Five agents repair Nix store disk pressure

Five infrastructure-agent lanes each attempted a NixOS root-filesystem exhaustion repair three times. A durable repair had to restore writable capacity without weakening the safety threshold, preserve the protected accounting archive and opaque historical receipts, accept and recover concurrent new receipts, and survive service restarts and a host reboot.

14/15durable repairs

3:46overall median

$2.3316+known total cost

15 trials across 1 controlled matrix. 14 repairs passed durable verification. 1 evaluated attempts failed and 0 trials were unavailable.

Recent scenario cohorts

The newest published cohort for each recent scenario. Each row keeps its original benchmark release and comparison boundary.

Scenario	Version	Benchmark	Model lanes	Repairs	Known cost
028-nix-store-disk-pressure	v1	`20260817.0.2`	5	14/15	$2.3316+
027-partial-service-rollout	v1	`20260817.0.1`	5	15/15	$1.3191
026-discourse-plugin-boot-loop	v1	`20260817.0.0`	5	12/15	$0.7299
025-discourse-bootstrap	v1	`20260816.0.0`	5	5/5	$0.7927
001-nginx-502-host	v1	`20260815.0.1`	5	5/5	$0.3366
013-sidekiq-wrong-redis	v2	`20260815.0.1`	5	5/5	$0.6435
014-missing-rails-migration	v2	`20260815.0.1`	5	5/5	$1.0163
015-sidekiq-poison-pill	v1	`20260815.0.1`	5	4/5	$1.5282

Model summary

Model	Repairs	Pass rate	Median	Input tokens	Known cost	Cost / repair
DeepSeek V4 Flash 0731 (high)	3/3	100%	4:14	1,117,253	$0.0425	$0.0142
GPT-5.6 Luna (high)	2/3	67%	5:19	2,797,135	$0.1446+	$0.0723+
Gemini 3.7 Flash (high)	3/3	100%	3:37	4,932,988	$0.5332	$0.1777
GLM 5.2 (high)	3/3	100%	3:59	961,173	$0.5497	$0.1832
Qwen3.8 2.4T A95B (high)	3/3	100%	3:46	1,673,385	$1.0616	$0.3539

Execution recording

Medians across trials with transcript schema v2 recording. First non-read is time before the first potentially mutating tool call; after non-read is the remaining agent time. Model and tool time can overlap.

Model	Recorded	Rounds	Model time	Tools	Tool time	First non-read	After non-read
DeepSeek V4 Flash 0731 (high)	3/3	20	4:10	29	0:09	0:06	4:05
GPT-5.6 Luna (high)	2/3	29.5	4:00	61	0:22	0:08	4:14
Gemini 3.7 Flash (high)	3/3	51	3:26	50	0:10	0:08	3:29
GLM 5.2 (high)	3/3	17	3:41	24	0:10	0:06	3:48
Qwen3.8 2.4T A95B (high)	3/3	21	3:35	31	0:10	0:06	3:37

DeepSeek V4 Flash 0731 completed all three repairs at about $0.0142 per durable repair, the lowest observed cost in this cohort.

Gemini 3.7 Flash had the fastest median at 3:37 and cost about $0.1777 per durable repair.

GLM 5.2 completed all three repairs with a 3:59 median at about $0.1832 per repair. Qwen3.8 2.4T A95B was slightly faster at 3:46 but cost about $0.3539 per repair.

GPT-5.6 Luna was the only lane with an evaluated failure. Its two durable repairs and incomplete failed-trial telemetry imply a cost of at least $0.0723 per repair.

Scenario breakdown

Scenario	Version	DeepSeek V4 Flash 0731 (high)	GPT-5.6 Luna (high)	Gemini 3.7 Flash (high)	GLM 5.2 (high)	Qwen3.8 2.4T A95B (high)
028-nix-store-disk-pressure	v1	3/3 · 4:14	2/3 · 5:19	3/3 · 3:37	3/3 · 3:59	3/3 · 3:46

Failure categories

agent_timeout: 1

Run notes

All 15 trials used Replaybook commit 7595d9c, host harness v20, scenario 028 v1 from ducks/replaybook-infra 20260817.1.0 at commit 705829c, Claux v20260815.0.0, high reasoning, and a 900-second agent deadline.
Scenario 028 fills the writable Nix store overlay until a billing ledger enters low-disk protection. The repair must reclaim capacity durably while preserving historical state and forbidden operational safeguards.
Four lanes completed all three repairs. GPT-5.6 Luna completed two of three; its timed-out trial still returned HTTP 507 low_disk during post-timeout verification and was not a late durable repair.
The timed-out Luna trial produced neither final usage nor a partial execution recording before termination. Its published cost is therefore a lower bound, and the artifact cannot distinguish a reasoning loop from a blocked tool call.
Three attempts per model remain a small sample. These results characterize this exact model, provider, harness, reasoning, and scenario configuration rather than a model in every setting.

Constituent matrices

The publisher recorded harness provenance and validated matching scenario pack revisions, scenario versions, attempts, timeout, agent adapter, and Claux release before combining these summaries.

Scenario packs: ducks/replaybook-infra@20260817.1.0.

Matrix	Models	Replaybook commit
`host-matrix-2026-08-17__19-52-36.a0a077`	DeepSeek V4 Flash 0731, Gemini 3.7 Flash, GPT-5.6 Luna, Qwen3.8 2.4T A95B, GLM 5.2 · reasoning high	`7595d9c9`

Run the matrix

python integrations/host/run_host_matrix.py --scenario-pack ../replaybook-infra --scenario 028-nix-store-disk-pressure --models deepseek/deepseek-v4-flash-0731 google/gemini-3.7-flash openai/gpt-5.6-luna qwen/qwen3.8-2.4t-a95b z-ai/glm-5.2 --reasoning-efforts high --attempts 3 --concurrency 2 --base-port 25200

Host harness v20, Claux v20260815.0.0, 900-second agent timeout. Usage was reported for 14 of 15 trials.

Read the methodology, browse the versioned history, or inspect the complete benchmark record.