Skip to content

v0.10.0 — Wiki pipeline hardening: terminating writer loop, self-healing jobs, editable headers

Choose a tag to compare

@dimknaf dimknaf released this 12 Sep 10:57
· 31 commits to main since this release
99cd121

Headline

The wiki pipeline is now dependable under long, unattended runs. The writer loop always terminates, crashed and restarted jobs recover on their own, and the page header is editable instead of frozen after the first write. Also in this release: a reasoning-effort knob for the wiki agents that cuts latency several-fold on self-hosted models, a per-call LLM timeout so slow local writes are no longer abandoned mid-flight, and both agents can now see how big a page is before deciding what to do with it.

Added

  • Append mode and paged reads for the wiki section tools — the model picks the edit that fits the content, not the one the tool allows.
  • check_members_cited — one shared "is this cited yet?" predicate for the writer tool and the router gate, replacing three copies that had drifted.
  • The page header is an editable section; it no longer freezes once a page grows past the inline-body limit.
  • AGENT_WIKI_REASONING_EFFORT — wiki agents only, blank by default, self-hosted vLLM only.
  • AGENT_REQUEST_TIMEOUT — per-call transport deadline, default 4800s.
  • Page sizes visible to the maintainer; neighbouring page names and sizes to the writer.

Changed

  • Writer handoff budget 20000 -> 30000. A budget set where it cannot fire silently disables the successor path.
  • Job lease 20 -> 120 min with a bounded reclaim ceiling — a slow write is no longer mistaken for an abandoned one.
  • vllm_workstation_qwen points at the Qwen model and port the wiki tuning constants were measured against. deepinfra remains the default profile.

Fixed

  • The writer loop terminates. Long runs were spending most of their time re-running a job whose work was already complete.
  • Self-healing restored — jobs past their lease re-enter triage instead of wedging, so an entity cannot be silently lost.
  • Crashes and restarts are recoverable: snapshots taken at claim time, orphaned jobs requeued on startup.
  • The token estimate counts tool-result payloads, so the handoff nudge can actually fire.
  • Per-call LLM timeout stops slow self-hosted writes being abandoned while the server is still completing them.
  • Duplicate page creation within the same window collapses to one.
  • update_entity can no longer overwrite a wiki body, and a blank body is a warned no-op rather than a silent wipe.