This is a large slice (57 files, ~6k lines), so I focused on the one part with real runtime-behavior risk: the rewritten "missing final assistant reply" detection in api/streaming.py. I read the new helpers _session_lacks_final_assistant_answer (api/streaming.py:3497-3527) and _agent_result_terminal_failure (api/streaming.py:3529-3540), the call site (api/streaming.py:5690-5693), and diffed against the old guard on origin/master:api/streaming.py:5608-5609. The compression-exhausted classification (_classify_provider_error, new compression_exhausted branch) and the post-compression tool-result pruning (_prune_context_tool_results_after_compression) both look sound and defensive. One behavior change deserves a second look before merge.

The _token_sent guard was dropped, which can turn a successful tool-terminal turn into an error

On master the guard suppressed the silent-failure error whenever any text was streamed:

# origin/master:5608-5609
# _token_sent tracks whether on_token() was called (any streamed text)
if not _assistant_added and not _token_sent:

The PR replaces that with:

# HEAD:5690-5693
_terminal_failure = _agent_result_terminal_failure(result) or _session_lacks_final_assistant_answer(_all_result_messages)
if _terminal_failure:
    _assistant_added = False
if _terminal_failure or not _assistant_added:

_session_lacks_final_assistant_answer returns True whenever the transcript ends on a tool row (api/streaming.py:3505-3506):

role = msg.get('role')
if role == 'tool':
    return True

Crucially it makes this decision purely from message shape and ignores the result's success status — _agent_result_terminal_failure(result) is OR'd in, so even a result whose status is done is overridden to _assistant_added = False if the last persisted message is a tool result. The downstream branch then classifies with silent_failure=not bool(_err_str); with no error string this yields the "No response from provider" apperror. So a turn that streamed visible text and ended cleanly on a final tool batch with no closing assistant sentence now surfaces an inline error to the user, where master let it complete (because _token_sent was true).

That interaction is sharpened by this very PR: the strengthened progress contract (_WEBUI_PROGRESS_PROMPT, new lines telling the model to "say what you just confirmed and what you will check next before continuing with more tools") makes "emit prose, then run a final tool batch, then stop" a more likely shape, not less. A model that does exactly what the new prompt asks and whose last action is a successful tool call would get flagged.

Suggestion

Gate the shape-based check on the result not already reporting success, and/or restore the streamed-text escape hatch:

_terminal_failure = _agent_result_terminal_failure(result)
if not _terminal_failure and not _token_sent:
    _terminal_failure = _session_lacks_final_assistant_answer(_all_result_messages)

That keeps the genuine target case (agent fails mid-tool-run, nothing streamed, no final answer) firing, while not penalizing a turn that produced visible progress and merely ended on a tool result. If ending-on-tool is intentionally treated as failure even when text streamed, please say so in a comment near api/streaming.py:3505 and add a test_live_stream_ux case pinning "streamed text + tool-terminal + success status" to the chosen outcome — right now _session_lacks_final_assistant_answer's status-blindness isn't covered by an assertion that distinguishes it from _agent_result_terminal_failure.

Everything else I sampled in streaming.py reads cleanly. I didn't execute the suite (cron policy), but the diff's own claim of 108 passed on the focused files plus the message-wording changes (Compressing context / Compression finished) line up with the test_auto_compression_card.py updates. Worth splitting the non-streaming doc/CHANGELOG churn from the behavior change if you want a tighter review surface, but that's process, not correctness.

sequenceDiagram
    participant FE as Frontend (sessions.js / messages.js)
    participant SC as StreamChannel (config.py)
    participant RJ as RunJournal (run_journal.py)
    participant SSE as SSE Handler (routes.py)
    participant Worker as Streaming Worker (streaming.py)

    Note over FE,Worker: Live turn in progress
    Worker->>SC: note_last_event_id(event_id)
    Worker->>SC: put_nowait((event, data, event_id))
    SC->>SC: "_last_event_id = event_id"
    SC-->>FE: broadcast 3-tuple to active subscribers

    Note over FE,Worker: User switches session — SSE disconnects
    SC->>SC: Buffer items in _offline_buffer

    Note over FE,Worker: User switches back — reconnect
    FE->>SSE: "GET /api/chat/stream?replay=1&after_seq=N&after_event_id=X"
    SSE->>SC: subscribe_with_snapshot()
    SC-->>SSE: "(subscriber_queue, {last_event_id: Y})"
    SSE->>SSE: "snapshot_cutoff_seq = parse(Y)"
    SSE->>RJ: "_replay_run_journal(after_seq=N, max_seq=Y, include_stale=False)"
    RJ-->>SSE: journal events N+1 to Y
    SSE-->>FE: replay events via SSE
    SSE->>SSE: "replay_cutoff_seq = Y"

    loop Live stream tail
        SC-->>SSE: item from subscriber_queue
        SSE->>SSE: "event_seq = parse(item.event_id)"
        alt "event_seq <= replay_cutoff_seq"
            SSE->>SSE: skip duplicate
        else "event_seq > replay_cutoff_seq"
            SSE-->>FE: emit SSE event
        end
    end

_{Reviews (17): Last reviewed commit: "Close #3401 merge review test gaps" | Re-trigger Greptile}

Pushed follow-up commits addressing review feedback. Summary of what changed on this branch:

207c09f9 — Fix false "no response" on streamed tool-terminal turns (addresses @nesquena-hermes's review)
The missing-final-assistant guard OR'd the status-blind _session_lacks_final_assistant_answer check in unconditionally, so a successful turn that streamed visible progress and ended on a final tool batch (or a leaked role=user control message) was reclassified as a terminal failure and surfaced a false error. The shape check is now gated behind not _token_sent, while _agent_result_terminal_failure(result) stays authoritative for explicit failure/partial/compression-exhausted status. Added behavioral unit tests for both helpers (tool-tail, user-tail, empty-messages, error-tail, success) and realigned the static guard to the corrected semantics.

4c48968b — Removed the reasoning-effort coercion from this PR. It's unrelated to the live-to-final reply experience and changes behavior for all reasoning-capable models, so it now ships as its own focused PR (#3505). The StreamChannel snapshot/event-id changes in api/config.py stay here because they're part of the reconnect-replay work.

b69bad5a — Addresses the Greptile review:

• P1 (reconnect dedup was a no-op): correct catch — the worker keeps the queue 2-tuple (Stage-364) and propagates the journal event id via the STREAM_LAST_EVENT_ID side-channel, so StreamChannel._last_event_id was never populated, snapshot_cutoff_seq/replay_cutoff_seq never engaged, and offline-buffer events could be redelivered after journal replay. Rather than switch to a 3-tuple (which would break the Stage-364 design and its static tests), added StreamChannel.note_last_event_id() and call it from the worker put() and cancel_stream, so the cutoff now actually engages while the 2-tuple queue shape is preserved. Added a unit test for the snapshot wiring plus a static guard so put() can't silently regress to inert.
• P2 (unbounded while True): the ownership-recheck loop is now capped at 3 attempts and returns 409 instead of spinning if stale cleanup keeps succeeding while another thread re-claims the session.
• P2 (misleading indentation in loadSession): re-indented the INFLIGHT block; whitespace-only (git diff -w shows no non-whitespace change, node --check passes).

Local verification: tests/test_cancelled_turn_status.py, test_webui_runtime_diagnostics.py, test_stage364_opus_live_sse_event_id.py, test_run_journal_routes.py, test_stale_stream_cleanup.py, test_inflight_stream_reuse.py, test_run_journal_streaming_static.py, test_sprint42.py, test_regressions.py all pass locally; full matrix runs in CI.

Note: this branch will periodically re-conflict on CHANGELOG.md only (the [Unreleased] block) whenever a new release lands on master — no code conflicts.

🤖 Generated with Claude Code

Updated with merge commit 53727cb to bring the branch onto latest master, resolve the CHANGELOG conflict, and remove the duplicate _active_stream_ids import noted in review. Local verification: git diff --check; pytest tests/test_cancelled_turn_status.py tests/test_webui_runtime_diagnostics.py tests/test_stage364_opus_live_sse_event_id.py tests/test_run_journal_routes.py tests/test_stale_stream_cleanup.py tests/test_inflight_stream_reuse.py tests/test_run_journal_streaming_static.py tests/test_sprint42.py tests/test_regressions.py -q (179 passed).

Conflict cleanup and CI fallout fixes pushed through 183e8258.

What changed:

• Merged latest origin/master into franksong2702/live-to-final-assistant-replies and resolved conflicts in the live-to-final scope.
• Restored merge-lost static UI invariants from current master:
  • partial tool-call assistant rows remain visible/anchorable;
  • thinking-only messages in simplified tool-calling mode render inline instead of inside a collapsed activity group;
  • persistent-state tool_complete notifications see tc.is_error before toast classification;
  • live compression card replacement restores the scroll snapshot before follow-settle.
• No fix(runtime): make cancelStream() owner-aware and close its SSE source #3345 active Stop changes or unrelated mobile titlebar changes were folded into this PR.

Verification:

• python -m pytest tests/test_issue401.py tests/test_issue3592_thinking_settlement.py -q -> 16 passed
• python -m pytest tests/test_issue3340_persistent_state_toasts.py tests/test_issue3479_ios_stream_scroll_jump.py tests/test_issue401.py tests/test_issue3592_thinking_settlement.py -q -> 27 passed
• python -m pytest tests/test_cancelled_turn_status.py tests/test_webui_runtime_diagnostics.py tests/test_stage364_opus_live_sse_event_id.py tests/test_run_journal_routes.py tests/test_stale_stream_cleanup.py tests/test_inflight_stream_reuse.py tests/test_run_journal_streaming_static.py tests/test_sprint42.py tests/test_regressions.py tests/test_webui_gateway_chat_backend.py -q -> 200 passed
• node --check static/messages.js static/ui.js static/sessions.js static/boot.js
• git diff --check
• GitHub Actions: 11/11 passed on 183e8258

AI assistance: Codex coordinated the merge cleanup with a sub-agent, reviewed the returned diff, fixed the CI-reported merge fallout, and reran focused verification before pushing.

Conflict cleanup pushed in 19598a70.

What changed:

• Merged latest origin/master (4c545a33) into franksong2702/live-to-final-assistant-replies.
• Resolved conflicts in CHANGELOG.md and static/messages.js.
• Kept this PR's live-to-final Unreleased notes while preserving current release history.
• Preserved the current master terminal-event stale-stream bail-out in done while keeping this PR's immediate _streamFinalized behavior, so the prior merge-lost static UI invariants remain intact.

Verification:

• git diff --check -> passed
• node --check static/messages.js static/sessions.js static/ui.js static/boot.js -> passed
• /Users/xuefusong/.hermes/hermes-agent/venv/bin/python -m pytest tests/test_cancelled_turn_status.py tests/test_webui_runtime_diagnostics.py tests/test_stage364_opus_live_sse_event_id.py tests/test_run_journal_routes.py tests/test_stale_stream_cleanup.py tests/test_inflight_stream_reuse.py tests/test_run_journal_streaming_static.py tests/test_cancel_stream_owner_guard.py tests/test_issue3587_intermediate_reasoning.py tests/test_session_events.py -q -> 132 passed
• GitHub Actions on 19598a70 -> 11/11 passed
• GitHub mergeability -> MERGEABLE / CLEAN

AI assistance: Codex performed the conflict cleanup and focused regression review.

Conflict cleanup pushed in 72b838fd.

What changed:

• Merged latest origin/master (f1211e1f, v0.51.267) into franksong2702/live-to-final-assistant-replies.
• Resolved the CHANGELOG.md conflict by keeping this PR's live-to-final notes in Unreleased and preserving the new v0.51.267 security release notes.
• api/routes.py auto-merged with the v0.51.267 security hardening changes; no manual behavior conflict was needed there.

Verification:

• git diff --check origin/master..HEAD -> passed
• node --check static/messages.js static/sessions.js static/ui.js static/boot.js -> passed
• python -m pytest tests/test_cancelled_turn_status.py tests/test_webui_runtime_diagnostics.py tests/test_stage364_opus_live_sse_event_id.py tests/test_run_journal_routes.py tests/test_stale_stream_cleanup.py tests/test_inflight_stream_reuse.py tests/test_run_journal_streaming_static.py tests/test_cancel_stream_owner_guard.py tests/test_issue3587_intermediate_reasoning.py tests/test_session_events.py tests/test_issue1909_csrf_token.py tests/test_issue2931_edge_tts_endpoint.py tests/test_sprint29.py -q -> 216 passed
• GitHub Actions on 72b838fd -> 11/11 passed
• GitHub mergeability -> MERGEABLE / CLEAN

AI assistance: Codex Autopilot performed the branch refresh, reviewed final scope, ran focused verification, pushed, and read back GitHub state.

Updated the #3401 Thinking/Worklog blocker.

Product model:

• Worklog remains the live-to-final record for an assistant turn.
• Thinking is preserved as its own Worklog Thinking Card, sibling to process prose and Tool Cards.
• Thinking is not promoted into final answer text and is not treated as a Tool Card.

Implementation:

• Live Thinking Cards are now segment-scoped, so later reasoning does not keep updating the first Thinking Card in the turn.
• Settled rendering keeps Thinking Cards in the folded Worklog.
• Duplicate suppression is intentionally narrow: exact / normalized-exact only against visible process/final text from the same assistant turn.
• Reasoning metadata is preserved even when the visible Thinking Card is suppressed.

Tradeoff:

• Partial overlap and semantic dedupe are intentionally not handled in this PR. That avoids live content jumping and avoids model/provider-specific behavior.

Verification:

• node --check static/messages.js static/ui.js
• pytest tests/test_regressions.py tests/test_ui_tool_call_cleanup.py tests/test_issue2565_reasoning_accumulation.py tests/test_issue3592_thinking_settlement.py tests/test_issue_progress_echo_dedupe.py tests/test_issue2454_active_session_spinner.py
• npx --yes eslint@10.4.0 --no-config-lookup -c eslint.runtime-guard.config.mjs "static/**/*.js"

Follow-up after the latest refresh/re-gate:

• Refreshed Redesign live-to-final assistant replies #3401 onto current origin/master and restored the reviewer-branch guardrails that were easy to lose during conflict resolution: idle attention-dot visibility/color, the dead settleLiveCompressionCards() removal, and stale test expectations around Fix compression-exhausted stream finalization #3316 terminal-failure semantics / Redesign live-to-final assistant replies #3401 Thinking replay.
• The intended product model is unchanged: process prose, Thinking Card, and Tool Card are sibling items inside the Worklog. Thinking is not silently dropped and is not treated as a Tool Card or Final Answer.
• Settled duplicate suppression is intentionally narrow: only exact / normalized-exact Thinking echoes are hidden from the folded Worklog. We keep the underlying reasoning metadata, preserve non-exact provider reasoning, and do not attempt partial-overlap, semantic, or model-specific dedupe. That avoids live-stream content jumping while still removing the obvious Spark/openai-codex exact echo in the settled view.
• Latest head 002a57a6 is green on Browser smoke and the full Tests matrix, including lint and the Scope / undefined-reference gate.