feat: stream claude-code tokens (reland #6664) + fix the orphan reasoning-end that killed runs

[pedrofrxncx]

Relands #6664 (reverted in #6669) with the bug that forced the revert fixed.

Stacked on #6669 — base is the revert branch, so this merges as: revert lands, then #6664 comes back fixed. Retarget to main if #6669 merges first.

What broke

A streamed text/reasoning block was ended after the finish-step that the next assistant message pushes (claude-code.ts pushed the step boundary before translator.translate(message), and translateAssistant is where the open block was closed). The AI SDK reducer clears its open text/reasoning parts on finish-step (process-ui-message-stream.ts:770-774), so the end arrived for a part it no longer knew:

[Decopilot] stream pump error for thread thrd_T9FUqcMjz9KaeubUZJfMi:
Received reasoning-end for missing reasoning part with ID "stream-2".

That throw kills the pump mid-run. No finish chunk ever reaches the JetStream log, so resolveCleanRunStatus(undefined) reads the truncated stream as a clean end: the thread is stored completed while the sandbox agent is still working, and the thread-finish hook advances the task board card to In Review. Observed in prod on board_x5gk0kjwXRfgmf3TFZ_JS — the thread's finish part carries no usage, unlike healthy runs.

The fix

Close whatever stream_event left open before pushing finish-step. closeOpenStreamBlocks() becomes public for that; translateAssistant still calls it (now a no-op on that path), so a block closed by content_block_stop is unaffected and the assistant restatement never emits a second end.

Testing

• bun test packages/harness-runner — 69 pass, incl. a new case: a block left open at the step boundary is closed once, and the assistant restatement adds nothing.
• bun run --cwd=packages/harness-runner check, bun run fmt.

---

Summary by cubic

Relands token-level streaming for Claude Code runs: text now arrives as the model writes it instead of one large frame after the turn finishes, and the block-ordering bug that made the first merge unsafe is fixed — an interrupted stream no longer throws, kills the run mid-stream, and records the thread completed while the agent keeps working. Also adds a live elapsed readout for long-running tool calls and two-minute working-tree checkpoints in the sandbox daemon.

Claude Code streaming

• Enables includePartialMessages; the translator turns raw stream_event messages into text/reasoning start-delta-end chunks and skips the assistant message's restatement of blocks already streamed.
• A coalescer concatenates adjacent deltas and flushes at ~200 characters, so token streaming does not become per-token frames through the daemon, JetStream, or SSE.
• Blocks still open at a step boundary are now closed before finish-step; an end emitted after that lands on a part the AI SDK reducer has already cleared.

UI and sandbox durability

• Any still-running tool call shows a live "Running for Ns" counter once it passes 10 seconds, replacing the static "Preparing…" that made long builds look frozen.
• The sandbox daemon checkpoints a run's working tree to its branch every two minutes, so a hard pod death loses at most that window of work.
• Checkpointed publishes never commit .env variants anywhere in the tree; .envrc is still committed.

^{Written for commit ef3975d. Summary will update on new commits.}