1.0.7: BOT_SELF_CHECK is swallowed by the BOT_PARALLEL comment and never written

[0xKr4kN]

In 1.0.7 (793f6ea), the new BOT_PARALLEL entry in write_env() does not close
its line, so the BOT_SELF_CHECK entry that followed it is now part of its
trailing comment:

f"BOT_PARALLEL={st.get('options', {}).get('parallel') or (1 if cap <= 60 else 2 if cap <= 150 else 4)}",   # générations simultanées : moins = machine plus fraîche f"BOT_SELF_CHECK={'1' if tier in ('A', 'B') else '0'}",

BOT_SELF_CHECK is therefore never written to engine/.env, and
engine/kibble-bot.py:286 falls back to its own default:

if why == "ollama" and len(out) >= 300 and env.get("BOT_SELF_CHECK", "1") == "1":

So the self-check pass now runs on tiers C and D, where app.py intends 0 —
the inverse of the intent, and on exactly the machines the setting exists for.
It costs a second generation per delivery on the small models that
kibble-bot.py's own comment says rewrite their answer anyway.

Reproduce: run 1.0.7 on a tier C or D machine, then look at engine/.env —
there is no BOT_SELF_CHECK line. Before 1.0.7 it read BOT_SELF_CHECK=0.

Fix: put it back on its own line.

f"BOT_PARALLEL={...}",   # générations simultanées : moins = machine plus fraîche
f"BOT_SELF_CHECK={'1' if tier in ('A', 'B') else '0'}",

Found on a tier D machine running 1.0.7 in a container: the line reads fine in
the source, only the missing variable in the generated .env gives it away.

Thanks for 1.0.7 — the temperature pause is the right shape for the problem. One
note from the container side: it reads gpu_temp, so on a machine with no
nvidia-smi the limit is settable but can never fire. A word in the UI when no
temperature is readable would save someone assuming they are protected.

[Ephrett]

Confirmed, and fixed in 1.0.8 (9c92cec). Your reading of it was exactly right: the two f-strings were on one source line, so BOT_SELF_CHECK became part of the BOT_PARALLEL trailing comment and never reached engine/.env, and engine/kibble-bot.py then fell back to its own "1" default — which is why the symptom was the self-check running on tiers C and D rather than an obvious error.

While checking that line I found a second problem in the same card, which I think matters more for you specifically. The power setting offered 1 parallel generation, and the worker only claims while backlog < BOT_PARALLEL:

g_due = len(st["claim_times"]) < cap and g_min < max(1, -(-cap // 60)) * 2 and backlog["ollama"] < MAX_PARALLEL_WRITES

At 1, the agent stops claiming as soon as a single generation is in flight, so a machine set to the coolest option goes quiet and delivers nothing — the opposite of a gentle setting. 1.0.8 offers 2, 3 or 4, the automatic value never picks 1, and /api/options rejects it.

If you set parallel to 1 to keep your PC cool, that is the reason it went quiet. Use 2 with a lower pace instead: the pace slider is the setting that actually reduces heat, and the thermal limit still pauses you above the temperature you choose.

Thanks for the report — the line number and the fallback made this a two-minute fix instead of an afternoon.