fix(tui): collapse encrypted reasoning runs into one updating line

[brentyates-swx]

Issue for this PR

Refs #53. The Luna session review in that issue stays open; this PR covers only the TUI rendering half.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

OpenAI Responses models (GPT-6 Luna and Sol through swx-azure) return encrypted reasoning. Each item arrives as a reasoning part with no title or body, and the TUI printed one Thought · <duration> line per part. A single turn could print dozens of these lines and push the useful output off screen.

Each run of consecutive encrypted parts now renders as one line with the count and the total duration, and the line updates as each part arrives:

Thought 3 · 6.0s

• groupOpaqueReasoning (new packages/tui/src/routes/session/opaque-reasoning.ts) is a pure function over one assistant message's parts. It returns the rows to render and the members of each run. A run's row is its first part, the same object on every update, so the run's view persists while the run grows.
• The run row is OpaqueReasoningRunView. It keeps one label node for the whole lifecycle, so a selection on the label survives the spinner appearing and going away. The spinner is a sibling that shows only while a part is in flight.
• A tool call, visible text, or visible non-encrypted reasoning ends a run. A message error ends a run too, because it renders after the message's parts.
• Parts that render no line (step markers, blank text, empty reasoning) stay in order and render nothing. They do not end a run.
• While a part is in flight, the line keeps the spinner and its count includes that part. The duration sums the finished parts.
• Non-encrypted reasoning renders as before. The done, unresolved, and duration rules sit in one helper, reasoningState, that the run row and single parts share. ReasoningHeader no longer has an encrypted branch, because nothing reaches it.
• AssistantMessage is renamed AssistantMessageView and exported for the render and activity tests. The old name clashed with the type-only SDK import once it was exported. The render test and the assistant activity test import it.

Not in this PR. Each item is tracked:

• The transcript export still writes an empty _Thinking:_ heading per encrypted part: tui: transcript export writes an empty _Thinking:_ heading for each encrypted reasoning part #160.
• Other rows still replace their label node when the spinner state changes: tui: spinner-or-label Switch replaces the label node, so a selection on it is lost #168.
• A part type with no renderer disappears silently from the assistant message: tui: a part type with no renderer disappears silently from an assistant message #169.
• A run never spans two assistant messages: tui: group visible opaque reasoning across messages and hidden tools #170.
• The encrypted-reasoning tests build their parts by hand rather than from producer output: tui: test encrypted reasoning from a producer-shaped fixture, not hand-built parts #171.
• The run row's label moves two columns when the spinner appears: tui: the run row's label moves two columns when the spinner appears #172.
• Branch names: AGENTS.md bans slashes, but firstmate's fm/ prefix breaks that rule: AGENTS.md bans slashes in branch names, but firstmate branches are fm/<task> #174.
• index.tsx already has prettier drift and 16 oxlint warnings on swxtch, which this PR leaves alone: tui: session route index.tsx has pre-existing prettier drift and 16 oxlint warnings #177.

How did you verify your code works?

• test/routes/session/opaque-reasoning.test.ts: 12 pure tests. They cover a run, a lone part and its identity, tool and text breaks, a failed tool call, parts that render no line, a part type the view does not render, mixed encrypted and plain reasoning, the [REDACTED] placeholder, and the predicate.
• test/routes/session/opaque-reasoning-render.test.tsx: 6 tests that render AssistantMessageView through the sync store. They cover a loaded run, parts arriving live, tool and text breaks, plain reasoning between runs, a live run whose label and selection survive a third part going in flight and then completing, and a message error.
• The lifecycle regression fails before the label fix. At the in-flight check, the label is a new node. It passes with the fix.
• Reverting each change turns the matching tests red: dropping transparent parts from the rows, copying a run's first part on every update, and reading run members once instead of reactively.
• test/routes/session/ passes: 22 tests, including the 4 tests that were already there. The assistant activity test now renders through AssistantMessageView.
• bun typecheck passes in packages/tui. oxlint reports no warnings on the new files, and prettier is clean on them.

Screenshots / recordings

Terminal frames from the render harness, 80 columns, for one transcript: a three-part run, a tool call, one more encrypted part, then text. The "before" frame renders the same parts with grouping disabled, which is how the base renders single parts.

Before:

   Thought · 1.0s

   Thought · 2.0s

   Thought · 3.0s

   $ ls

   Thought · 1.0s

   Here is the answer.

After:

   Thought 3 · 6.0s

   $ ls

   Thought · 1.0s

   Here is the answer.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

🤖 Generated with Claude Code

[brentyates-swx]

Session stamp for this PR:

• Session ID: e548b2ce-cde5-4bde-844e-535e815af08a
• Working repo (launch cwd): /home/byates/fm-worktrees/fm-llmrouter/opencode-thought-collapse-53/swx-opencode