Bound wiki-agent context use: output caps + distill-then-act for oversized threads

[sysread]

SYNOPSIS

Fixes the 2026-07-23 wiki-sweep skips ("maximum context length is 163840 tokens"). Caps agent output, distills oversized transcripts, revives the dead content-filter fallback.

PURPOSE

The wiki sweep skipped two long threads with a context-length 400. Diagnosis found the cause was not the model's window size: the headless tool loop never set max_completion_tokens, so the serving backend reserved 65536 output tokens out of the window for an agent that emits tool calls plus a two-sentence summary. Two aggravators: thread slices go to the model unbounded, against a ceiling that moves between backends serving the same model id (163840 enforced Jul 23; 1M-scale accepted a week later - the registry's contextWindow is not a contract); and the content-filter fallback model (arcee-trinity-large-thinking) no longer exists in Venice's catalog, so that path could never succeed.

DESCRIPTION

Today, runHeadlessAgent forwards no output cap, runWikiAgentOnThread / runExtractionOnThread push the whole thread slice plus the agent prompt into one conversation, every failure burns the 3-strike transient counter (even deterministic ones), and a classifier rejection retries against a vanished model id.

This PR changes each point in the same order:

• runHeadlessAgent gains maxTokens; the wiki fleet (article, records, librarian) caps every round at 8192, reclaiming ~57k tokens of input room.
• New agents/_accumulator.ts (port of fnord's accumulator): transcripts estimated over a pinned 96k working window are rendered to text, chunked, and folded into an accumulated notes buffer by read-only distill completions; the normal tool loop then runs once over the notes, so all writes keep the full toolbox and dedup discipline. A context-length 400 on a direct run retries distilled; a 400 during distillation backs off the chunk budget (0.2 steps, 0.6 floor).
• Context-length errors that survive distillation are flagged deterministic and the sweep skips the thread on the first failure (p_max_failures = 1) with an honest reason.
• CONTENT_FILTER_FALLBACK_MODEL re-pointed to venice-uncensored-1-2 (verified live: supports function calling; non-reasoning, so the fallback attempt keeps reasoning_effort off the wire).

Net effect: both skipped threads would have processed even on the 163840 backend (output cap alone suffices), and the distill path is the backstop for the genuinely huge tail - 410 of 412 threads in the live DB run byte-identical to before.

Notes:

• Intentional behavioral change: deterministic context-length failures now skip on the first strike instead of the third - the retry would fail identically, and the skip reason is surfaced in the Skipped panel either way.
• The distill-with-attempt-model choice means the rare oversized-AND-classifier-rejected thread re-distills on the fallback attempt; accepted cost, commented at the call site.
• reflection, summary, samskara_evaluation, and thread_topics still consume unbounded slices; the accumulator is shared and ready, wiring them is deliberate follow-up scope.
• Not verified end-to-end against a live oversized thread; the completion seam covers the logic (10 new Deno tests). Retrying the two Skipped-panel rows after merge is the cleanest live test.

---

Generated by Claude Code