[ayuxy027]

What was broken

Three places on the server where the result the user (or the agent) saw didn't match what actually happened:

• Dismissing an idea. If you dismissed an idea at the same moment it was being accepted (another device, or a second tap), the server answered with nothing and the idea card failed on it. Accepting already handles this case; dismissing didn't.
• Agent updates. When a delegated task's model said something, used a tool, then said something else, the update joined the two into one line, like "I'll make a plan.The plan is saved."
• Computer file errors. Every failed file operation showed the same "Check the path, permissions and file size" message, even though the computer had said exactly what was wrong (for example the file was too large). And if saving a file timed out, it was reported as failed, even though the file may already have been written.

What happens now

• A dismissal that loses that race returns the idea as it is now (accepted), the same way acceptance already does.
• Each model step's text is kept apart with a blank line.
• A failed file operation shows the reason the computer gave. If a file change (writing a file, writing a PDF, creating a folder) times out or is interrupted, it says the change may or may not have been saved and asks to check before trying again, instead of calling it a failure. A read that didn't finish keeps the general message, since nothing changed.

How I checked

• Added a test for each fix. Each one fails on main and passes with the change:
  • a dismissal that reads the idea right before an acceptance lands
  • a model that speaks, makes a plan with a tool, then speaks again
  • a file over the size limit, a write that times out (including when the computer's own time limit stops it), and a read that times out
• pnpm lint, pnpm typecheck and pnpm test pass.
• Not tested against a live Docker or E2B computer; the file tests use the existing computer test fixture.

Each fix is a separate commit.

[caspianchan31]

Read through the diff. All three fixes change the reported outcome to match what actually happened, and the race test (pausing db.get mid-dismissal) covers exactly the interleaving the first fix targets. A few non-blocking edge cases:

1. engine/model.ts — messageId keying. The separator treats undefined as just another id. A provider that never emits messageId compares undefined !== undefined on every delta, so steps stay joined — safe, but the fix silently no-ops there. A provider that emits messageId on only some deltas of one message is worse: the first delta that drops it compares undefined !== and inserts a blank line in the middle of a message. Separating only when both ids are defined and differ (otherwise append) removes the second case while keeping the first's behavior.

2. computer.ts — truncated on a write. stopped deliberately excludes truncated, but a truncated write confirmation has the same shape as a timeout (the change may have happened; only the JSON receipt got cut), yet it falls through to a plain "failed". Write receipts are small so it is unlikely, but treating a truncated write like the timeout case would close it. Same question for any mutating operation beyond write/write_pdf/mkdir if files.py exposes one: a stopped delete or move is equally ambiguous.

3. service.ts — the fallback get. It also makes a double-dismiss idempotent (CAS fails because the idea is already dismissed, the get returns it), which is correct. One assumption worth noting: if ideas can be deleted concurrently, the get returns undefined and the caller hits the same "answered with nothing" this PR removes. Fine if ideas are not deletable.

Nice that each fix is a separate commit with a test that fails on main first.

[ayuxy027]

thanks for going through this.. checked all three against the code and looks like none can happen today

• messageId: everything goes through tanstackAgent and splitTextAtToolCalls stamps an id on every text chunk, so a partial id never reaches this
• truncated write: limit is 2 MB and a write only prints a small path json.. files.py has no delete or move either
• fallback get: nothing removes ideas, only events and sessions get removed

gonna keep this as is unless a maintainer wants the extra guard