Add --save-dir finalization: auto-name, note, and sidecar

REFERENCE.md

@@ -116,3 +116,26 @@ object per dataset (not NDJSON; a single dataset is therefore one object):
 the row's `llm` key (the WER score still uses the raw transcript), and
 `--llm-reduce` runs one prompt over every item's result and adds a top-level
 `reduce` (`{"model","prompts","output"}`) to the object.
+
+## Recording streams to disk
+
+`assembly stream --save-dir DIR` auto-names a capture under `DIR/YYYY-MM-DD/`
+with a timestamped stem (`YYYY-MM-DD-HHMMSS[-slug]`) shared across every file it
+writes:
+
+- `<stem>.txt` — the transcript, one finalized turn per line (flushed live).
+- `<stem>.wav` — the recorded audio, 16-bit mono PCM. Suppress it with
+  `--no-save-audio` to keep only the text. Under `--system-audio` the two channels
+  can't share a file, so each gets its own `<stem>-you.wav` / `<stem>-system.wav`.
+- `<stem>.md` — written when `--llm "…"` is also passed: the final answer of the
+  live prompt chain, captured as a note next to the transcript.
+- `<stem>.aai.json` — a metadata sidecar so a list/browse UI needs no transcript
+  parsing: `{"title", "date", "duration_seconds", "speakers", "turns",
+  "transcript", "audio", "note"}`. `audio` is the list of WAV file names (empty
+  under `--no-save-audio`, two entries under `--system-audio`); `note` is `null`
+  when no `--llm` note was written.
+
+`--name "Title"` slugs an explicit title into the stem; `--auto-name` instead
+derives that title from the transcript via the LLM Gateway once the stream ends,
+renaming the files to match (the timestamp stem is kept if the title is empty).
+The two are mutually exclusive.

aai_cli/commands/agent/_exec.py

@@ -24,8 +24,8 @@
 from aai_cli.app.context import AppState
 from aai_cli.core import choices, client, errors, signals
 from aai_cli.core.errors import UsageError
-from aai_cli.streaming.session import resolve_output_modes
 from aai_cli.streaming.sources import FileSource
+from aai_cli.streaming.validate import resolve_output_modes
 from aai_cli.ui import output
 
 

aai_cli/commands/agent_cascade/_exec.py

@@ -25,8 +25,8 @@
 from aai_cli.core import choices, client, config_builder, errors, llm, signals
 from aai_cli.core.errors import UsageError
 from aai_cli.streaming import turn_presets
-from aai_cli.streaming.session import resolve_output_modes
 from aai_cli.streaming.sources import FileSource
+from aai_cli.streaming.validate import resolve_output_modes
 from aai_cli.tts import session as tts_session
 from aai_cli.ui import output
 

aai_cli/commands/dictate/_exec.py

@@ -19,7 +19,7 @@
 from aai_cli.core.config_builder import split_csv
 from aai_cli.core.hotkey import CTRL_C, CTRL_D, ESC, TerminalKeys
 from aai_cli.core.microphone import MicrophoneSource
-from aai_cli.streaming.session import resolve_output_modes
+from aai_cli.streaming.validate import resolve_output_modes
 from aai_cli.ui import output
 
 # Capture is resampled to one rate the Sync API accepts; 16 kHz mono PCM16 keeps

aai_cli/commands/stream/__init__.py

@@ -39,6 +39,10 @@
                 "Auto-name the transcript + WAV under a dir",
                 'assembly stream --save-dir ~/recordings --name "Standup"',
             ),
+            (
+                "Name from content + save a summary note",
+                'assembly stream --save-dir ~/recordings --auto-name --llm "summarize as a note"',
+            ),
             (
                 "Boost domain terms with keyterm prompts",
                 'assembly stream --keyterms-prompt "AssemblyAI" --keyterms-prompt "Claude"',
@@ -121,6 +125,18 @@ def stream(
         help="Title to slug into the --save-dir filename (e.g. a meeting title)",
         rich_help_panel=help_panels.OPT_SAVING,
     ),
+    auto_name: bool = typer.Option(
+        False,
+        "--auto-name",
+        help="With --save-dir, derive the filename from the transcript via the LLM",
+        rich_help_panel=help_panels.OPT_SAVING,
+    ),
+    no_save_audio: bool = typer.Option(
+        False,
+        "--no-save-audio",
+        help="With --save-dir, skip the WAV and save only the transcript",
+        rich_help_panel=help_panels.OPT_SAVING,
+    ),
     # model & input
     speech_model: SpeechModel = typer.Option(
         DEFAULT_SPEECH_MODEL,
@@ -398,5 +414,7 @@ def stream(
         save_transcript=save_transcript,
         save_dir=save_dir,
         name=name,
+        auto_name=auto_name,
+        no_save_audio=no_save_audio,
     )
     run_with_options(ctx, stream_exec.run_stream, opts, json=json_out)

aai_cli/commands/stream/_exec.py

@@ -23,18 +23,14 @@
 from aai_cli.core import choices, client, config_builder, signals, stdio, youtube
 from aai_cli.core.errors import UsageError, mutually_exclusive
 from aai_cli.core.microphone import MicrophoneSource
-from aai_cli.streaming import naming, record, transcript, turn_presets
+from aai_cli.streaming import naming, record, savedir, transcript, turn_presets
+from aai_cli.streaming.batch import stream_batch_sources
 from aai_cli.streaming.macos import MacSystemAudioSource
 from aai_cli.streaming.render import StreamRenderer
-from aai_cli.streaming.session import (
-    SourceOptions,
-    StreamSession,
-    resolve_output_modes,
-    stream_batch_sources,
-    validate_sources,
-)
+from aai_cli.streaming.session import StreamSession
 from aai_cli.streaming.sources import TARGET_RATE, FileSource, StdinSource
 from aai_cli.streaming.turn_presets import TurnDetectionPreset
+from aai_cli.streaming.validate import SourceOptions, resolve_output_modes, validate_sources
 from aai_cli.ui import output
 from aai_cli.ui.follow import FollowRenderer
 
@@ -90,6 +86,8 @@ class StreamOptions:
     save_transcript: Path | None
     save_dir: Path | None
     name: str | None
+    auto_name: bool
+    no_save_audio: bool
 
     def source_options(self) -> SourceOptions:
         """The audio-input subset, in the shape the validation/dispatch helpers read."""
@@ -205,57 +203,97 @@ class SaveTargets:
     ``audio`` tees a single source to one WAV; ``audio_by_label`` instead maps each
     parallel ``--system-audio`` channel ("you", "system") to its own WAV when the two
     streams can't share a file. At most one of the two is set; ``transcript`` is the
-    single shared transcript either way.
+    single shared transcript either way. ``plan`` is set only under ``--save-dir`` and
+    carries the post-stream finalization (auto-name rename, ``--llm`` note, sidecar).
     """
 
     transcript: Path | None = None
     audio: Path | None = None
     audio_by_label: dict[str, Path] | None = None
+    plan: savedir.SaveDirPlan | None = None
+
+
+def _save_dir_targets(opts: StreamOptions, sources: SourceOptions, save_dir: Path) -> SaveTargets:
+    """Resolve ``--save-dir`` into auto-named targets plus the finalization plan.
+
+    ``--save-dir`` owns filename assembly, so it rejects the explicit
+    ``--save-audio``/``--save-transcript`` paths and the conflicting ``--name``/
+    ``--auto-name`` title pair. Two parallel ``--system-audio`` streams can't tee to one
+    WAV, so each channel gets its own ``<stem>-{you,system}.wav`` (one shared transcript);
+    ``--no-save-audio`` drops the WAV(s) entirely.
+    """
+    mutually_exclusive(
+        ("--save-dir", True),
+        ("--save-audio", opts.save_audio is not None),
+        ("--save-transcript", opts.save_transcript is not None),
+        suggestion="--save-dir names the files for you; drop the explicit path.",
+    )
+    mutually_exclusive(
+        ("--name", opts.name is not None),
+        ("--auto-name", opts.auto_name),
+        suggestion="Both set the title — pass --name for an explicit one or "
+        "--auto-name to derive it from the transcript.",
+    )
+    # Local wall-clock time (what a meeting filename wants); the explicit utc-then-
+    # astimezone keeps the now() call timezone-aware for the linter.
+    now = datetime.now(UTC).astimezone()
+    plan = savedir.SaveDirPlan(
+        save_dir=save_dir,
+        now=now,
+        name=opts.name,
+        auto_name=opts.auto_name,
+        write_note=bool(opts.llm_prompt),
+    )
+    paths = plan.paths
+    naming.ensure_dir(paths.directory)
+    if opts.no_save_audio:
+        # Transcript + sidecar (+ note) only; no WAV teed for any source.
+        return SaveTargets(transcript=paths.transcript, plan=plan)
+    if sources.system_audio:
+        # Parallel mic + system: one WAV per channel beside the shared transcript.
+        return SaveTargets(
+            transcript=paths.transcript,
+            audio_by_label={
+                "you": naming.channel_audio(paths.audio, "you"),
+                "system": naming.channel_audio(paths.audio, "system"),
+            },
+            plan=plan,
+        )
+    if sources.system_audio_only:
+        # A lone system-audio stream; label its single WAV so it reads like the pair.
+        return SaveTargets(
+            transcript=paths.transcript,
+            audio=naming.channel_audio(paths.audio, "system"),
+            plan=plan,
+        )
+    return SaveTargets(transcript=paths.transcript, audio=paths.audio, plan=plan)
 
 
 def _resolve_save_targets(opts: StreamOptions, sources: SourceOptions) -> SaveTargets:
     """Resolve the save flags into the destinations the session writes.
 
-    ``--save-dir`` owns filename assembly — it auto-names the transcript and a matching
-    WAV under ``DIR/YYYY-MM-DD/`` — so it can't be combined with the explicit
-    ``--save-audio``/``--save-transcript`` paths, and ``--name`` only feeds that assembly.
-    Two parallel ``--system-audio`` streams can't tee to one WAV, so under ``--save-dir``
-    each channel gets its own ``<stem>-{you,system}.wav`` (one shared transcript), and the
-    explicit single-path ``--save-audio`` is rejected outright.
+    ``--save-dir`` owns filename assembly (see ``_save_dir_targets``); the explicit
+    ``--save-audio``/``--save-transcript`` paths are the fallback, with the save-dir-only
+    ``--name``/``--auto-name``/``--no-save-audio`` flags rejected outside it.
     """
     if opts.save_dir is not None:
-        mutually_exclusive(
-            ("--save-dir", True),
-            ("--save-audio", opts.save_audio is not None),
-            ("--save-transcript", opts.save_transcript is not None),
-            suggestion="--save-dir names the files for you; drop the explicit path.",
-        )
-        # Local wall-clock time (what a meeting filename wants); the explicit utc-then-
-        # astimezone keeps the now() call timezone-aware for the linter.
-        now = datetime.now(UTC).astimezone()
-        paths = naming.resolve(opts.save_dir, opts.name, now=now)
-        naming.ensure_dir(paths.transcript.parent)
-        if sources.system_audio:
-            # Parallel mic + system: one WAV per channel beside the shared transcript.
-            return SaveTargets(
-                transcript=paths.transcript,
-                audio_by_label={
-                    "you": naming.channel_audio(paths.audio, "you"),
-                    "system": naming.channel_audio(paths.audio, "system"),
-                },
-            )
-        if sources.system_audio_only:
-            # A lone system-audio stream; label its single WAV so it reads like the pair.
-            return SaveTargets(
-                transcript=paths.transcript, audio=naming.channel_audio(paths.audio, "system")
-            )
-        return SaveTargets(transcript=paths.transcript, audio=paths.audio)
+        return _save_dir_targets(opts, sources, opts.save_dir)
     if opts.name is not None:
         raise UsageError(
             "--name applies only with --save-dir.",
             suggestion="Pass --save-dir DIR to auto-name the files, "
             "or --save-transcript PATH for an explicit path.",
         )
+    if opts.auto_name:
+        raise UsageError(
+            "--auto-name applies only with --save-dir.",
+            suggestion="Pass --save-dir DIR so there's an auto-named file to title.",
+        )
+    if opts.no_save_audio:
+        raise UsageError(
+            "--no-save-audio applies only with --save-dir.",
+            suggestion="Omit --save-audio to skip the WAV, or pass --save-dir DIR.",
+        )
     if opts.save_audio is not None:
         if sources.system_audio:
             raise UsageError(
@@ -343,6 +381,8 @@ def _collect_batch_sources(opts: StreamOptions, *, text_mode: bool) -> list[str]
         ("--save-transcript", opts.save_transcript is not None),
         ("--save-dir", opts.save_dir is not None),
         ("--name", opts.name is not None),
+        ("--auto-name", opts.auto_name),
+        ("--no-save-audio", opts.no_save_audio),
         suggestion="--from-stdin streams many sources; saving applies to a single run.",
     )
     mutually_exclusive(
@@ -434,6 +474,7 @@ def run_stream(opts: StreamOptions, state: AppState, *, json_mode: bool) -> None
         save_audio=targets.audio,
         save_audio_by_label=targets.audio_by_label,
         save_transcript=targets.transcript,
+        save_plan=targets.plan,
         llm_interval=opts.llm_interval,
     )
     with signals.terminate_as_interrupt():

aai_cli/streaming/batch.py

@@ -0,0 +1,107 @@
+"""Drive a ``assembly stream --from-stdin`` list of sources, one realtime session each.
+
+The realtime API is one session at a time, so a list of files/URLs (read on stdin,
+one per line) streams sequentially. This lives beside ``StreamSession`` rather than
+inside it: a session owns *one* run, while this owns the sequence — fresh session per
+source, per-source failure accounting, and the batch-wide Ctrl-C/pipe handling.
+"""
+
+from __future__ import annotations
+
+from collections.abc import Callable, Iterable
+
+import typer
+
+from aai_cli.core.errors import CANCELLED_EXIT_CODE, CLIError, NotAuthenticated
+from aai_cli.streaming.render import StreamRenderer
+from aai_cli.streaming.session import StreamSession
+from aai_cli.ui import output
+
+# A batch source string resolved to its real-time audio chunks and declared rate.
+_OpenedSource = tuple[Iterable[bytes], int]
+
+
+def _stream_source(
+    source: str,
+    *,
+    index: int,
+    total: int,
+    make_session: Callable[[], StreamSession],
+    open_source: Callable[[str], _OpenedSource],
+    renderer: StreamRenderer,
+    json_mode: bool,
+) -> bool:
+    """Stream one batch source in its own session; return True when it failed.
+
+    A ``CLIError`` (bad path, missing ffmpeg, decode failure) is recorded as a warning
+    so the batch carries on — except ``NotAuthenticated``, which re-raises to abort the
+    whole batch (one rejected key fails every source identically, and auto-login should
+    trigger once).
+    """
+    renderer.source(source, index=index, total=total)
+    try:
+        audio, rate = open_source(source)
+        # handle_interrupt=False: let a Ctrl-C/pipe close bubble to the batch loop below so
+        # one interrupt stops the whole sequence. Flipping it to True is behavior-equivalent
+        # here (the session would convert the same interrupt to the same Exit(130)/Exit(0)),
+        # so no test can distinguish it.
+        make_session().run(audio, rate, handle_interrupt=False)  # pragma: no mutate
+    except NotAuthenticated:
+        raise
+    except CLIError as exc:
+        # Flatten newlines so a crafted path/URL can't inject extra log lines (CR/LF).
+        detail = f"{source}: {exc.message}".replace("\n", " ").replace("\r", " ")
+        output.emit_warning(detail, json_mode=json_mode)
+        return True
+    else:
+        return False
+
+
+def stream_batch_sources(
+    sources: list[str],
+    *,
+    make_session: Callable[[], StreamSession],
+    open_source: Callable[[str], _OpenedSource],
+    renderer: StreamRenderer,
+    json_mode: bool,
+) -> None:
+    """Stream each source in ``sources`` in turn — the ``assembly stream --from-stdin``
+    batch mode.
+
+    The realtime API is one session at a time, so a list of files/URLs streams
+    sequentially: each source gets a fresh ``StreamSession`` from ``make_session`` (its
+    own transcript and ``--llm`` chain state) via ``_stream_source``.
+
+    A Ctrl-C stops the batch with the cancel code (exit 130); a closed downstream pipe
+    stops it quietly (exit 0). When any source failed, raises a ``CLIError`` at the end
+    so a script can trust the exit code.
+    """
+    total = len(sources)
+    failures = 0
+    try:
+        for index, source in enumerate(sources, start=1):
+            failures += _stream_source(
+                source,
+                index=index,
+                total=total,
+                make_session=make_session,
+                open_source=open_source,
+                renderer=renderer,
+                json_mode=json_mode,
+            )
+    except KeyboardInterrupt:
+        # One Ctrl-C stops the whole batch, not just the current source. Exit 130
+        # (cancel) so the interrupt isn't mistaken for a clean run of every source.
+        renderer.stopped()
+        raise typer.Exit(code=CANCELLED_EXIT_CODE) from None
+    except BrokenPipeError:
+        # Downstream consumer (e.g. `| head`) closed the pipe; stop quietly.
+        raise typer.Exit(code=0) from None
+    finally:
+        renderer.close()
+    if failures:
+        raise CLIError(
+            f"{failures} of {total} sources failed.",
+            error_type="batch_failed",
+            suggestion="Check each failed path or URL, then re-run.",
+        )