Establish whether Claude Code's direct stream-json transport emits reasoning summaries, and under which setting

[Sandstorm831]

Question

Does Claude Code, driven the way Secant drives it today (claude -p with stream-json input and output), emit a provider-written reasoning summary during a Turn, and what launch flag, setting, or environment variable turns it on or off? Record the exact frames it arrives in, whether it streams, whether it carries a title, how it relates to redacted or encrypted thinking, and how Codex's app-server reasoning-summary deltas compare, so the Harness Seam's reasoning-summary event can be shaped from what both Harnesses really send.

[Sandstorm831]

Starting context

Input	Why it matters	Authority
ADR 0036 and its ADR 0022 note	A provider-written reasoning summary may cross the Seam and shows as a collapsed Thought: <title> · <duration> row; the raw chain of thought stays private	Canonical
Prototype the Run Workbench as an agent screen	Why this question exists and what the screen draws from the answer	Canonical decision
Study how T3 Code turns Claude Code and Codex events into a conversation ("Reasoning" and "Reasoning and ADR 0022")	T3 Code asks Claude for summarized thinking (showThoughts, thinking display summarized) through the Agent SDK and splits Codex raw and summary deltas; the starting lead, not proof for the direct CLI	Research
Establish Claude Code's viable structured transports	The direct claude -p stream-json transport Secant uses, which the answer must hold on; adopting the Agent SDK stays rejected	Research
src/harness/claude-code/frames.ts and src/harness/claude-code.ts	Which frames Secant parses today (only text and tool_use blocks and text_delta), and how it launches Claude Code	Canonical (current behaviour)
src/harness/codex.ts and src/harness/codex/required-schema.ts	Where Codex reasoning items and item/reasoning/* deltas are dropped or become generic activity today	Canonical (current behaviour)

Required methods

• /research, with findings under docs/research/ on a research/261-<name> branch

Reading rule

Read ADR 0036's reasoning-summary paragraph, then T3 Code's reasoning sections. Establish from primary sources, and a recorded run on the installed
Claude Code where you can, whether the direct stream-json transport emits summaries, under which flag, setting, or environment variable, and in which
frames. Record what Codex's app-server sends for comparison. Do not design the Seam event: that belongs to the ticket this one blocks.

[Sandstorm831]

Resolution

Yes: Claude Code's direct stream-json transport can emit provider-written reasoning summaries. On installed Claude Code 2.1.283 with claude-sonnet-4-6 on the first-party route, --settings containing showThinkingSummaries: true produced streamed summary chunks and completed thinking blocks. No Agent SDK or transport change was needed.

The full evidence, controls, sanitized frames, run table, source citations, and qualifications live in Claude Code reasoning summaries over direct stream-json, preserved by the merge-commit PR Establish Claude Code reasoning-summary transport.

• Controls. The documented showThinkingSummaries setting enables readable summaries. False or unset produced empty reasoning bodies with thinking_display: "updates" in this build; short progress updates between tools may still appear. Hidden --thinking-display summarized overrode the false setting in a direct probe. Hidden omitted is accepted by the installed parser but was not separately live-probed. No documented dedicated summary environment variable was established. MAX_THINKING_TOKENS=0 removed thinking blocks on Sonnet 4.6, which is thinking-off rather than summary-off and is not universal across models.
• Frames and streaming. With --include-partial-messages, the CLI wraps message_start, indexed content_block_start, content_block_delta/thinking_delta, signature_delta, and content_block_stop in stream_event frames. An assistant frame carries the completed thinking block; it can arrive before the forwarded block-stop and share its native message id with a later text-only assistant frame. Without partial messages the complete summary still arrives. The terminal result carries the final answer, not a separate reasoning summary. Summary deltas and their completed body represent the same content.
• Titles and private material. Normal summaries have no separate title or duration. The hidden internal highlights mode fell back to omitted in the probe and supplied no title, so dependable provider titles were not established. Opaque signature and safety-redacted redacted_thinking.data are encrypted material, not display text. The summary distinction is established for the probed Claude 4.6 model, not arbitrary historical thinking frames, custom aliases, or gateways.
• Codex comparison. Official documentation and locally generated codex-cli 0.155.0 schemas separate item/reasoning/summaryTextDelta and summaryPartAdded (identified by item and summary index) from raw item/reasoning/textDelta. Final reasoning items separate summary[] from content[]; neither schema supplies a title or duration. turn/start.summary and model_reasoning_summary control summaries. This is schema/documentation evidence, not a live Codex reasoning run.

Seven isolated no-tools Claude CLI runs, report formatting, and whitespace checks passed. The PR CI run passed the canonical gate, process runtime conformance, build, and release-consumer checks on all three operating systems. An earlier branch run exceeded the existing checkpoint test's 30-second timeout on Windows; the fresh PR run passed it without a source change. Hidden controls remain version-qualified, and no claim is made for untested providers or all models.

The research resolves transport facts only. Event shape, title fallback, timing, and launch policy remain with Decide the Harness Seam's tool, command, diff, and reasoning-summary events. No production behavior or ADR changed.