grenade

grenade commented on issue helexa/helexa#292

2026-08-27 09:19:55 +00:00

Prefix cache reuse plateaus mid-session — deep agentic turns re-prefill ~23k tokens every turn

Fixed and verified live.

budget_mb = 2048 on beast (4102cbe2, corrected from an initial 3072 in 9c1dfc1f — see below). Measured on the running host today:

prompt_tokens=28794 …

grenade pushed to feat/nginx-search-alias-and-sync at helexa/helexa

2026-08-27 09:10:52 +00:00

432f8009ce feat(nginx): accept POST on search, and keep queries out of the logs

grenade created branch feat/nginx-search-alias-and-sync in helexa/helexa

2026-08-27 09:05:40 +00:00

grenade pushed to feat/nginx-search-alias-and-sync at helexa/helexa

2026-08-27 09:05:40 +00:00

98ba45cbb4 feat(nginx): serve SearXNG's own path, and deploy the vhost from CI

grenade created branch fix/wire-surface-in-neuron-spans in helexa/helexa

2026-08-27 08:35:45 +00:00

grenade pushed to fix/wire-surface-in-neuron-spans at helexa/helexa

2026-08-27 08:35:45 +00:00

522682282c fix(observability): record which API surface a request arrived on

grenade created branch fix/295-ignore-tiny-prefill-samples in helexa/helexa

2026-08-27 07:45:26 +00:00

grenade pushed to fix/295-ignore-tiny-prefill-samples at helexa/helexa

2026-08-27 07:45:26 +00:00

176ce47349 fix(context): ignore prefill samples too small to measure throughput

grenade commented on issue helexa/helexa#297

2026-08-27 07:36:08 +00:00

Document preserve_thinking for operators, and audit the ~31 other undocumented neuron toggles

Done in 00acd283. 0 of 35 keys undocumented, from 17.

What landed

Written as guidance rather than a generated reference. Each lever states what it controls, the default *and why that…

grenade pushed to main at helexa/helexa

2026-08-27 07:35:49 +00:00

00acd283b6 docs(operating): document every operator setting a neuron exposes

grenade commented on issue helexa/helexa#299

2026-08-27 07:20:25 +00:00

Reasoning blocks repeat themselves: the repetition penalty sees a 64-token window inside a 28k think block

Survey of the recorded sessions

Measured across every ~/.pi/agent/sessions/**/*.jsonl on the operator box: 14 sessions, 42 substantial thinking blocks, ~1.1M characters of reasoning,…

grenade pushed to main at helexa/helexa

2026-08-27 07:04:26 +00:00

eb783c345e Merge branch 'fix/300-anthropic-streaming-thinking'

72d6a37986 fix(anthropic): carry thinking on the streaming path too

Compare 2 commits »

grenade created branch fix/300-anthropic-streaming-thinking in helexa/helexa

2026-08-27 06:43:40 +00:00

grenade pushed to fix/300-anthropic-streaming-thinking at helexa/helexa

2026-08-27 06:43:40 +00:00

72d6a37986 fix(anthropic): carry thinking on the streaming path too

grenade pushed to main at helexa/helexa

2026-08-27 06:29:52 +00:00

039a353adf Merge branch 'fix/300-return-the-models-reasoning'

4e4ffd9b60 fix(reasoning): return the model's work to the client on every surface

Compare 2 commits »

grenade created branch fix/300-return-the-models-reasoning in helexa/helexa

2026-08-27 05:46:42 +00:00

grenade pushed to fix/300-return-the-models-reasoning at helexa/helexa

2026-08-27 05:46:42 +00:00

4e4ffd9b60 fix(reasoning): return the model's work to the client on every surface

grenade opened issue helexa/helexa#300

2026-08-27 05:37:00 +00:00

Three of five API surfaces silently discard the model's reasoning

grenade pushed to main at helexa/helexa

2026-08-26 18:09:44 +00:00

7ad9f48cad Merge branch 'feat/operator-repetition-knobs'

6141befa16 feat(sampling): let an operator pin the seed

e55998d224 feat(sampling): presence and frequency penalties, OpenAI semantics

c7df005d37 feat(sampling): let an operator reach the repetition knobs

Compare 4 commits »

grenade pushed to feat/operator-repetition-knobs at helexa/helexa

2026-08-26 17:40:03 +00:00

6141befa16 feat(sampling): let an operator pin the seed