#5996 [GUI] helpers on one card: three read-only helpers ran 163, 266 and 351 minutes, with no cost warning and no ceiling

open · @xvcezrqpbxtpzblrt · 4 commentaires · Sur GitHub

BenchmarksSetup & installServer & APIAMD / HIPNVIDIA / CUDAModels & quantsDocumentationWindows

Description

### Area

[GUI] — the app's Subagents panel and what a fan-out costs on a small card. The engine numbers below come from this app's own records and from this repository's docs.

### What do you want to do?

Two things about helpers (subagents) on one 16 GB card:

1. **Know what a fan-out costs, before and while it runs.** I asked for three read-only helpers in one chat on this machine. They ran **163, 266 and 351 minutes**. The panel never said they share one decode stream, never said which one is decoding now, and gave me nothing to judge when any of them would end.
2. **Cap a helper** — a ceiling in minutes, steps or tokens — and have a run that hits it reported as *stopped by the ceiling* instead of ending whenever it ends. The three jobs here were "read this code, read that code, check the references". One of them ran almost six hours.

### What do you do today instead?

Either sit through it and hope — while not daring to press Stop, because stopping loses what the helper did and the panel does not show how much there is to lose (that half is #5254) — or start helpers before bed and read the result the next day. On this machine that is the honest option today, and it is why I understand people who turn helpers off in the agent rules.

### What this run recorded (read from the app's own files)

`C:\Users\123\.strata\data\chats\s_pjgpfh0m7v6x.json`, `session.agents`, recomputed into windows (local time):

| helper job | started | ended | wall clock | the row's token figure |
| --- | --- | --- | --- | --- |
| read the app's host layer (read-only) | 11-15 02:08 | 11-15 04:52 | **163.3 min** | 3,038,710 |
| read the bundled runtime's compaction code (read-only) | 11-15 02:08 | 11-15 06:34 | **265.7 min** | 3,099,729 |
| check official reference material (read-only, network) | 11-15 02:08 | 11-15 08:00 | **351.2 min** | 3,437,799 |

All three were open at the same time: the windows overlap 163.3, 265.7 and 163.3 minutes pairwise. The token figure is what the app counts, summed from the usage chunks of every request the helper made, input and output together (`server/agent/children.mjs:96`) — so a helper that re-reads a long prompt at each step grows that number even when it learns nothing new.

### Why three helpers do not mean three times faster, on this machine

- `docs/DETAILS.md:504` — the server setting `parallel` defaults to `1`.
- `docs/BATCHING.md`, first sentence: "By default Strata serves **one request at a time**: the others wait in the server's queue."
- this machine's model config `C:\Users\123\AppData\Local\Strata\install\strata-iq3_s.json` sets **no `parallel`** key at all (its `args` are the pack, the two model files, an expert profile, `--expert-cache auto --prefill auto --spec 4 --spec-min-p 0.5 --mtp … --max-context 393216 --rope-scaling yarn --rope-scale 1.5 --kv k8v4 --vision --vram-reserve-mib 512`).

So three helpers were live, and the engine decoded one request at a time. What one stream costs here, from the app's own engine log (`%LOCALAPPDATA%\Strata\install\strata-iq3_s.log:36165`), quoted whole:

```text
strata serve: prompt 201825 tokens = 6531 reused + 195294 read in 176512 ms (1106.4 tok/s), 724 generated in 19417 ms (37.3 tok/s), drafts accepted 355 of 576, 6 checkpoints
```

and the same log fills with lines such as `strata serve: conversation cache: skip parking (snapshot 4035 MiB exceeds available budget)` (at `%LOCALAPPDATA%\Strata\install\strata-iq3_s.log:18260`), which is why the prompt has to be read again at all (that is #3891 and #4226).

The copy does not prepare anyone for this. `apps/host/src/locales/en.json:132` and `zh.json:132` ("… 多个智能体可以**并行**研究、写作和审阅 …") are true at the session level, and the panel (`src/components/SubagentRow.tsx:36-39`) shows only a token figure and elapsed time — no "2 of 3 waiting for the engine", no estimate, no ceiling. The host already counts working helpers for the line `Helpers are working (N)…` (`server/agent/agent-bridge.mjs:611-612`), so the number is in hand and is not put to use for cost.

### What I would suggest

- **Say what is sharing what.** In the panel: which helper is decoding, how many are waiting, and one sentence that this model config serves one request at a time. The host knows both (the model config, and `turn.children` at `agent-bridge.mjs:466`).
- **An opt-in ceiling per helper** — minutes, steps or tokens — chosen when the helper is started, with the run reported as stopped by the ceiling. The shipped delegation budgets only depth (`maxDepth`; upstream `packages/subagent/tool-subagent/src/index.ts:77` says the rest "belongs to the child runtime or its own deployment"), so today nothing bounds how long or how much a helper may spend.
- **Warn before the fan-out** when `parallel` is unset: "3 helpers, one decode stream: expect about 3× the wall clock. `"parallel": 2` in the model config lifts that if the card has room (docs/BATCHING.md)."
- **Let me see what it has done** (#5254), so the `queue`/`steer` message the app already offers (`agent-bridge.mjs:1191-1203`) can be used to say "wrap up now" instead of waiting it out.

### Versions and hardware

Strata 0.2.13 · engine 0.1.59 · source read at main `13eedf552417a6a791a7d13c6597e24300ee3a09`. RTX 3080 Ti Laptop 16 GB (16384 MiB) · 64 GB RAM · Windows 11 Pro 10.0.26300 · i7-12800HX · driver 621.14, read 2026-11-15 with `nvidia-smi --query-gpu=name,memory.total,driver_version` and `Get-CimInstance Win32_ComputerSystem` / `Win32_OperatingSystem`. Model Qwen3.8-Flash-Next, IQ3_S pack, 384K context (`--max-context 393216`), KV `k8v4`, YaRN 1.5, MTP drafting.

### What I checked and what I did not

**Checked on this machine:** the chat record above and the pairwise overlap; the model config's missing `parallel`; the engine log line quoted; `docs/DETAILS.md:504` and `docs/BATCHING.md`; the panel code and the host's child bookkeeping.

**Not run:** no controlled comparison of `parallel: 1` against 2 or 3 on this card — at 384K context on 16 GB extra slots may not fit, and I did not restart the engine to find out; I did not measure how much of those 163-351 minutes was queueing and how much was decoding; I did not test how hard a per-helper ceiling is to add; the token figures are the app's counters from usage chunks, not a tokeniser count; I did not run the app's `npm test`.

### Related

#5254 (being able to read a helper's progress: the other half of the same problem, and I added code-level evidence there), #3891 and #4226 (keeping alternating conversations warm so they do not re-read the prompt), #5131 (documenting the `parallel` default), #5239 (a fit-to-card choice of 3 parallel slots that could not load — the same VRAM question from the setup side), #4218 (the same panel mis-reporting a stopped helper after a restart).

Sur le site

Liens install, modèles, releases.