The implemented bounded research workflow lets a human approve a fixed packet, topic, expiry and question/read budget. Your assistant can then ask varied questions within that scope, keep the discussion across restarts, and save a draft citing sources and Muse observations. Changed data or scope needs new consent. This has synthetic test coverage; it does not establish real-account LIVE success.

Local offline walkthrough

Run from the repository root after uv sync --python 3.12. This synthetic walkthrough collects data, previews a packet, opens a fixture Muse discussion, asks a question and a follow-up in the same discussion, saves host-written text and exports five formats. The host shows the human each E_SCOPE_REQUIRED approval preview. Only the human runs operator permit approve, reviews its stderr (including expiry and uses), and repeats the unchanged call with the returned grant. The agent never approves. Evidence is OFFLINE/FIXTURE, with no network or accounts.

These Bash blocks run in order. Python only edits configuration and copies public JSON fields; every product operation uses the shipped CLI. tests/test_readme_walkthrough.py executes these exact blocks in a temporary workspace.

export UV_OFFLINE=1
mkdir -p "$HOME/agent-work/tmp"
export SLICE_DIR="${SLICE_DIR:-$(mktemp -d "$HOME/agent-work/tmp/metamuse-at042-XXXXXX")}"
export SLICE_WS="$SLICE_DIR/workspace"
mkdir "$SLICE_DIR/inputs"
mm() { uv run metamuse --workspace "$SLICE_WS" --json --non-interactive "$@"; }
operator() { uv run metamuse --workspace "$SLICE_WS" --json "$@"; }
field() { uv run python -c 'import json,sys; print(json.load(open(sys.argv[1]))["data"][sys.argv[2]])' "$1" "$2"; }
approval_field() { uv run python -c 'import json,sys; print(json.load(open(sys.argv[1]))["approval"][sys.argv[2]])' "$1" "$2"; }
preview() {
  local output="$1"; shift
  local result=0
  mm "$@" > "$output" || result=$?
  test "$result" -eq 3
  uv run python -c 'import json,sys; e=json.load(open(sys.argv[1])); assert e["error"]["code"] == "E_SCOPE_REQUIRED" and e["approval"]; print(e["approval"]["summary"]); print(e["approval"]["next_action"])' "$output"
}
save_request() {
  uv run python -c 'import json,sys; a=json.load(open(sys.argv[1]))["approval"]; json.dump(a["request"],open(a["request_file"],"w"),ensure_ascii=False)' "$1"
}

mm init > "$SLICE_DIR/init.json"
uv run python - <<'PYTHON'
import json, os
from pathlib import Path
root = Path(os.environ['SLICE_DIR'])
config = Path(os.environ['SLICE_WS']) / 'metamuse.toml'
text = config.read_text().replace('[app]\n', '[app]\nimport_roots = ' + json.dumps([str(root / 'inputs')]) + '\n')
text = text.replace('[muse]\nenabled = false', '[muse]\nenabled = true\nbackend = "fixture"')
config.write_text(text)
corpus = Path('tests/fixtures/corpus/import/comments_tree.jsonl')
(root / 'inputs/comments.jsonl').write_text('\n'.join(corpus.read_text().splitlines()[:3]) + '\n')
PYTHON
mm doctor > "$SLICE_DIR/doctor.json"
mm project create --name 'Synthetic slice' --request-key project > "$SLICE_DIR/project.json"
export SLICE_PROJECT=$(field "$SLICE_DIR/project.json" id)
mm add --project "$SLICE_PROJECT" --file "$SLICE_DIR/inputs/comments.jsonl" --kind jsonl --request-key import --run > "$SLICE_DIR/import.json"
mm search --project "$SLICE_PROJECT" --query '這是' > "$SLICE_DIR/search.json"
uv run python - <<'PYTHON'
import json, os
from pathlib import Path
root = Path(os.environ['SLICE_DIR'])
hits = json.loads((root / 'search.json').read_text())['data']['results']
(root / 'inputs/selection.json').write_text(json.dumps([hit['ref'] for hit in hits]))
PYTHON
mm packet prepare --project "$SLICE_PROJECT" --selection-file "$SLICE_DIR/inputs/selection.json" --request-key packet > "$SLICE_DIR/packet.json"
export SLICE_PACKET=$(field "$SLICE_DIR/packet.json" id)
mm packet preview "$SLICE_PACKET" > "$SLICE_DIR/packet-preview.json"
cat "$SLICE_DIR/packet-preview.json"

Call without a grant, show the review and its human command, save the public request, then approve the exact digest. Credential-like values are redacted separately in every text field. A redacted or edited request fails the original digest; it must be reviewed again. Expired, revoked, exhausted or mismatched grants return E_SCOPE_DENIED without a preview. Repeat without a grant to get a fresh review.

preview "$SLICE_DIR/open-preview.json" muse open --project "$SLICE_PROJECT" --topic 'Synthetic activity review' --request-key open
save_request "$SLICE_DIR/open-preview.json"
operator permit approve --operation muse.open --request-file "$(approval_field "$SLICE_DIR/open-preview.json" request_file)" --exact-digest "$(approval_field "$SLICE_DIR/open-preview.json" action_digest)" --request-key open-grant > "$SLICE_DIR/open-grant.json"
mm muse open --project "$SLICE_PROJECT" --topic 'Synthetic activity review' --access-grant "$(field "$SLICE_DIR/open-grant.json" id)" --request-key open > "$SLICE_DIR/open.json"
export SLICE_DISCUSSION=$(field "$SLICE_DIR/open.json" discussion_id)
mm jobs get "$(field "$SLICE_DIR/open.json" job_id)"
mm worker run --once
mm jobs get "$(field "$SLICE_DIR/open.json" job_id)"

printf '%s' 'List evidence and gaps.' > "$SLICE_DIR/inputs/question-1.txt"
printf '%s' 'Which missing replies would change those conclusions?' > "$SLICE_DIR/inputs/question-2.txt"
for exchange in 1 2; do
  preview "$SLICE_DIR/send-preview-$exchange.json" muse send --discussion "$SLICE_DISCUSSION" --packet "$SLICE_PACKET" --question-file "$SLICE_DIR/inputs/question-$exchange.txt" --request-key "send-$exchange"
  mm packet preview "$SLICE_PACKET"
  save_request "$SLICE_DIR/send-preview-$exchange.json"
  operator permit approve --operation muse.send --request-file "$(approval_field "$SLICE_DIR/send-preview-$exchange.json" request_file)" --exact-digest "$(approval_field "$SLICE_DIR/send-preview-$exchange.json" action_digest)" --request-key "share-grant-$exchange" > "$SLICE_DIR/share-grant-$exchange.json"
  mm muse send --discussion "$SLICE_DISCUSSION" --packet "$SLICE_PACKET" --question-file "$SLICE_DIR/inputs/question-$exchange.txt" --share-grant "$(field "$SLICE_DIR/share-grant-$exchange.json" id)" --request-key "send-$exchange" > "$SLICE_DIR/send-$exchange.json"
  export SLICE_EXCHANGE=$(field "$SLICE_DIR/send-$exchange.json" exchange_id)
  mm jobs get "$(field "$SLICE_DIR/send-$exchange.json" job_id)"
  mm worker run --once
  mm jobs get "$(field "$SLICE_DIR/send-$exchange.json" job_id)"
  preview "$SLICE_DIR/read-preview-$exchange.json" muse read --exchange "$SLICE_EXCHANGE" --refresh
  save_request "$SLICE_DIR/read-preview-$exchange.json"
  operator permit approve --operation muse.read --request-file "$(approval_field "$SLICE_DIR/read-preview-$exchange.json" request_file)" --exact-digest "$(approval_field "$SLICE_DIR/read-preview-$exchange.json" action_digest)" --request-key "read-grant-$exchange" > "$SLICE_DIR/read-grant-$exchange.json"
  mm muse read --exchange "$SLICE_EXCHANGE" --refresh --access-grant "$(field "$SLICE_DIR/read-grant-$exchange.json" id)" > "$SLICE_DIR/refresh-$exchange.json"
  mm worker run --once
  mm muse read --exchange "$SLICE_EXCHANGE" > "$SLICE_DIR/read-$exchange.json"
  mm muse close-exchange "$SLICE_EXCHANGE" --outcome needs_more_data --request-key "close-$exchange" > "$SLICE_DIR/close-$exchange.json"
done

The fixture replies are progress, so the operator closes each exchange as needs_more_data. Draft text comes from the host and cites observations from both exchanges. CLI invocations have already restarted the application between steps; the separate MCP acceptance test also restarts a real stdio server and checks that no send repeats.

uv run python - <<'PYTHON'
import json, os
from pathlib import Path
root = Path(os.environ['SLICE_DIR'])
facts = json.loads((root / 'inputs/selection.json').read_text())
citations = []
for exchange in (1, 2):
    read = json.loads((root / f'read-{exchange}.json').read_text())['data']
    assistant_ids = {message['id'] for message in read['messages'] if message['role'] == 'assistant'}
    observed = next(o for o in read['observations'] if o['message_id'] in assistant_ids and o['raw_event'].get('payload',{}).get('role') == 'assistant')
    assert read['mock'] and read['research_complete'] is False
    facts.append(observed['evidence_ref'])
    citations.append(observed['id'])
(root / 'inputs/facts.json').write_text(json.dumps(facts))
host_text = '=活動回顧（host 撰寫）\r\n首輪列出缺口；追問指出仍需要留言與子回覆。😀 e\u0301\n'
(root / 'inputs/draft.txt').write_bytes((host_text + '\n引用：' + ', '.join(citations) + '\n').encode())
PYTHON
mm draft save --project "$SLICE_PROJECT" --text-file "$SLICE_DIR/inputs/draft.txt" --platform generic --facts-file "$SLICE_DIR/inputs/facts.json" --request-key draft > "$SLICE_DIR/draft.json"
export SLICE_DRAFT=$(field "$SLICE_DIR/draft.json" draft_id)
mm draft preview "$SLICE_DRAFT" --revision 1 > "$SLICE_DIR/preview.json"
uv run python - <<'PYTHON'
import json, os
from pathlib import Path
(Path(os.environ['SLICE_DIR']) / 'inputs/export-refs.json').write_text(json.dumps([{'kind':'draft_revision','id':os.environ['SLICE_DRAFT'],'revision':1}]))
PYTHON
for format in bundle jsonl markdown text csv; do
  mm export --project "$SLICE_PROJECT" --refs-file "$SLICE_DIR/inputs/export-refs.json" --format "$format" --request-key "export-$format" > "$SLICE_DIR/export-$format.json"
done
cat "$SLICE_DIR/preview.json" "$SLICE_DIR/export-bundle.json"
if [ "${SLICE_KEEP:-0}" != 1 ]; then rm -rf -- "$SLICE_DIR"; fi