…t render

The previous commit opted Discord into non-stream reasoning unconditionally
(streamReasoningInNonStreamModes: true). With progress.thinking off (default),
the onReasoningStream callback still fired in plain mode and ran side effects —
statusReactions.setThinking() and windowReasoningOpen — before the compositor's
render gate, so a thinking status reaction leaked despite the config being off.
(The compositor render + 🧠 counter were already gated; the reaction was not.)

Make channels.discord.streaming.progress.thinking the single source of truth:
the draft-preview controller now exposes reasoningProgressEnabled (the resolved
flag it already used for the compositor reasoningGate), and the Discord handler
passes it as streamReasoningInNonStreamModes. In plain mode with the config off
the callback no longer fires at all; stream mode keeps main's behavior.

Adds resolver coverage asserting progress.thinking defaults off (unset, explicit
false, missing entry, and non-progress stream mode) and is true only when
explicitly enabled in progress mode.

Also the first time the Discord extension typecheck (against built dist) was run
for this change set; it passes.

Addresses the ClawSweeper [P1] 'gate the Discord reasoning opt-in before side
effects' finding.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…c placeholder

extractAssistantThinking substituted a diagnostic string ("Native reasoning was
produced; no summary text was returned.") onto the reasoning lane whenever a
thinking block carried a signature but no summary text. That fires on GPT
constantly via the codex harness and about half the time via the openclaw
harness, and on Anthropic when display:"summarized" still returns empty —
surfacing a placeholder bubble on every channel. It is diagnostic, not reasoning
content; return "" so the downstream .filter(Boolean) drops the bubble. The
signed block stays on the message for API replay; this only governs display.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…alize

Anthropic has no native commentary channel; the only signal that leading text is
pre-tool narration is that a tool_use block follows it. The finalize loop tagged
it only once stop_reason:tool_use landed at message end — after the unphased text
already streamed. Tag it the moment the first tool_use content block opens, so
the commentary phase rides that partial and the subscriber can route it to the
narration lane far earlier. Finalize stays as an idempotent backstop for
replay/non-streaming assembly.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

…is known

On block-reply surfaces (onBlockReply + blockReplyBreak "text_end" — the default)
the subscriber durably sends a text block at its text_end. For a tool turn with
leading narration, that fired before the commentary tag was applied at the next
event (tool_use content_block_start), leaking pre-tool narration as a final reply
that the later tag could not retract. Progress mode was unaffected (no mid-stream
durable send), which is why it looked clean.

Hold each leading text block's text_end in the transport until the disposition is
known: flush it right after tagPendingCommentaryText at the tool_use boundary (so
it carries the commentary phase and the subscriber suppresses it from the durable
reply), or at message_delta / finalize for a non-tool turn (where the text is the
answer and is emitted unphased). Cleared on error/abort. Text deltas still stream
live, so the preview is unchanged; only the text_end (and thus the durable commit)
waits one event. Subscriber and all channels are untouched — they now simply see a
correctly-phased text_end, so every streaming-capable channel inherits the fix.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

The transport now defers a pre-tool text block's text_end until it is tagged
commentary, but the earlier text_delta chunks were still appended to the
block-reply buffer (handleMessageUpdate). On block-reply channels
(blockReplyBreak "text_end" — Feishu/SMS/IRC and other append-style surfaces)
that buffer is drained mid-stream by the chunker and flushed before tool
execution (handleToolExecutionStart), so the narration could still post as a
durable answer that the later commentary tag cannot retract.

Gate the durable block-reply append for phase-pending Anthropic text: while an
anthropic-messages assistant block is unphased mid-stream (no deliveryPhase, no
streamItemId), do not append it to the block-reply buffer. A tool turn's
deferred commentary-tagged text_end then suppresses it; a non-tool answer is
delivered in full from handleMessageEnd (lastBlockReplyText stays null, so the
message_end safety-send fires). Unlike the OpenAI phase-pending guard this does
not early-return — the live preview keeps streaming; only the durable append is
withheld. Discord progress mode and other providers are unaffected.

Unit tests reproduce the leak (narration posted without the gate) and prove the
non-tool answer is still delivered in full.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

… payloads

Mirror json's reasoningPayloadsEnabled seam (openclaw#97875) for the 💬 commentary
lane: add GetReplyOptions.commentaryPayloadsEnabled and an isCommentary payload
tag next to isReasoning across the payload types; default-suppress isCommentary
in dispatch-from-config (block stream + final loops) unless the channel opts in;
exclude commentary from TTS exactly like reasoning; thread the tag through the
block-reply coalescer, reply-delivery direct-lane, and realtime consult so it
cannot be dropped mid-pipeline. Discord opts in when its commentary lane is
active. No durable producer is wired yet (preserves the obviyus-approved 3-part
Anthropic withhold invariant; handlers.messages.ts unchanged).

Also regenerates docs/.generated/config-baseline.sha256 and bumps the plugin-SDK
surface budgets to actual post-openclaw#97875-rebase values.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…unts only streamed items

Reasoning durable/window selection now follows the session /reasoning level
(off|on|stream) instead of the streaming.progress.thinking config flag:
on -> durable blockquote, stream -> live window lane, off -> nothing.
reasoningPayloadsEnabled is opted in only for 'on'; the window opt-in only
for 'stream'. Collapse-bar counters increment only when a line is actually
streamed to the window (post-yield), so durable 🧠/💬 lanes are never counted.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Anthropic commentary has no native item id, so the transport re-emits the full
commentary text under an empty itemId on every later stream event. The verbose
standalone commentary lane dedups only on a defined matching itemId, so each
snapshot posted a duplicate persistent message (observed 8x). Skip re-emitting
an unchanged commentary snapshot per item key.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…italic

The progress-draft compositor wraps reasoning and commentary window lines in
_..._ for italics. Telegram renders that as italics; Discord shows a literal
underscore on commentary. Add a commentaryItalics compositor flag (default on)
and opt Discord out: streamed thinking stays italic, commentary renders plain.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…in live transport (opus-4-8 was sent no display → blank summary); mirror llm/providers/anthropic.ts

(cherry picked from commit 378a06da3a6061a867213148e7f0de7d216b0618)

…rs (pioneer)

Restores 363f544c4e (dropped in the min-rebase), re-applied onto the
tag-at-tool-boundary structure:
- stream-resolution: route anthropic-messages models with a non-"anthropic"
  provider onto the boundary-aware managed transport (the only place commentary
  gets tagged); the base SDK stream never tags it. Scoped to provider!=anthropic
  so direct-anthropic paths are unchanged.
- transport backstops: tag when the turn CONTAINS a toolCall, not only when
  stop_reason maps to toolUse — Bedrock/Vertex routes report end_turn on tool
  turns. No-op for direct Anthropic (already toolUse when a tool call is present).
- regression test for the stop_reason-mislabel case.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Aligns the two assertions with the commentaryItalics:false opt-out; reasoning
window lines stay italic.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Restores e0c838314b (dropped in the min-rebase), adapted to the count-on-stream
bar: when /reasoning on posts durable 🧠 messages, they land after the early
progress draft, so editing the draft in place strands the summary at the top
above the thoughts. Track persistentReasoningDelivered and, when set, clear the
early draft and post the summary fresh below the thoughts (thoughts → summary →
answer). Durable 🧠 stays uncounted in the bar (no progressReasoningSteps bump).

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

… non-work tools out of bar

[P2] reasoning-level default now resolves the agent's own reasoningDefault first
(mirrors core get-reply-directives + Telegram), then global, then off — an
agent-scoped default no longer runs reasoning on while Discord drops the durable
🧠 lane.

[P3] exclude non-work tools (message/react/typing) from the collapse-bar tool
count via the canonical isChannelProgressDraftWorkToolName predicate, so they no
longer inflate the bar. (The earlier render-result approach was wrong — the
compositor returns false for stored-but-deferred lines, which undercounts.)

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

…surface

Reasoning display is gated by the session /reasoning level after the rework, so the
streaming.progress.thinking config key/resolver/docs were a no-op (ClawSweeper P2).
Remove the key, type, zod, generated metadata, resolver, tests, and docs mention;
regenerate the docs map. progress.commentary and session-level reasoning unchanged.

Re-pin the plugin-sdk public-surface budgets (10406->10401 / 5224->5220) to match
the source counts after removing the exported resolver.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>