything drawn for the person.
…tion

A new area, api/openapi/agents.yaml, under bot-agents: list, add, read, change and delete the agents Bots hand work to, issue and revoke an agent's callback token, and read and set a Bot's delegation (self, on its one agent, or orchestrate, handing work to the agents it may use). The client is regenerated. The harness drift test covers the agent's harness list too.
…ing an agent, its callback token, and how a Bot hands out work
…cker on a Bot that hands work out

The Agents page lists the agents Bots hand work to, with where each runs, how many Bots use it and whether it is private, and adds or changes one in a sheet that reuses New Bot's framework and address steps. /agents?new and /agents?agent still lead to the Bots. A Bot's Agents page switches it between doing the work itself on one agent and handing work to an ordered list of agents, with its own model and instructions. The conversation's model picker is hidden for a Bot that hands work out. Agents joins the main menu beside Bots, and the copy glossary gains agent and Agents.
…tests that open a Bot

A Bot's Agents row read how it works even after the Bot was deleted from this browser, so the page asked about a Bot that was gone. It now waits on the same delete as the Bot's other reads, and a delete drops what was read of its delegation. The fixtures that open a Bot answer its delegation, and the command menu lists Agents.
…work to agents switch on its own

A task handed to an agent says so in one quiet line under its card's state: who is on it while it works, who did it once done, and who it was handed to otherwise. The thread's brief reads Sent from the Bot to the agent. A task the Bot does itself names nobody. On a Bot's Agents page the switch is its own row and saves when switched, keeping the agent the Bot ran on as its default; the model, instructions and agents stay a form with Save.
…t the rail scroll

Agents joined the rail's destinations, and in a 600px window the rail's buttons were squeezed to fit, so the Bots badge rode 2px over the button above it (opendots-rail-badge.test.tsx, failing on the fused branch). The buttons keep their size and the rail scrolls where they don't all fit.
Every tool OpenBot offers a Bot has words for its line; follow_up and stop came with cards and none, so where no card is drawn the line read the bare name.
A run of tool lines with a step still unanswered once the turn is over is held for the person, and the approval card for it is drawn under the run's line. The fold took the run in, so a task waiting on a payment's approval showed "Worked for 5s" and nothing to answer (the approvals journeys). Such a run, or one whose answer asks the person to connect or consent, now stays where it is; while the turn runs, an unanswered step is only working and folds as before.
… for row too

A turn the person typed into a task's thread can still read as running while its step waits on their approval, so the fold took that run in and the card with it (the approvals journey's third order). The fold now also keeps out any run with a step the approval inbox holds for the person, read from the inbox the conversation already keeps.
Slices 4 and 5 met here: the pane said the Bot ran the task, on its own kind and model, while the card and the brief said Fact finder did. A task handed to an agent now reads Handed to <agent> under Ran by, as its card says; a task the Bot did itself names none.
…sk of its own

While a task wrote a guide for EU customers, "Also write a guide for US customers" was passed into that task with follow_up, so "Stop the EU one" could only steer the task writing both. The delegating section now says another thing to make or do starts its own task whatever word it is asked with, and follow_up is for changing or adding to the work a task already does.
…ow hint

A Bot's Agents page lost its Saved. confirmation when the save's refetch drew
the form again; the confirmation now outlives that. How a Bot hands out work
is shown only to people who can change the Bot, since its owner's words can
name an agent that others see only as Private agent; the spec says the server
sends it empty to everyone else. An agent's Bot count now counts the Bots the
caller can see, and a refused delete names the ones they can see. An owner's
row on People says why its Admin switch and Remove can't be used.
…'s count

A working or waiting task's row on the board offers Stop, the same stop as its card and its
Activity row, folded into the row's menu where the line has no room. The board's "N waiting"
reads the badge's own count, so a task under a Bot no longer shared with the person is listed
there to mark done and not counted. A task waiting on a question says the question in full at a
phone's width; the row's menu is as wide as its longest item; Task details drops "yet" once the
task has ended.
…with the whole text

The board and a Bot's page listed a waiting hand-off as the approval's raw fields (Hand off, Preview, Audience, Recipients) in a box cut off at its foot, without the sentence that says who reads the text. They now show what the card in the conversation shows: the row titled as the card is (Hand this to Ledger Bot), the sentence, and the text itself. The text is never in a box that scrolls on its own: a long one is cut to a few lines with a fade and Show all, as a step's output is in the conversation.

This branch has not been deployed

No deployments
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant