fix: extract <think> blocks to m.reasoning + populate LLM Wiki Last writer (#1257)

[gsurenull]

Two related fixes for v0.51.222+.

1. Reasoning-only thinking-block extraction (commit 1)

Reasoning-only providers (e.g. MiniMax-M3) were leaving the thinking trace
inline in m.content, bloating persisted session files by 30-50% and
bypassing m.reasoning that the thinking card reads on reload.

• static/messages.js: new _splitThinkFromContent() handles <think>...,
  <|channel>thought\n...<channel|>, <|turn|>thinking\n...<turn|>
• api/config.py: register MiniMax-M3 in the minimax (international)
  catalog, bump _MODELS_CACHE_SCHEMA_VERSION 3 → 4
• Note: MiniMax-M3 is intentionally not added to minimax-cn — the
  China endpoint catalog is sourced from hermes-agent's catalog registry
  and does not yet expose M3

Refs #1852

2. LLM Wiki status panel Last writer (Issue #1257, commit 2)

The status panel was always rendering "Not available" for last_writer
because the field was reserved as None with no reader wired up.

• api/routes.py: new _llm_wiki_last_writer() with 3-tier fallback
  1. most-recent page frontmatter updated_by / writer / author
  2. most recent log.md action verb (returns "ai-agent (<action>)")
  3. static "ai-agent" fallback so the UI never shows "Not available"
     for a configured wiki
• tests/test_issue1257_llm_wiki_status.py: assert the new behavior
  instead of the locked "is None" baseline

Reads only page frontmatter and log.md headings, never page bodies, so
the private-safe status contract is preserved.

Closes #1257

[nesquena-hermes]

Two commits here — I read both at HEAD (7a8833e4 think-extraction, b90caa0f wiki last-writer) against origin/master, plus the agent's wiki skill to check the contract. The <think>-extraction half (commit 1) is sound: _splitThinkFromContent in static/messages.js reuses the existing _thinkPairs table (messages.js:716-720) and mirrors _parseStreamState's lstrip-then-startswith semantics (messages.js:1030-1056), so the live renderer and the persist path agree on what counts as a think block. Adding MiniMax-M3 to the minimax catalog with the _MODELS_CACHE_SCHEMA_VERSION 3→4 bump (api/config.py:1110, :2649) is the right way to force a cache rebuild. No notes there.

The wiki last-writer half (commit 2) has a real correctness bug. Let me ground it in the agent's actual wiki format.

Code reference

api/routes.py _llm_wiki_last_writer, Priority 2 branch:

for line in log_path.read_text(encoding="utf-8", errors="replace").splitlines():
    stripped = line.strip()
    if not stripped.startswith("## ["):
        continue
    if "|" not in stripped:
        continue
    tail = stripped.split("]", 1)[1].strip() if "]" in stripped else ""
    action = tail.split()[0] if tail else "update"
    return f"ai-agent ({action})"

This returns on the first ## [ heading. But log.md is append-only and chronological (oldest first). The agent's skill spells this out — skills/research/llm-wiki/SKILL.md:245 ("Append-only") and the template at :250 always seeds the file with:

## [YYYY-MM-DD] create | Wiki initialized

So for any wiki that has done more than initialize, this field — labeled "Last writer" in static/panels.js:3308 — will permanently render ai-agent (create), i.e. the oldest action, the literal opposite of "last". The test fixture (tests/test_issue1257_llm_wiki_status.py:24) only has a single ## [...] update heading, so it passes — but it doesn't exercise the multi-entry log that every real wiki has.

Fix sketch

Walk to the last matching heading instead of returning on the first:

last_action = None
for line in log_path.read_text(encoding="utf-8", errors="replace").splitlines():
    stripped = line.strip()
    if not stripped.startswith("## [") or "|" not in stripped:
        continue
    tail = stripped.split("]", 1)[1].strip() if "]" in stripped else ""
    last_action = tail.split()[0] if tail else "update"
if last_action:
    return f"ai-agent ({last_action})"

and add a fixture entry (a create line before the update line) so the test proves last-wins rather than first-wins.

Second, smaller note — Priority 1 is dead code

The frontmatter scan looks for updated_by / writer / author. I grepped the agent's wiki skill: the page frontmatter schema (SKILL.md:131-144) is title, created, updated, type, tags, sources, confidence, contested, contradictions — there is no author/writer key, and nothing in the skill ever writes one. So for any Hermes-authored wiki, Priority 1 never matches and every result falls through to the log.md branch (the one with the bug above). Not harmful, but it means the frontmatter path can't be the thing that saves you — the log.md scan has to be correct. Worth either dropping Priority 1 or aligning it with a key the agent actually emits (updated: is the closest real signal, though it's a date, not a writer).

Verification

After the fix: a fixture whose log.md has ## [2026-05-01] create | … then ## [2026-05-04] update | … should yield ai-agent (update), and adding a later ## [2026-05-06] ingest | … should yield ai-agent (ingest). That's the assertion that would have caught this.

[nesquena-hermes]

Shipped in v0.51.230 ✅ — both fixes cherry-picked onto fresh master.

1. <think> extraction → m.reasoning: evolved during review from client-side-only to client + server-side at save time — _split_thinking_from_content() in api/streaming.py runs on the final assistant message before s.save(), so the persisted session file (not just the browser copy) is compacted. Matches the live renderer exactly: leading-only single block, mid-body/code-fence tags stay as content, partial blocks left intact, existing on_reasoning preserved + merged.

2. LLM Wiki Last writer: now populated via the 3-tier frontmatter/log/static fallback (closes #1257).

Note: the original branch re-added MiniMax-M3 + bumped _MODELS_CACHE_SCHEMA_VERSION — both dropped as stale (MiniMax-M3 already shipped in v0.51.223). Gates: Opus + Codex SAFE, full suite 7422/0, browser smoke + QA harness clean. Thanks @gsurenull! Closing as merged-via-release.