fix(client-runtime): Pi subagent rows show their thinking level#30489

Problem

The hover card on a Pi subagent row shows only the model, so you can't tell what reasoning level a delegated child runs at. Claude (effort), Codex (reasoningEffort), Cursor (reasoning) and OpenCode (variant) rows already show it. Pi names its reasoning option thinking, and that id was missing from the shared list the label reads (REASONING_EFFORT_OPTION_IDS).

To reproduce on main (7c22e24):

1. Start a Pi thread on anthropic/claude-haiku-5-5.
2. Have it call delegate_task with options: [{ "id": "thinking", "value": "high" }].
3. While the child runs, hover its row. The card reads "Haiku 5.5".

An OpenCode child delegated with variant: max reads "Beam Free · max" in the same card.

Change

• thinking is now the last entry of REASONING_EFFORT_OPTION_IDS, so a model that has a dedicated effort select still shows that.
• Claude and Cursor also have an option called thinking, but it's an on/off switch. Every consumer skips on/off values, so those never become a label.
• modelTraits.ts and handoff.ts each added thinking to that list themselves. They now use the shared list as it is, with the same ids in the same order, so their behaviour doesn't change.

Scope and approval

Fixes pingdotgg/t3code#30442

Verification

Tests:

• threadExecution.test.ts › formatModelSelectionEffort › labels a Pi thinking level but not Claude's boolean thinking toggle fails on main (expected null to be 'Low') and passes here.
• The same test checks that Claude's on/off thinking still gets no label.
• vp-check-changed passes on the head: format, lint, knip, typecheck of 7 packages, and 41 tests across 2 files.

Evidence. Before is 7c22e24. After is 7715109; the head adds only a merge of main.

• Both sides ran the real web dev server with a fresh database and T3CODE_DEV_SEED=0.
• Setup: Pi enabled; the parent on anthropic/claude-haiku-5-5 delegates one async child at thinking: high that runs sleep 3600.
• Browser: Chromium 156 through agent-browser, DPR 1, no device emulation, 100% zoom, reduced motion, browser clock frozen.

The diffs below mask four regions that come from real runs and can't match across two servers: the changed digits of the parent's "Worked for" duration and of three elapsed timers. Each mask is a cyan box drawn around exactly the changed digit pixels, plus 1 px; no static text is masked. With them masked, the only change at zero threshold in every pair is one 31×12 px box: the · high suffix. The unmasked diffs are kept with the run's scratch files.

R1: running Pi child, timeline row hover card

Steps (before / after): 1.0 s hover the "Pi high proof" row; card readable at about 1.5 s; dwell to 4.5 s; pointer leaves.

1440×900 dark	Before (7c22e24)	After (7715109)
Recording

Viewport / theme	Before (7c22e24)	After (7715109)	Diff0, timer digits masked	Diff16, timer digits masked
1440×900 dark
1440×900 light
390×900 dark
390×900 light

R2: same child, Related threads hover card in thread details

Steps: 1.0 s open thread details; 1.7 s hover "Pi high proof" under Related threads; dwell to 5.2 s; pointer leaves.

1440×900 dark	Before (7c22e24)	After (7715109)	Diff0, timer digits masked	Diff16, timer digits masked
Still
Recording

C1: control harness, an on/off thinking: true stays model-only

This is a labelled component harness, not a real Claude turn. It mounts the real SubagentTooltipContent inside the real hover card, with a Claude-style on/off thinking option and a saved thinking: true. Both sides read model only, and the two frames are pixel-identical at zero threshold.

1440×900 dark	Before (7c22e24)	After (7715109)	Diff0	Diff16
Still
Recording

What else changed (all inside the masks):

• The parent's "Worked for" went from 5.4 s to 6.2 s.
• The child's elapsed timers moved by one second.
• Nothing else.

Seeding: in each side's fresh home, userdata/settings.json gets "pi": {"driver":"pi","enabled":true,"config":{"enabled":true}} under providerInstances. Then:

• node src/bin.ts project add <empty fixture repo>
• node src/bin.ts thread start --provider pi --model anthropic/claude-haiku-5-5 --json <project> '<prompt>'

The prompt was: "Call orchestrator_capabilities, then delegate_task exactly once with target provider pi, model anthropic/claude-haiku-5-5, options [{"id":"thinking","value":"high"}], mode async, title "Pi high proof", and task "Run sleep 3600 in the shell, wait for it to return, then reply done. Make no file changes." Keep the delegation available. Do not replace it with the provider-native subagent tool."

Not checked:

• Mobile: no current producer gives a mobile row a Pi thinking level. Pi's own subagents have no T3 child thread, and plain mobile agent rows carry no effort.
• Desktop: it renders the same web components.

Model and harness: Claude Opus 5.5 (Claude Code) as lead, test writer, implementer and reviewer B; GPT-6.1 Sol (Codex) as reviewer A, proof auditor and evidence capturer, all orchestrated through T3 Code.

🤖 Generated with Claude Code