Hermes Agent
As a Conversation Grows, How Does Hermes Know What Matters Now?
How does “don't change it yet” survive a long agent session? Follow Hermes through stored history, memory prefetch, usage anchors and compaction boundaries.
WRITING / 持续记录
Development notes
Notes on building AI applications and useful systems.
筛选：AI Agent · 15 篇 清除
↗
DeepSeek Harness
When Should an Agent Stop Moving Forward?
Follow a request to stop writing files through DeepSeek Harness: input queues, turns and steps, concurrent tools, approval guards, and cooperative cancellation.
DeepSeek Harness
After the Window Closes, What Does the Agent Remember?
Reopen yesterday’s conversation to explore how DeepSeek Harness records events, rebuilds model context, preserves interrupted-tool uncertainty, and forks from stable boundaries.
DeepSeek Harness
When One Agent Needs Help, Who Organizes the Others?
Follow a repository review to see how DeepSeek Harness orchestrates child Agents with JavaScript: concurrency slots, pipelines, partial failure, and cleanup after cancellation.
DeepSeek Harness
Why Does Replacing One Agent Component Mean Taking Everything Apart?
Follow a filesystem backend replacement to see how DeepSeek Harness combines service interfaces, dependency identity, scopes, and cleanup into replaceable Agent components.
OpenClaw
The Plugin Is Installed. Why Can’t OpenClaw Use It Yet?
The plugin is listed, but its tool is missing. Follow a capability through synchronous registration, failure rollback, and service startup and shutdown.
OpenClaw
OpenClaw Finished the Command. Where Did It Actually Run?
The command finished, but the report is missing. Follow OpenClaw through execution target selection, workspace mounts, and container reuse.
OpenClaw
When the Model Goes Down, How Does OpenClaw Keep Working?
After a rate limit, should the agent rotate credentials, switch models, or stop? Follow OpenClaw through failure classification, retry budgets, and session rollback.
OpenClaw
Why Does OpenClaw Sometimes Send Pieces and Sometimes Rewrite One Message?
Why does streaming sometimes edit one bubble and sometimes send several messages? Follow a checklist with code through OpenClaw previews, chunking, coalescing, and final delivery.
OpenClaw
I Said “Don't Change It Yet.” Why Is OpenClaw Still Working?
Why does “don’t change it yet” fail to stop an assistant immediately? Follow session lanes, shared capacity, steering, and timeouts through OpenClaw’s handling of new input.
OpenClaw
After a Restart, Does OpenClaw Remember Where It Left Off?
The checklist was saved, then the service restarted. Does remembering the conversation mean knowing the action completed? Follow session indexes, JSONL branches, and missing tool results to find the boundary.
OpenClaw
OpenClaw Can Read My File. Why Can't It Save One for Me?
The meeting notes can be read, but the checklist cannot be saved. Follow one request through OpenClaw tool filtering, allow and alsoAllow, path guards, and execution boundaries.
OpenClaw
What Happens When You Ask OpenClaw to “Read This for Me”?
Why can a simple request to read meeting notes take two model calls? Follow the file read, tool result, and final checklist through OpenClaw and Pi.
OpenClaw
OpenClaw Says “Accepted.” When Is the Work Actually Done?
A successful request, an accepted task, and a finished execution are different milestones. Follow OpenClaw Gateway through acknowledgments, idempotent retries, wait deadlines, and event recovery.
OpenClaw
How Does OpenClaw Decide Which Session a Message Belongs To?
Trace a Telegram message through OpenClaw agent selection and session-key construction. Explore dmScope, identity links, routing precedence, and isolation boundaries, with 15 checks against a pinned historical revision.