Wiki + librarian workers: emit operator-facing reasoning to the log drawer

[sysread]

SYNOPSIS

The wiki per-conversation worker and the wiki librarian now end each cycle with a one-or-two-sentence operator summary of WHY they edited / merged / skipped, surfaced as reasoning="..." on the log-drawer line. The journal worker already did this; this PR lifts the other two to parity.

PURPOSE

Currently the wiki and librarian cycle logs show only a tool-call count - "finished thread X (3 tool calls over 25 messages)", "librarian finished (5 tool calls over 87 articles)". A user skimming the log drawer cannot tell:

• whether a no-op cycle was correct (generic Q&A, no user-centric subject) or a silent regression,
• which articles the librarian considered for merge but deliberately left separate, and why,
• whether the per-conversation agent linked out to Wikipedia instead of creating a sterile article.

The journal worker already solves this via its structured {worthy, reasoning, entry} output, which the loop inlines as reasoning="...". The wiki and librarian are tool-driven (no JSON output), so historically their prompts closed with "reply with a single word; the word is discarded" purely to terminate the tool loop. The final-text channel was free; the prompt just wasn't asking for anything useful on it.

DESCRIPTION

Layer 1 - how existing code behaves

• Journal worker parses {worthy, reasoning, entry} from a JSON completion. JournalOutput.reasoning is set from that field; the loop logs (wrote=true/false, ..., reasoning="...") at info tier.
• Wiki per-conversation worker runs runHeadlessToolLoop and returns WikiOutput.finalText - the model's final assistant turn after the tool loop terminates. The prompt's closing block says "reply with a single word; the word is discarded". The loop logs (N tool calls over M messages) with no reasoning= field; finalText is captured into the output struct but never surfaced.
• Wiki librarian is the same shape: runHeadlessToolLoop, WikiLibrarianOutput.finalText, "reply with a single word", loop logs (N tool calls over M articles).

Layer 2 - what this PR changes

• Wiki autonomous prompt (src/lib/agents/wiki/prompt.ts) - replaces the "single word" closing block with a "Final reply: one or two sentences explaining your choices" block. Three worked examples cover update, no-op-because-generic-Q&A, and no-op-because-not-user-centric. Asks for plain text, no Markdown, under two sentences, and explicitly tells the model the text surfaces in the user's log drawer.
• Wiki cycle driver (src/lib/agents/wiki/loop.ts) - whitespace-normalises runResult.output.finalText (collapses \s+ to a single space, trims), falls back to (none) on empty, and inlines it as reasoning="..." on the existing finished thread ... info log. Matches the journal worker's format exactly so grep / scrollback search lines up across all three.
• Librarian prompt (src/lib/agents/wiki-librarian/prompt.ts) - same change. Examples deliberately call out the "considered for merge but left separate" case, which is the user's stated motivating example - that decision is currently invisible in the logs.
• Librarian cycle driver (src/lib/agents/wiki-librarian/loop.ts) - same reasoning="..." surface on the librarian finished ... line.
• Type docs (src/lib/agents/wiki/types.ts, src/lib/agents/wiki-librarian/types.ts) - finalText doc comments no longer call the field "discarded by production callers"; they describe the operator-facing role and reference the prompt's "Final reply" block.
• Agent file headers (src/lib/agents/wiki/agent.ts, src/lib/agents/wiki-librarian/agent.ts) - updated to match.
• Dev doc (docs/dev/wiki.md) - the "Entry points" section's WikiAgent.run paragraph now describes the operator-summary contract instead of calling final-text discarded. A new Gotcha entry notes that the prompt's "Final reply" instruction and the loop's reasoning="..." surface move together - reverting one without the other silently drops a user-facing diagnostic.

The journal worker is not touched - it already does this.

Notes for AI reviewers:

• The \s+ -> ' ' whitespace normalisation in both loops is deliberate. Models occasionally emit a multi-line final turn even when asked for one or two sentences; the single-line log convention is load-bearing and a stray newline would break the (field=value, field=value) "title" shape the journal worker established.
• The || '(none)' sentinel is also deliberate. An empty finalText shouldn't happen in production (the prompt now mandates a sentence), but a dangling reasoning="" reads as a logger bug; the sentinel makes a missing summary self-explanatory.
• The new "Final reply" prompt blocks are additive - they sit after the existing scope / preserve-facts / be-conservative discipline blocks. The "single word" closing instruction is removed in the same edit; leaving both would tell the model conflicting things.

Test plan

• mise run check green: 1069 tests passed, build clean (the pre-existing 500kB chunk warning is unchanged).
• No tests reference WikiOutput.finalText or WikiLibrarianOutput.finalText directly - the field is unchanged in shape, just newly load-bearing for log output.
• Manual smoke: after merge, watch the log drawer during the next wiki cycle and the next librarian cycle (12h cadence by default). Look for reasoning="..." on the finished thread / librarian finished lines. A short, decision-focused sentence is the expected shape; if the model emits filler ("Great work!") or stays under one sentence, the prompt may need tightening.

---

Generated by Claude Code