[danshapiro]

What users see now

In a freshclaude or kilroy pane, a running turn is visible while it runs instead of only after it ends:

• The answer and the model's thinking appear as they are written (about ten updates a second), in place. When the saved transcript catches up, the text stays where it is: no flicker, no second copy, no blank moment.
• A tool row appears as soon as the model starts the call. Its result row follows within about half a second of the tool finishing (it used to wait for a 3-second poll).
• While Claude retries an overloaded API request, the pane shows "Retrying (attempt N) — overloaded (529)", with the Stop button, and the row goes away when the answer starts.
• Rate-limit warnings, a failed /compact and the CLI's own warnings show in the pane's notice banner without ending the turn.
• Subagent (Agent/Task) delegations update their activity rows while they run.
• After approving ExitPlanMode, the settings gear shows the mode Claude went back to, and the next message keeps it.
• The session cost no longer adds up turn totals twice (the CLI already reports a running total).
• Other panes and the rest of the app (tab bar, sidebar, pane headers) do not re-render while text streams.

Phase A (the saved-transcript projection: delegation blocks, compaction cards, /cost output, the diff panel, [image], API error cards) shipped in #860; this PR is the live part on top of it.

What changed, by phase

• B, sidecar (crates/freshell-claude-sidecar): the SDK message mapping is a pure module (sdk-message-frames.mjs). The sidecar announces each transcript line it expects the CLI to write (sdk.session.changed {reason:'transcript', uuid}), stamps stream events with their API message id, and streams only what the pane renders. API retries become sdk.retry; rate limits, failed compactions and the CLI's prominent informational messages become sdk.notice (never a turn end). The CLI's own permission-mode change becomes sdk.session.metadata.
• C, Rust (freshell-freshagent, freshell-ws): retries and notices reach every client (freshAgent.retry, freshAgent.notice) with structured logs; the CLI's permission-mode change folds into the session's settings and is stated in metadata. Snapshots read the transcript store the lane's sessions write, and the projection is incremental (only lines appended since the last read are parsed, off the async runtime).
• D, client: a separate freshAgentLive store slice holds the live tail (streaming blocks, the retry, notices), so token-rate updates never touch the sessions map. A 100 ms batcher feeds it. The tail renders through the transcript's own turn article and hides each block in the same frame its transcript line lands (by line uuid, or by message id and position when the uuid frame was missed). Panes refetch the snapshot per announced line, re-poll a bounded number of times for lines the CLI is slow to write, and clear the tail once the turn's lines are on screen.
• E, browser tests and docs: a shared browser lane (test/e2e-browser/helpers/claude-sim-lane.ts) runs the production sidecar over the CLI simulator; freshclaude-live-transcript-rust.spec.ts covers streamed text, tool rows, retries, /compact, delegations, the diff panel, /cost, the plan-mode exit and a reload, for freshclaude and kilroy. docs/index.html shows a turn rendering live.

Review fixes in this PR (after the phase work)

1. An error no longer wipes the live tail. A failed turn (result, then its error, about 100 ms before the CLI writes the last line) kept its answer only by luck, and a refused settings change in the middle of a block dropped the rest of the block. The turn's end is now its own signal (a result or an idle status); an error drops only the retry row, and the pane's clear keeps a block whose turn is still streaming.
2. Tails are bounded for sessions this browser does not show: visible panes register as the tail's viewers, and a tail no visible pane settles is dropped when its turn ends. A stopped recovery applies queued stream events before it removes the session.
3. A message sent right after a turn ends shows below the previous answer, not above it, until that answer's line lands; the next turn's text shows below the new message.
4. A snapshot refetch re-renders only the turns that changed: unchanged turns and items keep their objects across commits and the item card is memoized. Three refetches of a 40-answer transcript went from 240 markdown renders to none.
5. A permission mode the pane cannot show (the SDK's auto or dontAsk) makes the pane drop its own mode, so the next send no longer puts Claude back into plan mode.
6. The CLI's quiet informational messages (info, notice, per-tool progress) no longer raise an alert banner; only warnings, suggestions and messages that stop the turn do.
7. An API retry marks the turn running, so a new session's first request shows Stop while it retries (its transcript does not exist yet, so no snapshot can say so). The first streamed block of a just-created session (status connected, not idle) now marks it running too.
8. Simulator conformance: an approved ExitPlanMode returns to the mode before plan mode (new plan-return scenario), and API retries are compared against the real CLI through a shared RETRY:<status>x<n> prompt marker (new retry scenario). That comparison found the CLI replays the prompt only after a retried request is answered, which is what exposed item 7.
9. Test hygiene: the live-frames Rust fake no longer leaves sidecar or store paths in the process environment; a test now holds the client's guard that ignores subagent stream events.

origin/main (#870, core history) is merged in; the conflicting snapshot helpers now use main's disk_snapshot (genesis and rollback aware) with this branch's incremental projection.

Tests (all at the final commit unless noted)

• Client, sidecar and jsdom flows: pnpm run test:vitest run test/unit/claude-sidecar test/unit/client/components test/unit/client/lib test/unit/client/store test/e2e test/unit/contracts --config config/vitest/vitest.config.ts — 6949 passed, 1 failed: terminal-file-link-same-tab (terminal links, unrelated to this work; the host's load average was about 20), which passed on 3 reruns alone. An earlier full run had a similar one-off in open-tab-session-sidebar-visibility, also unrelated and also green on 3 reruns.
• Provider contract and simulator conformance (pinned CLI 2.1.237 against the loopback provider only): pnpm run test:vitest run test/integration/providers/ --config config/vitest/vitest.config.ts — 89 passed (after the merge).
• Rust: cargo test -p freshell-freshagent --lib — 1496 passed; cargo test -p freshell-ws --test freshagent_claude_live_rendering --test freshagent_claude_real_cli --test freshagent_claude_lifecycle --test freshagent_claude_rollback — 2, 15, 12 and 17 passed (after the merge).
• Browser, cloud backend: GCLOUD_ROBOT_REQUIRE=1 pnpm run test:e2e <freshclaude-live-transcript-rust, fresh-agent-control-rust, fresh-agent-error-turn-attention, freshclaude-durable-transcript-rust, fresh-agent-rollback-rust> --project=chromium — 57 passed, none retried (Cloud Run, image of the final commit). Earlier runs in this PR found the retry case failing (Stop missing during a new session's first-request retries), which fixes 7 address.

Follow-ups

• kata 38ve: move fresh-agent-error-turn-attention.spec.ts onto the shared ClaudeSimLane helper once core PR-D (which edits that spec) lands.
• The simulator's tool-input streaming step and replay workspace option are still checked only by the simulator's own tests, not against the real CLI.
• When the CLI is in a mode the pane cannot show (auto, dontAsk), the gear shows the pane's default mode rather than the CLI's.
• The plan's adaptive minimum gap between refetches was not added; sharing unchanged turns removed the cost it was meant to bound. Worth revisiting if very long sessions still feel slow.
• The Rust side does not yet broadcast "running" when a turn starts (core's G21); the client-side retry mark in fix 7 covers the gap this PR found.

🤖 Generated with Claude Code