Skip to content

FIX-1629: Tasks tab rows open in place and run the flow's task actions - #2347

Merged
jhoffner merged 26 commits into
mainfrom
fix/fix-1629
Sep 29, 2026
Merged

jhoffner merged 26 commits into
mainfrom
fix/fix-1629

Conversation

@jhoffner

@jhoffner jhoffner commented Sep 28, 2026 •

Copy link
Copy Markdown
Collaborator

Requested by Jake · project thread

The problem

In the DevTool's Tasks tab, opening a task put a raw JSON block over the table, cut off at the edge of the pane. On a laptop, a parked row's reason was off screen, and at 1280 wide its status was too (FIX-1523). Nothing in the tab could change a task. To cancel, re-prioritize or answer one while debugging, you had to write a curl command or a throwaway action.

What this does

Each task is now a row that opens in place. Status comes first, then the goal, then the reason. Opening a row shows every field the task carries, with the raw JSON folded at the bottom. An open row lists the flow's actions that take a taskId, with the id already filled in. Running one goes through the panel's normal action path, and the row shows the result. A refusal is shown as a refusal, in the tool's own words.

  • taskToolActions(board) (orchestration) turns a durable board's eight task tools into ordinary flow actions.
  • A channel whose CHANNEL.md says boardActions: true exposes its boards' tools the same way, and each works only in that channel's own session.
  • The kitchen-sink support.help channel opts in.

Two panels. Left, today: a task table whose Details expander floats a clipped JSON block over the rows. Right, after: the same row opened in place, fields listed with the full goal wrapped, the raw JSON folded underneath, and an Actions strip offering cancelTask, updateTask and answer with the task id already filled.

The shipped row matches the figure, with one exception: the dispatch-run link sits after the latest-change kind, outside the row's toggle button, because a button can't hold another button.

flowchart LR
  L["multi-seat-collab hire · no model · shipped DevTool"] --> R["Tasks tab · 1280 wide · row collapsed"]
  R -->|"status and reason in view"| E["expand the parked row"]
  E -->|"every ledger field readable in the pane"| A["run an action from the row"]
  A -->|"ledger changed · row shows it · no reload"| P["PASS"]
  C["control · main, or channel-actions-off"] -.-> R
  A -.->|"under a control"| F["must FAIL · names its leg"]
Loading

Goal check: PASS. The four legs are read, answer, tool and refuse. channel-actions-off fails at tool and refuse only. Today's main DevTool fails at read, answer and tool. The results are in the verdict log.

What's asked of you

This builds the approved spec. The open fork was merged as ON, so support.help turns on boardActions. One point is a deviation from the spec's wording and needs your call:

After you answer a parked row, that row reads "in progress", not "completed". Accept that for now?

  • Plain terms. In the lab, workers hand each row to a run of its own. Answering from the seat's row really does move the task. The row you clicked leaves parked and shows the answer's claim (in_progress). The finish happens in that child run, and the row's own run link opens it showing completed, with no reload.
  • Trade-off. Making the clicked row show completed needs rows to update live from changes made in other sessions. That is FIX-1506, which the spec named as out of scope.
  • Recommendation. Accept. The goal check grades exactly this: the row leaves parked, and the run it links to shows the ledger's final status.
  • What would change my mind. If people answer from the seat's row and trust its status as final.
  • If this is wrong. Someone may read in_progress as "my answer didn't finish", for as long as FIX-1506 is open. Nothing is lost or written wrongly.

Parts worth reviewing closely

This section is for the code reviewer.

  • How a row reads an action's outcome. See packages/devtool/src/react/lib/task-actions.ts → outcomeOf. The action route returns only a request id, and a refused verb emits no change item. So the row reads the tool's { ok, error } from the root block_trace of the request it sent. Of everything I built, this matching is the part I'm least sure of. The goal check exercises it for real, and unit tests cover the other outcome states.
  • The channel's session fence. See packages/workforce/src/channel/channel-flow.ts → boardTaskActionsFor. Every channel shares one flow, so the resolver refuses unless ctx.session.identity.id is the board's own channel. That identity is set by the server, so the check follows BP-031.
  • How taskToolActions(board) finds the ledger. See packages/orchestration/src/task-board/board-resolver.ts. taskBoard() records its resolver in a WeakMap keyed by the handle, so the handle's public type is not widened. Each action also composes board.capability, so a flow doesn't have to declare the collection.

Fixes FIX-1629 · Spec: specs/issues/FIX-1629/ on main · Spec review: #2346 · Goal: goals/devtool-workforce-visibility/works-a-task-from-its-row/ (run with PLAYWRIGHT_BROWSERS_PATH=/opt/pw-browsers pnpm tsx goals/devtool-workforce-visibility/works-a-task-from-its-row/run.mts; verdict log in its goal.md)

How to review this

How to review this. This implements an approved spec — the approach and the numbered
decisions in its DECISIONS.md are already signed off by a human, so please review the
code against that direction, not the direction itself. The canonical set is retained
at specs/issues/<ISSUE-ID>/ on main; Linear links to it and the original merged spec
PR, which keeps review history. Follow-up amendment PRs explain later direction changes.
Current architecture docs remain the authority for shipped behavior (BP-037).

Most valuable here: correctness on the second path (the legacy shape, the null
boundary, the concurrent case, the cancel path), anything deriving an auth or routing
decision from caller-controllable input, and a behaviour the tests assert around
rather than on.

Already settled upstream: the approach and every decision in the spec's DECISIONS.md
(human-approved — if one is wrong, say so as a spec finding and it gets folded back; don't
re-argue it inline), and the conventions in docs/contributing/best-practices.md.

Engineering calls
  • BR-25 wording. When an action settles a row a worker holds, the worker's later result is declined with reason terminal, not lost-claim. The verb sees a terminal row before it checks the claim. The row is still refused and dropped, and the test asserts terminal. If this is wrong, a caller keying on lost-claim misses the case.
  • "Lease holder" is shown as "Lease until". The server redacts claimedBy, so the stream has no holder to show. The raw JSON shows whatever the stream carries. If this is wrong, a field is missing from the list, but it is still one click away.
  • Handle-form actions compose the board's capability. A flow therefore doesn't declare the collection, and the spec's resources: { [todos.id]: todos } example is gone, because DefinedTaskCollection has no .id. If this is wrong, the only effect is that the resource is also installed on the action.
  • Raw JSON is a lazily mounted <pre>, not JsonViewer. It shows the full record, byte-complete, with no re-serialization on each streamed change while it is folded. If this is wrong, it is cosmetic.
  • The board tools are also in the channel's internal.actions, beside fileTask and readBoard. If this is wrong, one more internal dispatch path exists, still behind the same session fence.
  • Two robustness fixes the spec implied. Each row has its own render-error boundary, from the failure taxonomy's "a render error in one row must not blank the tab". A second run of an action clears the first run's answer before sending, which the goal check exposed as a stale "Done". Both were written red-first.
  • Changesets are patch for all three packages, not the spec's minor. Every change is additive, and the changeset README reserves minor for pre-1.0 changes that break consumers. If this is wrong, it only affects version numbering.
  • Review was done inline. This session had no sub-agent tool, so I ran the coherence, restraint, correctness and completeness lenses myself against the spec's BR table. The findings were the two robustness fixes above.
Verification
  • pnpm typecheck passes across the workspace, including the goals package and its control-shape guard.
  • Tests on the affected packages: devtool 397, orchestration 1466, workforce 1030 and kitchen-sink 232, all passing. Every dependent of orchestration and devtool also passes, except two wall-clock assertions in integration-tests (side-chain-pool-fix554, task-board-resource-wake-stale-ref). They fail only under parallel load, and pass 3/3 when run alone.
  • Red first: the per-row boundary test failed with "Objects are not valid as a React child" before the fix. The stale-outcome test failed with expected 'ok' to be 'pending' before its fix. The kitchen-sink BR-21 test failed without the CHANNEL.md line. The earlier commits' red runs are recorded in their tests.
  • Goal works-a-task-from-its-row:
    • PASS, twice, plus once on a swapped fixture.
    • channel-actions-off: FAIL at tool and refuse only.
    • main's DevTool source: FAIL at read, answer and tool. Refuse was ungraded.
  • Re-pointed checks (C7), each on a run of its own:
    • the-checklist-rows: PASS. silent-park fails at row 4 only, and no-inventory at row 5 step 3 only. At 1280px, 92px of the reason is now visible, where before none was.
    • it-hands-a-row-between-two-seats-in-view: PASS, and all five controls fail at exactly their named legs.
  • Other CI checks: pnpm docs:build succeeds, knip:ci is clean, and the changeset, model-string, NUL, spec-folder, isolation and publish-set validators all pass.
Changes by area
  • devtool: the Tasks tab as accordion rows (task-collections-view.tsx, task-row-body.tsx); the Actions strip (task-row-actions.tsx, lib/task-actions.ts); a locked field in SchemaForm; sendAction now returns its response to the panel.
  • orchestration: taskToolActions, taskToolSuffix, and the board-resolver record.
  • workforce: the boardActions key (the seventh), withBoardActions, and the fenced per-board actions. channel-board now uses the shared suffix.
  • kitchen-sink: support.help declares boardActions: true, with a test for it.
  • goals:
    • the new goal;
    • the lab's opt-in and the channel-actions-off control;
    • two checks re-pointed from table cells to row slots.
  • docs: the DevTool's "Task boards" section, a new "Changing tasks from outside a run" section in the task board page, and boardActions in the channels docs; both package READMEs; one changeset.

🤖 Generated with Claude Code

https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N


Generated by Claude Code


Note

Medium Risk
Introduces public flow and opt-in channel actions that mutate task boards (same checks as workers, but callable by anyone who can reach the flow or channel session), plus DevTool UI that drives those actions from live sessions.

Overview
FIX-1629 adds ways to change durable tasks outside a draining worker run, and rebuilds the DevTool Tasks tab so debugging matches that model.

Orchestration & workforce: New taskToolActions(board) (plus taskToolSuffix) turns a durable board’s eight task tools into ordinary flow actions named like cancelTask_<board>. Channels can opt in with boardActions: true in CHANNEL.md (seventh allowed key); each board’s tools are then channel actions scoped to that channel’s session. Kitchen-sink support.help enables it; docs cover DevTool task boards, “changing tasks from outside a run,” and channel boardActions.

DevTool: Tasks are accordion rows (data-task-id, slot-based status/goal/reason/assignee) instead of a table with a floating JSON expander—status-first layout so parked reasons stay visible at laptop width. Opening a row shows the full record (TaskRowBody) and an Actions strip (TaskRowActions) for flow actions whose schema requires taskId, with the id locked in the form. Runs use the same dispatch path as the action bar; handleSendAction returns a request id (or error), and outcomeOf reads ok/refused/failed from the root block_trace. SchemaForm supports locked fields; action-bar helpers are exported for reuse.

Verification: New Playwright goal works-a-task-from-its-row, updated checklist/multi-seat-collab reads for row slots, lab channel-actions-off control, and expanded devtool unit tests.

Reviewed by Cursor Bugbot for commit f0a9724. Configure here.

The Tasks tab's table becomes a list of disclosure rows. A collapsed row
leads with status, then goal and reason sharing the width as
minmax(0, 1fr) tracks, so the reason stays in view on a laptop-width pane
(absorbs FIX-1523). Clicking the row, or Enter/Space on it, opens it in
place: every field the task carries, with the raw record folded at the
bottom. Open rows are keyed by board and task id and held above the rows,
so a streamed change never closes one.

The floating <details> JSON expander and its fixed 28rem block are gone.
FIX-1481's reason tests are re-pointed from the Reason column to the
row's reason slot; the presence predicate and verbatim text are kept.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…actions (FIX-1629)

taskToolActions(board) returns the eight task tools as a flow actions map,
named <tool>_<suffix>, each running the same guarded verb a model's tool
does. A refusal comes back as { ok: false, error } and writes nothing; it
is readable from the root block_trace of the request that ran it. No
action claims or drains the board, and an action settling a row a worker
holds lands like a coordinator's call, so the worker's own result is then
declined.

Request- and sequencer-backed boards are refused by name: their rows do
not outlive the request that filed them. Each handle-form action composes
the board's capability, so the ledger is installed without the flow
declaring it again. taskToolActions(collectionId, resolve) serves a caller
that resolves its own ledger. taskToolSuffix is the one qualifier rule.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…ns (FIX-1629)

A CHANNEL.md may declare `boardActions: true` (the seventh key in the
closed list, boolean only, refused by name otherwise). Each of that
channel's boards then gains the eight task tools as channel actions,
named <tool>_<board id with dots as underscores>, built with
orchestration's taskToolActions. Channels that say nothing keep exactly
today's actions.

Channels share one flow, so each action first checks the session is the
board's own channel, then resolves the ledger through the same check
fileTask and readBoard make; another channel's session is refused before
any row is read. The local boardToolSuffix copy is replaced by
orchestration's taskToolSuffix.

The reference app's support.help opts in, so its escalations can be
worked from the DevTool's Tasks tab.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…(FIX-1629)

An open Tasks-tab row lists the viewed flow's actions whose input requires
a string taskId. A board's suffixed task tools show only on that board's
rows; a generic action such as an app's `answer` shows on every board.
Picking one opens its form in the row with taskId filled and locked, and
submitting runs it through the panel's existing action dispatch, the path
the action bar uses. There is no other write.

The dispatch answers with a request id and no output, and a refused verb
writes nothing, so the row reads the tool's { ok, error } from the root
block_trace of the request it sent: success, a refusal in the tool's own
words, a failure, or "outcome isn't visible" when the request left no
trace. A dispatch that throws shows its message and keeps the form filled.

handleSendAction now returns what it dispatched; useActionDispatch
resolves a thrown dispatch to { error }. SchemaForm gains locked fields.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
… (FIX-1629)

New goal devtool-workforce-visibility/works-a-task-from-its-row: on the
multi-seat-collab hire, in Chromium on the shipped DevTool at 1280x800, the
parked row is read collapsed and open, answered through the seat's own
action, and a board task tool is run and refused from the channel's row.

The lab's channel opts in with `boardActions: true`; GOAL_CONTROL=
channel-actions-off hires it as if it had not. The two existing checks that
read the Tasks tab's Reason column now read the row's reason slot by task id.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…tions (FIX-1629)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…X-1629)

The spec's failure taxonomy: a render error in one row must not blank the
tab. Each row now sits in its own boundary, which shows a one-line note in
the row's place and retries when the task changes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
… again (FIX-1629)

Running an action a second time left the first run's "Done" on the row until
the new dispatch came back, so a refusal read as success for that long.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…r (FIX-1629)

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
@jhoffner jhoffner self-assigned this Sep 28, 2026
@vercel

vercel Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

The latest updates on your projects. Learn more about Vercel for GitHub.

3 Skipped Deployments
Project Deployment Actions Updated
flow-state-dev-docs Ignored Ignored Sep 29, 2026 12:53am UTC
flow-state-dev-kitchen-sink Ignored Ignored Sep 29, 2026 12:53am UTC
flow-state-dev-ui Ignored Ignored Preview Sep 29, 2026 12:53am UTC

Request Review

@chatgpt-codex-connector

chatgpt-codex-connector Bot commented Sep 28, 2026 •

Copy link
Copy Markdown

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review Status Commit Review trigger
📝 Code Review ✅ Completed 2026-09-29T00:59:14.004579Z 002609c Manual request
ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

@linear-code

linear-code Bot commented Sep 28, 2026

Copy link
Copy Markdown

FIX-1629

Copy link
Copy Markdown
Collaborator Author

Second-look and architecture review (FIX-1629)

I read the source diffs in devtool, orchestration and workforce. I did not run tests, and I skimmed the goals and docs. The overall shape holds up. Task changes go through the flow's own actions, the outcome is read off the root trace, and taskToolActions reuses the existing guarded verbs. I found no correctness blockers. These are points for whoever picks this up.

Worth changing in this PR

  1. TaskRowBody's dispatchRun prop is dead. task-row-body.tsx accepts it and adds a "Dispatch run" field. TaskRow renders <TaskRowBody entry={entry} id={bodyId}> without passing it (task-collections-view.tsx:349). The link stays in the collapsed row. Either pass it, or remove the prop and the field.
  2. The devtool's taskToolSuffix is copied from orchestration, and nothing checks that they agree. The package-boundary reason for the copy is fine. The failure mode is quiet, though: if the rules drift, board scoping in taskActionsFor breaks with no test failing. Add a small parity test somewhere that can import both (for example the integration-tests package). The alternative is to have the flow list expose which board an action belongs to (see the design note below).
  3. bodyIdOf builds a DOM id from a 32-bit hash plus the key length. Collisions are unlikely but possible, and they would give two rows the same aria-controls target. Use useId() in TaskRow, or encode the key directly.

Design note: a leaner approach for the board scoping

taskActionsFor decides which board an action belongs to by parsing action names. It matches _<suffix>, checks the suffixes of the other boards, and falls back to a hardcoded TASK_TOOLS prefix list. This is the most fragile part of the PR. It duplicates the naming rule and the list of eight tools in a second package. A user action that happens to end in a board's suffix would be misattributed.

The cleaner version has taskToolActions tag each action with its board, for example { block, description, meta: { board: collectionId } }. It would ride through the flow list's actionSchemas or a sibling map. The devtool would then filter on the tag, and the mirrored suffix rule, the tool list and the "other boards" inference would all go away.

I would not block this PR on it, because the change crosses the flow-list wire shape. Filing it as a follow-up seems right. If you keep the name heuristic, item 2 becomes more important.

Refactor opportunities that ride along with this feature

  • withUses is spread into all eight tool definitions in buildTaskTools. It could be applied once where taskToolActions maps the tools into actions, or in buildTaskTools' return, so the eight definitions stay untouched. This is the one place where the diff touches every tool for a single-caller concern.
  • The WeakMap in board-resolver.ts is a side channel. It exists to avoid widening TaskBoardHandle. That is a defensible trade. A symbol-keyed or @internal field on the handle would do the same job with less indirection and would survive a handle copy or spread, which the WeakMap would not. If you keep the WeakMap, note the copy case in the taskToolActions throw message, since "was not built by taskBoard()" would be misleading for a spread handle.
  • taskToolActions has two overloads, one for a board and one for a collection id plus resolver. The second exists for the channel fence. That is acceptable. Its only caller is channel-flow, so it could stay @internal and off the public barrel, which would keep the public API to one form.
  • DevToolPanel.handleSendAction now returns a union (RowDispatch). The sendAction hook's return type also changed from null to { error }. I found only one caller of sendAction, so this is safe. The "request" in response narrowing is worth a comment, or better a discriminant field, so it survives future response shapes.

Positive

  • Moving resolveFor and removing the orphaned docblock in channel-board.ts fixes a real doc-attachment problem. It is a good small cleanup that fits this change.
  • Fencing channel board actions by session identity in boardTaskActionsFor is the right call. Every channel shares one flow, so the actions would otherwise be callable across channels.

Generated by Claude Code

@cursor cursor Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Simplify review (FIX-1629)

Verdict: The core product bet looks right — reuse the existing action bar dispatch path, expose the eight verbs as ordinary flow actions via taskToolActions, and discover row actions from taskId in the schema rather than inventing a DevTool-only mutation API. That matches the spec diagram (no new route, refusal from root trace) and keeps Workforce on the same guarded verbs as models. I’m approving on direction; the notes below are about cumulative complexity and follow-ups, not blockers.

Did the approach go too far?

Probably not on the framework slice. taskToolActions reusing buildTaskTools is the main alternative to a second implementation of cancel/complete/fail. The board-resolver WeakMap is a smell, but it’s the same cycle-breaking pattern used elsewhere in orchestration; putting resolve on TaskBoardHandle would be cleaner at the type level and would avoid the side channel — at the cost of widening a handle every consumer sees. I’d keep the WeakMap unless you’re already planning a handle shape change.

Where complexity stacks: (1) devtool mirrors taskToolSuffix and a hard-coded TASK_TOOLS list with multi-board suffix heuristics; (2) a large goal harness (works-a-task-from-its-row/run.mts); (3) Tasks-tab UI that re-derives on every stream tick because rowActions embeds full requestGroups. Items (1) and (3) are the main maintainability/perf risks; (2) is justified by the spec’s anti-game bar but could share helpers when a second goal needs the same Playwright steps.

Simpler alternatives considered

Alternative Trade-off
Accordion/read-only first, actions later Smaller PR; spec and user ask explicitly rejected this
DevTool hard-codes only the eight suffixed tools (no takesTaskId discovery) Less filtering code; breaks parked-row answer and other app actions
Import taskToolSuffix from @flow-state-dev/orchestration in devtool Removes duplication; widens devtool’s runtime dependency graph (may conflict with intentional narrow deps)
Move pinned suffix + tool basenames to @flow-state-dev/contracts Best long-term DRY for models, actions, and DevTool; zero/heavy-dep friendly
Dev-only parity test (devtool mirror vs orchestration export) Cheap drift guard without changing prod deps
Stabilize rowActions (don’t pass entire requestGroups; lookup by requestId or memoized outcome map) Less Tasks-tab churn during unrelated streaming; optional devtool perf pass

What’s already in good shape

  • Reusing SchemaForm + buildInputFromForm / getDefaults from the action bar.
  • taskActionsFor tests for multi-board suffix cases.
  • Workforce boardTaskActionsFor using the two-arg taskToolActions(collectionId, resolve) instead of the WeakMap path.
  • Open-row state lifted above rows so stream updates don’t collapse accordions.

Skipped (needs owner call or larger refactor)

  • Replacing RowBoundary class boundary.
  • Refactoring the goal runner until a second consumer exists.
  • Tightening TASK_TOOLS.some(prefix) unless you’ve seen real collisions with app action names.

No changeset concern from this review — the orchestration surface change is appropriately covered.

Open in Web View Automation 

Sent by Cursor Automation: Review For Best Approach

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: f0a9724df7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

@cursor cursor Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cursor Bugbot has reviewed your changes using high effort and found 3 potential issues.

Fix All in Cursor

❌ Bugbot Autofix is OFF. To automatically fix reported issues with cloud agents, enable autofix in the Cursor dashboard.

Reviewed by Cursor Bugbot for commit f0a9724. Configure here.

@cursor cursor Bot left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Code Snob. A smaller shape meets the same contract. Draft POC, do not merge: #2349

Fold these three subtractions into this branch:

  • Delete bodyIdOf. It hashes the open-state key into a DOM id. useId() is the id aria-controls needs, and the row tests only check that the toggle points at the body.
  • Delete TaskRowBody's dispatchRun prop. Nothing passes it. The run link already sits on the collapsed row, which is the layout this PR shipped.
  • Stop spreading uses into all eight tools. handler.withDefaults({ uses }) applies the board capability once inside buildTaskTools. The action that reaches the ledger without the flow declaring the collection still passes.

Left alone on purpose: the DevTool's copy of taskToolSuffix (package boundary this PR already chose), the name-suffix filter (the spec's rule for which actions a row offers), and the WeakMap on the board handle.

Agent on this PR: fold that subtraction, then close the POC unmerged after this lands.

View PR

Open in Web View Automation 

Sent by Cursor Automation: Code Snob

…prop, one withDefaults (FIX-1629)

The row's aria-controls id comes from useId() instead of a hash of its key.
TaskRowBody's dispatchRun prop is gone; nothing passed it and the run link
sits on the collapsed row. buildTaskTools applies the board capability once
through handler.withDefaults({ uses }) instead of spreading it into all
eight tools. From the draft POC #2349.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…d writes, concurrent row dispatch, suffix collisions (FIX-1629)

- taskActionsFor matches a task tool's suffix whole, and gives an app's
  suffixed action to the board with the longest matching suffix, so
  cancelTask_feature_work is no longer offered on the work board's rows.
- outcomeOf reads the last root trace (the in_progress entry comes first)
  and reads a declined TaskWriteOutcome as a refusal.
- handleSendAction fences a dispatch on the workspace only (isCurrent), so a
  second row's action no longer retires the first row's answer.
- A channel kind refuses two opted-in boards whose task actions would share
  names (eng.feature.work / eng_feature.work) instead of letting one replace
  the other; taskToolActions documents the non-injective qualifier.
- TaskToolActionsBoard.capability is DefinedCapability, not <any, any>.
- Code quality: unused screen import; a null check the caller already makes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

jhoffner commented Sep 28, 2026 •

Copy link
Copy Markdown
Collaborator Author

from: fsd-claude
session: fix-1629-pr-feedback-r1
kind: reply

Re Code Snob (review 5345338114, POC #2349): all three subtractions held, and they are folded in 7ca1aa2.

  • bodyIdOf is deleted. The row's aria-controls id comes from useId(), and the row tests pass.
  • TaskRowBody's dispatchRun prop is deleted.
  • buildTaskTools uses handler.withDefaults({ uses }) once in place of the eight ...withUses spreads. I checked that the test "reaches the ledger when the flow does not declare the collection" actually depends on it: with uses dropped, that test fails. With the fold in, it passes, and orchestration is 1466/1466.

#2349 stays open for now. It gets closed after this lands.


Generated by Claude Code

… names

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

from: fsd-claude
session: fix-1629-lifecycle
kind: review

@codex review

Round-1 fixes (b8ad96f, 7ca1aa2) and a docs line (6cac5bb) landed after your review of f0a9724. Please re-review the current head.


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 6cac5bb7eb

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

… carried field, row request refresh, no submit lock (FIX-1629)

- outcomeOf resolves a ref/structure root output against the request's own
  items before classifying; an unretained target reads as unknown, never ok.
- An open row shows every field the task carries, not only the named list.
- The panel re-reads the session's requests while any row's request is still
  running, so a row whose request is not the streamed one still settles.
- A row sends a second submit while the first is in flight (BR-16: no
  client-side lock) and shows each submission's own answer, newest first.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review

Please re-review the new head, b610984 (PR feedback round 2).


Generated by Claude Code

…erated board, drops a stale pick, keeps stream-only traces (FIX-1629)

- taskToolActions builds its tools claimless: an action invoked in-process
  from inside a claimed worker's async chain no longer presents that
  worker's ticket (BR-25). Model-facing task tools are unchanged.
- The row's action matcher also knows every board the flow's generated task
  tools name: all eight `<tool>_<suffix>` present is a board, listed on this
  session or not. A sibling channel's tools stay off this channel's rows,
  and an app's lone `cancelTask_now` is still generic.
- A picked action the row no longer offers loses its form and cannot submit.
- A finished request's outcome is read from its polled log plus the items
  only its stream saw (transient traces are never persisted); rendered items
  still come from the poll.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review

Please re-review the new head, b2cb8e5 (PR feedback round 7).


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: b2cb8e5fc0

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…tream state to its request; name tasks by title

- outcomeOf: a request that ended in anything but completed reads as failed
  even when the action refused first (the refusal is carried in the message);
  a completed action waits while the request still runs.
- useRequestStream hands back only the snapshot and status of the request it
  was asked for, so the render after a switch cannot file the old request's
  log or status under the new id; the panel's request groups and Continue
  seed check the stream's request id too.
- Collapsed task rows show title ?? goal, text and tooltip.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8abb43b9b7

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…ches any board of the kind, opted in or not

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: c6448a1e98

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

Comment thread apps/docs/docs/devtool/overview.md Resolved
…r's decline to terminal; say the row action settles as a coordinator (FIX-1629)

- outcomeOf falls back to the request's latest failed error item before the
  generic "The request ended failed." (BR-15).
- The BR-25 test asserts the worker's settlement declines `terminal`, and a
  new test shows why: blockTask on a held row is an illegal transition, so
  every settlement over a claim finishes the row.
- DevTool overview: the action doesn't claim the task, settles it as a
  coordinator would, and a running worker's later result is dropped.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 8c56857b56

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

…and the worker's later result is dropped (FIX-1629)

The displaced worker's complete declines `disallowed` and its fail declines
`lost-claim`; neither writes. The blockTask test no longer claims every
settlement over a claim finishes the row.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
jhoffner pushed a commit that referenced this pull request Sep 29, 2026
…ains the retryable failTask leg

After a retrying failTask re-queues a held row, the displaced worker's
completion declines `disallowed` and its failure `lost-claim`, neither
writing (task-tool-actions.test.ts on #2347).

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N
…ser into another workspace (FIX-1629)

The panel already mounts the view under its workspace key; the test goes red
when that key is removed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 643fbdfb38

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

… (FIX-1629)

The row poll now treats `suspended` as open, through one isRequestOpen helper
that outcomeOf's pending check shares. A request resumed elsewhere has no
open stream, so without the poll the row sat on "Running…".

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014LohgYyHsXCYQd6kABH11N

Copy link
Copy Markdown
Collaborator Author

@codex review


Generated by Claude Code

@chatgpt-codex-connector chatgpt-codex-connector Bot left a comment

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: 002609cc12

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

  • Open a pull request for review
  • Mark a draft as ready
  • Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

"@flow-state-dev/devtool": patch
---

New `taskToolActions(board)` exposes a durable task board's eight task tools as flow actions named `<tool>_<board>` (with `taskToolSuffix` for the naming rule). A `CHANNEL.md` can declare `boardActions: true` to expose its boards' task tools as channel actions, off by default. The DevTool's Tasks tab now opens a task in place with its whole record, and runs the flow's actions that take a `taskId` from the row, showing a refusal as a refusal (FIX-1629).

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P1 Badge Keep the changeset summary to one sentence

This release fragment contains three user-facing sentences, while the repository requires each changeset summary to be a single sentence naming the Linear issue. Collapse the three package-visible outcomes into one sentence before publishing.

AGENTS.md reference: AGENTS.md:L36-L39

Useful? React with 👍 / 👎.

req.items!,
liveRawItems.get(req.id) ?? streamFor(req.id)?.rawItems ?? streamRawItems.get(req.id),
)
: (liveRawItems.get(req.id) ?? streamFor(req.id)?.rawItems ?? req.items ?? []),

Copy link
Copy Markdown

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

P2 Badge Retain streamed traces when the persisted log is empty

When a transient: true row action produces no persisted items—for example, it returns { ok: false } without changing the task—snapshotSupersedesLive is false because req.items is empty, and this fallback ignores streamRawItems. Once another request takes the stream slot, the previously captured root trace therefore disappears and the row regresses from the visible refusal to “this request left no trace.” Fresh evidence beyond the earlier transient-cache fix is that the terminal-empty-log branch explicitly bypasses the only retained raw-stream cache; include that cache in this fallback or reconcile raw logs independently of the rendered-item decision.

Useful? React with 👍 / 👎.

Copy link
Copy Markdown
Collaborator Author

Paused at the feedback cap. This PR has had twelve rounds of review feedback, all handled. Automated fixes stop here until the owner decides what to do next.

Two Codex threads on 002609c are still open:

  • 4128539563: the changeset summary is three sentences, and AGENTS.md asks for one.
  • 4128539571: a transient: true row action that is refused without writing anything can lose its visible refusal once another request takes the stream slot. The row then falls back to "this request left no trace".

Assessment: this is converging slowly, not looping. Nothing has been re-argued. Each round has found a new edge case, and since round 3 almost all of them are in the same place: how the DevTool reconstructs a row action's outcome from the live stream, the polled snapshot and trace refs. The structural fix is for the engine to report that outcome explicitly. That is out of scope here, because it needs the BR-22 amendment already listed as a follow-up.


Generated by Claude Code

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.

2 participants