OR-252 feat: show the coding agent's task list and counted tool activity in chat

[myles332]

Summary

Users could not tell which step the coding agent was on, or what it had already done, because progress lived only in a collapsed "Used tools" row. This PR surfaces it three ways:

• Checklist card in the transcript for the agent's own task list (Claude Code TaskCreate/TaskUpdate/TaskList or legacy TodoWrite, OpenCode todowrite, Codex update_plan). Only the newest state of a turn renders; a failed write stays an error row.
• Docked strip above the composer while the turn runs. With a task list it shows the current step, done/total and a progress bar. Without one it shows a progress outline read off the turn itself: each narration paragraph is a phase, annotated with the tool activity it produced, and expands to the full list of phases.
• Counted tool-group headlines ("Read 3 files · Edited 2 files · Ran 4 commands") replace the generic "Used tools" label.

Load-bearing details:

• Claude Code 2.1 exposes no task-list tool to newer models (Sonnet 5, Opus 4.8, …) unless CLAUDE_CODE_ENABLE_TODO_TOOLS is set, and then exposes incremental TaskCreate/TaskUpdate rather than TodoWrite. Chat sessions now set the flag, and ui/src/taskProgress.ts folds those calls across turns. Even so, Sonnet 5 usually skipped the tools in testing, which is why the narration-derived outline exists as the fallback.
• The shared playbook (SYSTEM_PROMPT.md) asks every harness to write its task list first and keep it current.
• Codex: turn/plan/updated (the only way update_plan reaches the client) becomes a task-list part, for main and sub-agent turns.
• A task-list write is excluded from the streaming tool tail so the Thinking status still shows; prior task lists are threaded only to messages that touch the list so memoized messages stay stable while streaming.
• ui/tsconfig.json gains allowImportingTsExtensions because runtime imports between tested modules need the extension under Node's test runner. ui/dist is rebuilt.

Test plan

• cargo fmt --all --check, cargo clippy --all-targets -- -D warnings, cargo build --locked, cargo test --locked (700 passed)
• ui: i18n lint, style lint, paraglide compile, tsc --noEmit, node --test (78 passed), vite build
• Isolated dev slot, injected Claude Code TodoWrite transcript: checklist card and counted headline render
• Isolated dev slot, live Codex turn with update_plan: docked strip updates step by step through "All tasks done"
• Isolated dev slot, live Claude Code (Sonnet 5) implementation turn without a task list: the outline strip shows six narrated phases with their tool activity
• Live Claude Code turn that actually calls TaskCreate/TaskUpdate (covered by unit tests against the CLI's real output formats; not observed live because Sonnet 5 skipped the tools)
• Live OpenCode turn with todowrite
• zh-CN and fa rendering of the new strings

🤖 Generated with Claude Code