llm: accept a directory argument, recursing for .md/.txt

README.md

@@ -321,10 +321,11 @@ ffmpeg -i talk.mp4 -f wav - | assembly transcribe -
 git log --oneline -30 | assembly llm "write release notes grouped by feature/fix"
 ```
 
-Pass files straight to `llm` instead of building the pipeline yourself — each is read, prefixed with a `===== name =====` header, and concatenated as the prompt's context (so the answer can cite which note it came from):
+Pass files — or a whole directory — straight to `llm` instead of building the pipeline yourself — each is read, prefixed with a `===== name =====` header, and concatenated as the prompt's context (so the answer can cite which note it came from). A directory argument recurses for its `.md`/`.txt` files:
 
 ```sh
 assembly llm "answer using only these notes: who owns the deploy?" notes/*.md
+assembly llm "summarize the key decisions" transcripts/
 ```
 
 ## 📚 Documentation

aai_cli/commands/llm/__init__.py

@@ -44,8 +44,8 @@ def _list_models(output_field: choices.TextOrJson | None, json_mode: bool) -> No
             ),
             ("Pipe any text in", 'echo "meeting notes" | assembly llm "turn into action items"'),
             (
-                "Read one or more files as context",
-                'assembly llm "answer using only these notes: who owns the deploy?" notes/*.md',
+                "Read files or a whole directory as context",
+                'assembly llm "summarize the key decisions" transcripts/',
             ),
             (
                 "Pick a model and add a system prompt",
@@ -60,8 +60,9 @@ def llm(
     prompt: str | None = typer.Argument(None, help="The prompt to send to the model"),
     files: list[Path] | None = typer.Argument(
         None,
-        help="Optional input files to read as the prompt's context (each is header-prefixed "
-        "with its name and concatenated; takes priority over piped stdin)",
+        help="Optional input files or directories to read as the prompt's context "
+        "(a directory recurses for .md/.txt files; each is header-prefixed with its "
+        "name and concatenated; takes priority over piped stdin)",
     ),
     # Note: text piped on stdin is injected into the prompt (e.g. `cat notes | assembly llm "summarize"`).
     model: str = typer.Option(

aai_cli/commands/llm/_exec.py

@@ -27,6 +27,9 @@
     '`assembly stream -o text | assembly llm -f "summarize action items as I talk"`.'
 )
 
+# Suffixes a directory argument recurses for (matched case-insensitively).
+_DIR_SUFFIXES = (".md", ".txt")
+
 
 @dataclass(frozen=True)
 class LlmOptions:
@@ -82,17 +85,46 @@ def _validate_follow_args(
     return prompt
 
 
+def _expand_paths(files: tuple[Path, ...]) -> tuple[Path, ...]:
+    """Expand any directory argument into the ``.md``/``.txt`` files it holds.
+
+    A directory recurses (depth-first via ``rglob``, path-sorted so the header
+    order is deterministic) for files whose suffix is in ``_DIR_SUFFIXES``; a plain
+    file passes through unchanged. A directory holding no matching files is a usage
+    error — the empty-check the justfile's ``transcripts/**/*.md(.N)`` glob used to
+    carry, now that the CLI can. This lets `assembly llm "…" transcripts/` stand in
+    for the glob-and-guard the caller would otherwise hand-roll.
+    """
+    expanded: list[Path] = []
+    for path in files:
+        if not path.is_dir():
+            expanded.append(path)
+            continue
+        matches = sorted(
+            p for p in path.rglob("*") if p.is_file() and p.suffix.lower() in _DIR_SUFFIXES
+        )
+        if not matches:
+            raise UsageError(
+                f"No {' or '.join(_DIR_SUFFIXES)} files found in {path}.",
+                suggestion="Point at a directory containing .md or .txt notes, "
+                "or pass files directly.",
+            )
+        expanded.extend(matches)
+    return tuple(expanded)
+
+
 def _read_files(files: tuple[Path, ...]) -> str:
     """Read each file and join them, each prefixed with a ``===== name =====`` header.
 
-    The header names each source (the file's stem) so a multi-file prompt can cite
-    which note an answer came from; it's applied uniformly, even for a single file,
-    so the format the model sees is predictable. A missing or unreadable path is a
-    usage error raised before any auth or network — the same fail-fast ordering as
-    the --transcript-id check.
+    A directory argument is first expanded into its ``.md``/``.txt`` files
+    (see ``_expand_paths``). The header names each source (the file's stem) so a
+    multi-file prompt can cite which note an answer came from; it's applied
+    uniformly, even for a single file, so the format the model sees is predictable.
+    A missing or unreadable path is a usage error raised before any auth or
+    network — the same fail-fast ordering as the --transcript-id check.
     """
     sections: list[str] = []
-    for path in files:
+    for path in _expand_paths(files):
         try:
             text = path.read_text(encoding="utf-8")
         except OSError as exc:

tests/__snapshots__/test_snapshots_help_run.ambr

@@ -592,9 +592,11 @@
 
   ╭─ Arguments ──────────────────────────────────────────────────────────────────╮
   │   prompt      [PROMPT]    The prompt to send to the model                    │
-  │   files       [FILES]...  Optional input files to read as the prompt's       │
-  │                           context (each is header-prefixed with its name and │
-  │                           concatenated; takes priority over piped stdin)     │
+  │   files       [FILES]...  Optional input files or directories to read as the │
+  │                           prompt's context (a directory recurses for         │
+  │                           .md/.txt files; each is header-prefixed with its   │
+  │                           name and concatenated; takes priority over piped   │
+  │                           stdin)                                             │
   ╰──────────────────────────────────────────────────────────────────────────────╯
   ╭─ Options ────────────────────────────────────────────────────────────────────╮
   │ --model                  TEXT                  LLM Gateway model             │
@@ -634,9 +636,8 @@
    $ assembly llm "summarize the key decisions" --transcript-id 5551234-abcd
    Pipe any text in
    $ echo "meeting notes" | assembly llm "turn into action items"
-   Read one or more files as context
-   $ assembly llm "answer using only these notes: who owns the deploy?"
-   notes/*.md
+   Read files or a whole directory as context
+   $ assembly llm "summarize the key decisions" transcripts/
    Pick a model and add a system prompt
    $ assembly llm "draft a follow-up email" --model claude-opus-4-7 --system "Be
    concise."

tests/test_llm_command.py

@@ -128,130 +128,6 @@ def fake_complete(api_key, *, model, messages, max_tokens, transcript_id=None, e
     assert seen["transcript_id"] is None
 
 
-def test_llm_reads_file_argument_as_context(monkeypatch, tmp_path):
-    _auth()
-    seen = {}
-
-    def fake_complete(api_key, *, model, messages, max_tokens, transcript_id=None, extra=None):
-        seen["content"] = messages[0]["content"]
-        seen["transcript_id"] = transcript_id
-        return _payload("done")
-
-    monkeypatch.setattr("aai_cli.commands.llm.gateway.complete", fake_complete)
-    note = tmp_path / "alpha.md"
-    note.write_text("bob owns the deploy")
-    result = runner.invoke(app, ["llm", "who owns the deploy?", str(note), "--json"])
-    assert result.exit_code == 0
-    # The file content is injected, under a header naming the file's stem.
-    assert "who owns the deploy?" in seen["content"]
-    assert "bob owns the deploy" in seen["content"]
-    assert "===== alpha =====" in seen["content"]
-    assert seen["transcript_id"] is None
-
-
-def test_llm_concatenates_multiple_files_with_headers_in_order(monkeypatch, tmp_path):
-    _auth()
-    seen = {}
-
-    def fake_complete(api_key, *, model, messages, max_tokens, transcript_id=None, extra=None):
-        seen["content"] = messages[0]["content"]
-        return _payload("done")
-
-    monkeypatch.setattr("aai_cli.commands.llm.gateway.complete", fake_complete)
-    first = tmp_path / "first.md"
-    first.write_text("ship friday")
-    second = tmp_path / "second.md"
-    second.write_text("freeze monday")
-    result = runner.invoke(app, ["llm", "summarize", str(first), str(second), "--json"])
-    assert result.exit_code == 0
-    content = seen["content"]
-    assert "===== first =====" in content
-    assert "===== second =====" in content
-    assert "ship friday" in content
-    assert "freeze monday" in content
-    # Both note bodies appear under their own header, in the order passed.
-    assert content.index("===== first =====") < content.index("===== second =====")
-    assert content.index("ship friday") < content.index("freeze monday")
-
-
-def test_llm_files_take_priority_over_stdin(monkeypatch, tmp_path):
-    _auth()
-    seen = {}
-
-    def fake_complete(api_key, *, model, messages, max_tokens, transcript_id=None, extra=None):
-        seen["content"] = messages[0]["content"]
-        return _payload("done")
-
-    monkeypatch.setattr("aai_cli.commands.llm.gateway.complete", fake_complete)
-    note = tmp_path / "note.md"
-    note.write_text("from the file")
-    result = runner.invoke(
-        app, ["llm", "summarize", str(note)], input="from stdin, should be ignored"
-    )
-    assert result.exit_code == 0
-    assert "from the file" in seen["content"]
-    assert "from stdin, should be ignored" not in seen["content"]
-    assert "Ignoring piped stdin; file arguments take priority." in result.output
-
-
-def test_llm_missing_file_exits_2_without_network(monkeypatch, tmp_path):
-    # A bad path (e.g. an unmatched shell glob passed through literally) is a usage
-    # error raised before auth or the gateway, not a crash.
-    _auth()
-    monkeypatch.setattr(
-        "aai_cli.commands.llm.gateway.complete",
-        lambda *a, **k: (_ for _ in ()).throw(AssertionError("must not call the gateway")),
-    )
-    missing = tmp_path / "nope.md"
-    result = runner.invoke(app, ["llm", "summarize", str(missing)])
-    assert result.exit_code == 2
-    assert "Couldn't read" in result.output
-    # The clean OS reason (errno's strerror) is shown, not the raw exception repr —
-    # so no "[Errno N] …: '/path'" bracket leaks into the message.
-    assert "[Errno" not in result.output
-
-
-def test_llm_files_with_terminal_stdin_emits_no_warning(monkeypatch, tmp_path):
-    # With files given and stdin a terminal (not piped), there's nothing being
-    # ignored, so the "Ignoring piped stdin" warning must not fire.
-    _auth()
-    monkeypatch.setattr("aai_cli.commands.llm._exec.stdio.stdin_is_piped", lambda: False)
-    seen = {}
-
-    def fake_complete(api_key, *, model, messages, max_tokens, transcript_id=None, extra=None):
-        seen["content"] = messages[0]["content"]
-        return _payload("done")
-
-    monkeypatch.setattr("aai_cli.commands.llm.gateway.complete", fake_complete)
-    note = tmp_path / "note.md"
-    note.write_text("only the file")
-    result = runner.invoke(app, ["llm", "summarize", str(note)])
-    assert result.exit_code == 0
-    assert "only the file" in seen["content"]
-    assert "Ignoring piped stdin" not in result.output
-
-
-def test_llm_transcript_id_takes_priority_over_files(monkeypatch, tmp_path):
-    _auth()
-    seen = {}
-    # Pin stdin to a terminal so only the file argument is the ignored source.
-    monkeypatch.setattr("aai_cli.commands.llm._exec.stdio.stdin_is_piped", lambda: False)
-
-    def fake_complete(api_key, *, model, messages, max_tokens, transcript_id=None, extra=None):
-        seen["content"] = messages[0]["content"]
-        seen["transcript_id"] = transcript_id
-        return _payload("s")
-
-    monkeypatch.setattr("aai_cli.commands.llm.gateway.complete", fake_complete)
-    note = tmp_path / "note.md"
-    note.write_text("file content here")
-    result = runner.invoke(app, ["llm", "summarize", str(note), "--transcript-id", "t_9"])
-    assert result.exit_code == 0
-    assert seen["transcript_id"] == "t_9"
-    assert "file content here" not in seen["content"]
-    assert "Ignoring file arguments; --transcript-id takes priority." in result.output
-
-
 def test_llm_transcript_id_takes_priority_over_stdin(monkeypatch):
     _auth()
     seen = {}