feat(sheets): an agent column can run a base agent, and arrives already pointing at one - #22
Merged
Merged
Conversation
…dy pointing at one A new sheet was never runnable. The builder cannot see the person's own agents, so it wrote harness_id "" and said to pick one; the shipped templates ship blank for the same reason. Every route to a sheet therefore ended at a Run button that refused on every column until you opened each menu in turn. The missing piece was that a BASE is a perfectly good answer. Its id IS its base name, which the server already accepts as a harness id, so a column can run one without anybody having configured anything. And unlike a chrn_ id, a base id is stable, so the builder can write one without inventing anything. - The column picker lists base agents beside the person's own, under a heading for each. Bases come from /v1/bases and are filtered on what this deployment reports ready with an available model, never a list written down here: offering one that is not installed would produce a sheet that looks configured and fails on the first row. - The builder now names a base suited to the work. The skill documents the ids and its validator accepts them, so a sheet arrives runnable. - The app fills any blank agent column with a base it has just confirmed this deployment can run, which covers the templates and every sheet written before this. Blanks only: a column pointing at an agent that no longer exists keeps refusing, because quietly re-pointing it would run something other than what the sheet says it runs. The kit's own Harness stays excluded from the list, since a sheet whose column runs the sheet's own agent is recursion with a file-write race in it. The base it happens to sit on is not excluded: that is a different agent, with its own session and no interest in sheet.json. Verified on the public VM: a sheet asked for in one sentence came back with its agent column bound to codex, chosen by the builder itself, and Run enabled without touching a menu. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01X7VoW2QzfDkxEoudktbjVY
…nswer was thrown away Two things the base-agent change walked into, both visible on the first run of a sheet built the new way. The banner. The app carries its own copy of the sheet rules in model.js, and only the skill's Python copy learned about base ids — so a sheet the builder had just written correctly was reported broken by the app that asked for it: 'harness_id: is "codex". Leave it "" unless you were given a real agent id.' The empty cells. A cell was filled only when the terminal status read exactly "completed". A turn read a moment before its label catches up comes back terminal with its answer already in the record, and the cell threw that answer away and wrote "the turn ended without an answer" over it. An answer is an answer whatever the turn was labelled, so the content decides now and the status only chooses the wording when there is nothing to show. A Stop stays a Stop: partial output under a green tick would read as a success nobody got. Verified on the public VM: the same three rows that failed came back done, with real sentences, and no banner. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01X7VoW2QzfDkxEoudktbjVY
An agent cell keeps its files under its answer, which makes the row tall. Every other cell in that row was showing one ellipsised line above a block of empty row, because the line count was one number shared by the whole row. The measuring now happens per cell, in SheetGrid (reifyui 0.11.2). This side is the half the kit owns: the answer text marks itself with data-shg-clamp-text so the grid can tell it apart from the file cards beside it and work out how much room each actually has. NOTE: needs reifyui 0.11.2 on npm. The fix is in that package; this bump is what picks it up. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01X7VoW2QzfDkxEoudktbjVY
… we ship with The previous commit asked for reifyui ^0.11.2, which is where the fix belongs but is not on the registry — so `npm ci` failed with ETARGET and the sheets build went red. Doing it here instead makes the fix work against the version already published, and CI green without waiting on a release. --shg-clamp is set per <td>. A custom property resolves from the nearest ancestor, so this wins over the row-level value SheetGrid sets, and it reaches both this kit's answer text and the grid's own value cell. Three things have to be true at once, and each is a trap on its own: - The basis cannot be the RENDERED height. More lines makes the row taller, which allows more lines: measured live the rows climb 92 → 137 → 182 and never settle. Every clamp goes back to one line before measuring, so the basis is the height the row's OTHER content needs — which the clamp cannot influence. - The basis cannot be the <td>. A table cell is stretched to its row and reports the row height for every column, so they all look equally full. Measured on the text's own box, which sizes to its content. - The count cannot be shared by the row. A cell carrying files under its text has far less room than one carrying nothing; one number for both is the bug restated. Verified on the public VM against reifyui 0.8.1: a 137px row gives its file-card column 4 lines plus the card and every other column 7, a 272px row gives 13 and 16, and a second pass computes the same answer rather than ratcheting the rows upward. The same fix is on ReifyUI main for 0.11.2. When that is published this copy should go, so the measuring lives in the grid that owns the markup. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01X7VoW2QzfDkxEoudktbjVY
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
The problem
A new sheet was never runnable. The builder agent cannot see the person's own agents, so it wrote
harness_id: ""and told them to pick one in the column menu. The shipped templates ship blank for the same reason. So every route to a sheet ended at a Run button that refused on every column until you opened each menu in turn.The missing piece
A base is a perfectly good answer, and the server already supports it: a base's id is its base name (
codex,opencode, …), and_backend_of_builtinaccepts it as a harness id directly. So a column can run one without anybody having configured anything first.And unlike a
chrn_id, a base id is stable. That is what makes it safe for the builder to write one: it is not inventing an id it cannot see.What changed
/v1/basesand are filtered on what this deployment reportsreadywith an available model, never a list written down in the kit. Offering an uninstalled base would produce a sheet that looks configured and fails on the first row.The kit's own Harness stays excluded, since a sheet whose column runs the sheet's own agent is recursion with a file-write race in it. The base it happens to sit on is not excluded: different agent, own session, no interest in
sheet.json.Verified on the public VM
codex, chosen by the builder itself (confirmed in the agent's own end-of-turn capture), and Run enabled without touching a menu.🤖 Generated with Claude Code
https://claude.ai/code/session_01X7VoW2QzfDkxEoudktbjVY