[MarkusNeusinger]

Summary

• Owner decision 2026-10-21: chat windows (sessions) of one user work in parallel. Before, a plot rendering in window A refused window B's paste at once with 409 run_active, which read as "a run or another paste is in progress". Now paste, chat turns, data steps, file conversion and the theme toggle run in every session immediately. Only sandbox work (plot_pipeline, run_data_code) goes through the existing run queue, with its queued, position N status. Within one session it stays one thing at a time.
• Per-user cap: at most AGENT_USER_MAX_RUNS sandbox runs (default 3, owner's choice) may be queued or running per user, across all sessions. A 4th run ends with the new stream error run_limit, shown as "you already have 3 plots or data runs in progress in other windows; wait until one finishes, then try again". The cap lives in RunQueue.submit (QueueUserLimit). Queue capacity and ordering are unchanged.
• Locks are keyed by session id: _active(runtime, sid, *, paste=True) is checked on every route, _data_route_free is folded into it, and Runtime.pasting and Runtime.converting hold session ids. Each per-session protection is kept, including the paste-versus-data-route race inside one session. Only the cross-session locks are gone.
• Budget overlap: counts that are checked and booked without an await in between (messages, pastes, conversions, pipeline runs) cannot overshoot. The daily token budget can overshoot by the model calls already in flight in parallel turns. The daily data-run count can overshoot by at most 2. Code comments document both.
• Plumbing:
  • run_limit is added to telemetry ErrorCode, to the BFF allowlist _ERROR_CODES (otherwise it would reach the browser as internal), to STREAM_ERROR_CODES and to the page texts.
  • The run_active texts now say "in this window".
  • Docs: docs/reference/api.md, docs/concepts/agent-network.md and agents/README.md are updated.

Plan

N/A. These are owner decisions from the session of 2026-10-21, implemented by a delegate. Gemini 4 Argon reviewed it first; its verdict is in a comment.

Test plan

• uv run pytest tests/unit tests/integration -n 4: 12,239 passed, 3 skipped.
• Queue tests:
  • per-user cap: the 4th entry is refused while another user's still queues;
  • the cap is checked before the queue length;
  • finished, withdrawn and expired entries free their place.
• Flow tests:
  • window B pastes and chats while window A renders;
  • A's own second turn still gets 409;
  • B's plot queues;
  • the 4th sandbox run ends with run_limit and no pipeline model call, while another user is unaffected;
  • the toggle blocks only its own session;
  • two sessions convert at once.
• BFF: run_limit passes through. Frontend: event parsing, hook and page texts.
• ruff, mypy api core, tools.changelog check; frontend lint, fm:check, type-check, test (1,166) and build.
• After deploy, live check: start a plot in window A, then paste and chat in window B. B's plot should show queued.

🤖 Generated with Claude Code

https://claude.ai/code/session_01GdMSqLR5ww4ji74EUSmk9M

[Copilot]

🔵 Needs a closer look

Session concurrency and shared-budget behavior require final human validation, including the pending live multi-window check.

No open issues 🎉

What changed in this PR

Enables parallel agent-chat sessions while limiting each user’s queued and running sandbox work.

Changes:

• Makes activity locks session-specific.
• Adds a configurable per-user sandbox cap, defaulting to three, with run_limit reporting.
• Updates concurrency tests, browser messages, and documentation.

File	Description
tests/​unit/​api/​test_agent_router.py	Tests run_limit forwarding.
tests/​unit/​agents/​test_settings.py	Checks the default cap.
tests/​unit/​agents/​runtime/​test_turn_limits.py	Tests session-specific paste protection.
tests/​unit/​agents/​runtime/​test_service_flow.py	Tests independent sessions.
tests/​unit/​agents/​runtime/​test_run_queue.py	Tests cap enforcement and release.
tests/​unit/​agents/​runtime/​test_run_queue_flow.py	Tests parallel windows and queue limits.
tests/​unit/​agents/​runtime/​test_data_code_tool.py	Tests data-run limit errors.
tests/​unit/​agents/​runtime/​test_convert_route.py	Tests parallel file conversions.
docs/​reference/​api.md	Documents session locks and errors.
docs/​concepts/​agent-network.md	Updates concurrency and resource bounds.
changelog.d/​agent-parallel-windows.md	Records the behavior change.
app/​src/​sections/​agent-chat/​messages.ts	Adds actionable limit messages.
app/​src/​sections/​agent-chat/​DataPanel.test.tsx	Updates busy-message expectations.
app/​src/​sections/​agent-chat/​ChatThread.test.tsx	Tests session and cap messages.
app/​src/​lib/​agent.ts	Accepts run_limit events.
app/​src/​lib/​agent.test.ts	Tests error parsing.
app/​src/​hooks/​useAgentSession.ts	Updates session-lock commentary only.
app/​src/​hooks/​useAgentSession.test.ts	Tests limit-error completion.
api/​routers/​agent.py	Allows run_limit through the proxy.
agents/​stream.py	Documents the new stream error.
agents/​README.md	Documents parallel sessions and configuration.
agents/​main.py	Applies session locks and handles cap errors.
agents/​anyplot/​tools/​session.py	Propagates data-run cap errors.
agents/​anyplot/​theme_render.py	Updates session-exclusion documentation.
agents/​anyplot/​telemetry.py	Recognizes run_limit telemetry.
agents/​anyplot/​settings.py	Defines the configurable user cap.
agents/​anyplot/​run_queue.py	Enforces per-user queue limits.

_{Review effort: Balanced}

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[MarkusNeusinger]

First review: Gemini 4 Argon

Verdict: sound with changes. Gemini confirmed that every single-session lock is kept, that the counts which are checked and booked without an await (messages, pastes, conversions) cannot overshoot, and that the cap is checked before the queue length. (Its file reads were refused, so it reviewed from the diff and marked unverified claims; I verified the ones below.)

Applied (d7718da12):

• TurnSlot.take now sweeps stale runs right before submit. A run of the user that goes stale during another turn's root call no longer counts toward AGENT_USER_MAX_RUNS until the next request sweeps. I added a flow test for a stale run in another window. It covers the sweep at the request's start; the take-time sweep covers a window of seconds that the test cannot reach.

Checked, not a problem:

• "plot_pipeline may attribute capacity on run_limit." The pipeline maps no code. On a refused slot it returns, and the route's TurnSlot._fail(code) sets the ledger error (agents/anyplot/pipeline.py:1321, agents/main.py:466). The flow test asserts error run_limit and chat_turn.error == run_limit.

Accepted for now (admin-only; listed for the public launch):

• Daily token budget: a user's parallel turns can overshoot it by the calls in flight. The rate limits (10 messages/min, 10 pastes/min) bound this, and a code comment documents it.
• Refused runs still count: a run_limit (like a capacity) refusal still counts as a daily pipeline run, because BudgetPlugin books the run before the slot, as it already did for capacity.
• Theme toggles: they are no longer limited to one per user. They wait in the render slot's toggle lane behind pipeline renders, give up after 120 s, and cost no tokens.
• The hardcoded 3 in the page text matches the other page texts that quote server limits.

Fable was not involved: no finding was serious, and the chat is admin-only. Copilot is still to come.