The Timeout Was Honest. The Active State Was Not.
When an adapter exposes tool items only at completion, a watchdog can see zero active work during a real upstream continuation. Fix the missing lifecycle fact first.
distributed ML runtime systems · agent ops · self-hosted AI
Engineering notes from systems that have to work
Deep dives on distributed execution, backend/runtime reliability, self-hosted AI agents, debugging, and evidence-driven engineering workflows.
Start with a subject
Browse by topic
14 posts
OpenClaw & Self-Hosting
Operating, upgrading, and debugging self-hosted OpenClaw and adjacent agent tooling.
39 posts
Agent Architecture & Operations
Reusable patterns for agent orchestration, handoffs, panels, memory, and operational contracts.
22 posts
Reliability & Security
Incidents, recovery, authorization boundaries, supply-chain risk, and fail-closed engineering.
10 posts
Developer Workflow & Open Source
PR evidence, review discipline, GitHub workflows, and public-safe engineering collaboration.
8 posts
Distributed Runtime Systems
Distributed execution, model serving, performance, capacity, and runtime debugging.
6 posts
Personal Automation & Tutorials
Practical setup guides, personal automation, migrations, and lessons from building in public.
Newest by publication date
Latest publications
A Cheaper Agent Is Not a Cheaper Workflow: Delegate by Workload Class
Cheap agent delegation should be promoted by workload class, not model label. Tiny-task savings earned shadow expansion; a larger failure locked its class.
Yielded Is Not Finished: Guarding Partial-Clone Git Scans Without Breaking Git
A yielded Git command can remain alive and materialize missing history. The proportionate fix is narrow classification, explicit process ownership, and independent non-execution proof.
Archive Before You Retire: A Transaction Model for Git Worktree Backlogs
Stale worktrees are not safe to delete just because they look old. Archive-preserve transactions make recovery requirements and protected-tip drift checks explicit.
High Draft Acceptance Does Not Guarantee a Throughput Gain: A Lean CPU Speculative-Decoding Test
Across two output lengths, every tested draft depth accepted many proposed tokens but still lost on both runtime-reported generation throughput and request wall time.
Approval Is Not Execution: Deferred Side Effects Without a Second Control Plane
A deferred-effects manifest can preserve proposed intent, suspend dependency-bound work, and compile a review bundle while native action approval remains the only approval authority for the external effect.
When Reasoning Eats the Answer: Debugging Empty LLM Completions
Empty final content with finish_reason=length can be consistent with reasoning exhausting a shared completion budget. A paired-cap test helps separate that case from model and runtime failure.
A Default Is Not Caller Intent: Preserve Provenance in Agent Settings
A resolved value is not enough. Preserve request provenance, selection authority, runtime-reported evidence, constraints, and rejection reasons across adapters.
The Harness Passed. The Claim Did Not: Why Independent Evidence Matters
A green harness is not enough when its “observation” merely repeats the input. Runtime claims need an independent evidence path and a real stop rule.
Meaningful revisions
Recently updated
A Default Is Not Caller Intent: Preserve Provenance in Agent SettingsUpdated August 24, 2026
A Visible Message Does Not Prove an Agent WakeUpdated August 24, 2026
A Local LLM Router Is Not a Panel Lane YetUpdated August 23, 2026
LLM Panel Orchestration in OpenClaw: Config-Backed Routing, Timeout Classes, and Honest Dissent Without ChaosUpdated August 23, 2026
Built for readers, crawlers, and agents
Every post is readable as static HTML with a stable URL. The site also publishes RSS, a sitemap, Markdown mirrors, Pagefind search, and an llms.txt index.