[LivXue]

Summary

A DAG node that stops on a decision is announced to the main agent with the runner's brief as the fenced report (_exception_report): the verdict, the resolve_dag_node calls it may make, and that tool's whole schema. The web UI took the fence's text as the delivery row's body, so What came back on a row waiting on a decision opened onto that brief, run together into one paragraph, and never showed what the node had said. A spawn never reaches this state; only a DAG node's own report does.

• The runner passes what the node returned on the attempt being reported (its error, when it raised) through every announcer -- the background lane, a foreground run's outbox including its re-send on release, and a stint driver's pass-through -- to announce_dag_exception, which carries it as output on the delivery mark. The injected text the model reads is unchanged.
• The mark is stored with the conversation and sent on every replay, so output is cut at 20,000 characters, the head dag.node reads by default; the run dir keeps the whole text. It is present even when empty, so a node that said nothing is told apart from a runtime that never sent it.
• output is declared on subagent.delivered, on turn.started's delegated and on a stored entry's delegated (OpenRPC, the Pydantic models, both generated clients), and the descriptions that named the fence as the only readable body now name output as well.
• The web fold shows output under its own caption, What the node returned. A stored entry without it keeps the fold it had.
• A report about a node that has already failed (its continuations spent, or no route to answer it) was marked exception and read Waiting on a decision; it is now marked error and reads Failed. On the web, notice (a stall notice about a node still running) and cancelled fell through to ok and read Finished; they now read Still running and Stopped.
• The run tally (N completed, N failed, ...) is read only off a graph's own result. A node's report carries its own status, and its output may quote a sub-graph's tally; such a row would otherwise read Partly done.
• The offline fixture's suspended node now fences the brief the runner really writes and carries the node's words as output, so the web UI's offline mode shows what ships.

Not changed: the TUI does not read output (its rows already word error), and the model-facing report and ask are byte-identical.

Type

• Fix
• Feature
• Docs
• CI / tooling
• Refactor
• Other

Verification

At the head, on origin/main 4315993:

• uv run --frozen --all-extras pytest -q -n 2 over tests/test_subagent_*.py, tests/test_rpc_*.py, tests/test_spine*.py, tests/test_agent_loop_session*.py, tests/test_playbook*.py, tests/test_agent_loop_subagent*.py and tests/test_security_untrusted_context.py: 7322 passed, 2 failed. tests/test_subagent_node_runtime.py::test_what_cannot_be_read_names_nothing fails whenever the suite runs as root (it builds a file it expects to be unreadable) and imports only raven.agent.subagent.node_runtime, which this change does not touch. tests/test_subagent_acp_remote.py::test_a_daemon_killed_outright_mid_turn_ends_the_turn_with_the_reason_and_leaves_no_agent hit its time budget under load and passed three times alone; it reaches none of the announcer code.

• npm test in ui-web, both test roots: 253 files, 4220 of 4222 tests passed. The two others timed out at 5 s under load (src/state/session/pipeline.events.test.ts and src/state/session/registry.test.ts), and both files then passed twice alone, 68 of 68; the same suite passed 4222 of 4222 on the content before the rebase.

• ui-web: npm run gen:check, npm run type-check, and npm run lint (0 errors; its 4 warnings are in files this change does not touch). ui-tui: npm run lint:rpc, npm run lint:i18n, npm run type-check, and src/__tests__/chatStream.test.ts (54 passed).

• ruff check, ruff format --check and ty check over the touched Python files; lint-imports: 10 kept, 0 broken.

• scripts/check_commit_messages.py origin/main..HEAD, commitlint on the commit and on the prospective squash header, scripts/check_source_language.py origin/main...HEAD and scripts/check_large_files.py origin/main...HEAD: all exit 0.

• Every new test was watched failing before its fix. A pass of 22 single-construct mutants -- each forwarding hop, the read of a crash's error before the judge's reason overwrites it, the error status, empty versus absent, the cap and its boundary, the live and replay pass-throughs, the caption, the tally predicate and the two new labels -- left none surviving.

• Real Chrome on the web UI's offline mode (vite, src/page.html?stub=1): the scheduled conversation's suspended row reads Waiting on a decision and folds open onto What the node returned and the node's own two sentences.

• Relevant tests pass locally

• Relevant lint / type checks pass locally

• User-facing docs or screenshots are updated when needed

The local failures are the ones named above: a root-only test, and three timing-bound tests that pass alone. The changelog carries the entry.

Risk

• A field is added to the mark, nothing is removed: a client that ignores output (the TUI, an older web build) keeps showing the fence's text, and an entry stored before this keeps its old fold.

• A failed node's mark changes from exception to error. The mark's status is read only by the two clients' row labels; nothing in the runtime routes on it.

• A stored conversation grows by up to 20,000 characters per node report. The text is the sub-agent's own and is kept unredacted, by the same rule the spawn announce applies to a result.

• ExceptionAnnouncer, the outbox's AnnounceReport and put_report, and the DAG tool's three adapters gain a defaulted output keyword, and every in-tree implementation takes it. An announcer written against the old fixed signature would raise on it; none exists in the tree.

• Rollback: revert the squash commit. Nothing validates a stored mark at runtime, so entries that already carry output are passed through and ignored by the old code.

• Security impact considered

• Backward compatibility considered

• Rollback path is clear for risky changes

Related Issues

N/A

[LivXue]

The fix in real Chrome, on the web UI's offline mode (vite, src/page.html?stub=1): the scheduled conversation's row for a node waiting on a decision, folded open. It shows what the node returned, under its own caption; before this change the same fold opened onto the runner's brief to the model, resolve_dag_node schema included.

[gloryfromca]

No blockers; this can merge as far as I am concerned.

Reviewed the full diff and traced the node output through the runner, foreground outbox/release path, manager injection, stored transcript mark, live RPC boundary, generated schemas, and web live/replay rendering. I also checked relevant callers and history, the repository rules and canonical context terms, backward compatibility for older entries without output, status handling, test changes for weakening, and the architecture boundary between runtime, RPC, and UI. I tried to refute the edge cases around empty output, crashed nodes, repeated attempts, stall notices, released foreground runs, truncation, and graph-tally text; the implementation and tests account for them.

Verification: uv run --extra dev pytest tests/test_subagent_dag_adjudication.py tests/test_subagent_dag_runner.py tests/test_subagent_manager.py tests/test_playbook_stint.py -q (787 passed); focused web Vitest run (231 passed); web and TUI TypeScript type checks; web and TUI RPC generation checks; TUI i18n generation check; source-language check; and git diff --check all passed. An earlier Python invocation without the dev extra failed during shared fixture setup because raven_everos was absent, before any test body ran; rerunning with the repository's dev environment passed.

[0xKT]

Not a blocker (review of 72effe49, main 43159933): the node's output reaches the delivery mark on the paths that carry a mark, the text the model reads is byte-identical to main's in every scenario the review drove, and the four schema copies agree. Three small things the web/TUI side may want, read at the head:

Severity of this one finding, not a verdict on the pull request. Any blocking findings from this pass are review threads on the changed files; GitHub renders those collapsed, as a file name with no text.

1. The TUI and the web now word a notice differently. ui-tui's deliveredMessageKey (ui-tui/src/domain/messages.ts:88-99) maps error, exception and cancelled and lets notice fall through to gui.deleg.delivered ("Finished"); the web row now says it is still running.
2. The cut marker reads as the node's own words. manager.py:2419-2424 appends "\n... (output truncated)" inside output, so the fold shows it under the node's returned text (and the mark is 20,000 characters plus the marker, where CHANGELOG says "cut at 20,000"). dag.node reports the same cut as a separate output_truncated flag.
3. The fold has no height bound. .sdlv .sdbd (ui-web/src/styles/page.css:1481-1484) sets overflow-x: auto and no max-height, so a 20,000-character output opens as one very tall block in the transcript.

[LivXue]

No new Windows compatibility issue found in the inspected production changes.

Reviewed head 72effe49b98845a8c0c50e9d3be19f4a92e12154, base/merge base 4315993300e538ec954c79565cbc8585006c8053: complete 24-file diff, runner/outbox/stint forwarding, manager injection, UTF-8 transcript storage, RPC live/replay projection, generated clients, Web rendering, tests and discussions. This is a Windows-scoped assessment, not a merge verdict.

Native Windows verification (Python 3.13.12, Node 25.9.0):

• Original focused Python run: 48 passed, 1 failed, stopping at test_a_re_sent_report_still_carries_what_the_node_returned. The unchanged _foreground_tool fixture invokes external cat, unavailable here (FileNotFoundError: [WinError 2]); the new assertion needs its echoed output. This is a test-environment limitation, not evidence of a runtime regression. The older foreground-handoff control at the exact base passed, but does not assert echo output.
• Remaining focused Python cases: 12 passed, covering forwarding/status/schema/empty-output/truncation and stint routing.
• The failing release-output case with only the external echo command replaced by a native Python echo process in a scratch pytest plugin: 1 passed. The script path contained spaces and non-ASCII characters; no repository source was edited. This validates that release/re-send behavior with that substitute, not the original cat fixture.
• Web transcript and session-event suites: 231 passed.

The existing general UI observations are not re-filed. Limits: no full repository suite, live model/remote dispatch, browser visual check or interactive TUI. Tests used detached exact-revision checkouts and isolated profile/Raven/OAuth state.

Exact verification commands and results

uv sync --frozen --python 3.13 --extra dev
# native; exit 1; cwd C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\pr1063\head
uv run --frozen --no-sync pytest tests/test_subagent_dag_adjudication.py tests/test_subagent_dag_runner.py::test_a_node_report_carries_what_the_node_itself_returned tests/test_subagent_dag_runner.py::test_a_crashed_node_report_carries_its_error_as_what_came_back tests/test_subagent_dag_runner.py::test_a_continued_nodes_second_report_carries_the_second_attempts_output tests/test_subagent_dag_runner.py::test_a_re_sent_report_still_carries_what_the_node_returned tests/test_subagent_dag_runner.py::test_a_stint_driver_passes_a_notification_on_with_its_output tests/test_subagent_dag_runner.py::test_the_announcer_is_told_whether_a_decision_is_pending tests/test_subagent_manager.py::test_an_informational_node_notice_is_headed_as_a_notice_not_a_failure tests/test_subagent_manager.py::test_a_node_report_marks_what_the_node_returned_for_the_reader tests/test_subagent_manager.py::test_a_node_report_marks_an_empty_return_as_empty_not_absent tests/test_subagent_manager.py::test_a_node_report_bounds_the_output_it_marks tests/test_subagent_manager.py::test_a_stall_notice_marks_no_output tests/test_playbook_stint.py::TestWhoAnswersARoundsSuspendedNode -n 0 -x --tb=short --basetemp C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-native\pytest --junitxml C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-native\junit.xml

# native-rest; exit 0; cwd C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\pr1063\head
uv run --frozen --no-sync pytest tests/test_subagent_dag_runner.py::test_a_stint_driver_passes_a_notification_on_with_its_output tests/test_subagent_dag_runner.py::test_the_announcer_is_told_whether_a_decision_is_pending tests/test_subagent_manager.py::test_an_informational_node_notice_is_headed_as_a_notice_not_a_failure tests/test_subagent_manager.py::test_a_node_report_marks_what_the_node_returned_for_the_reader tests/test_subagent_manager.py::test_a_node_report_marks_an_empty_return_as_empty_not_absent tests/test_subagent_manager.py::test_a_node_report_bounds_the_output_it_marks tests/test_subagent_manager.py::test_a_stall_notice_marks_no_output tests/test_playbook_stint.py::TestWhoAnswersARoundsSuspendedNode -n 0 -x --tb=short --basetemp C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-native-rest\pytest --junitxml C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-native-rest\junit.xml

# base-cat; exit 0; cwd C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\base431599
uv run --frozen --no-sync pytest tests/test_subagent_dag_runner.py::test_a_foreground_call_returns_the_first_report_while_the_graph_keeps_running -n 0 -x --tb=short --basetemp C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-base-cat\pytest --junitxml C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-base-cat\junit.xml

# native-python; exit 0; cwd C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\pr1063\head
uv run --frozen --no-sync pytest tests/test_subagent_dag_runner.py::test_a_re_sent_report_still_carries_what_the_node_returned -p native_echo_fixture -n 0 -x --tb=short --basetemp C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-native-python\pytest --junitxml C:\Users\livxu\AppData\Local\Temp\raven-windows-review\93f210e71847404cb53894de7ac36f62\checks\1063-native-python\junit.xml

# cwd: isolated head checkout/ui-web; exit 0; 231 passed
npm.cmd ci --ignore-scripts --no-audit --no-fund
node node_modules/vitest/vitest.mjs run src/features/transcript/TranscriptPage.test.tsx src/state/session/pipeline.events.test.ts --reporter=default --reporter=json --outputFile.json="C:/Users/livxu/AppData/Local/Temp/raven-windows-review/93f210e71847404cb53894de7ac36f62/checks/1063-web/results.json"

For native-python, the external native_echo_fixture plugin replaces only the test helper's cat configuration with subprocess.list2cmdline([sys.executable, str(script)]); that UTF-8 script writes sys.stdin.read() to stdout.