markdown-board-duckdb-migration

Answer

1. Namespace basename bug (home → project key). The old code derived the namespace from the home directory basename, collapsing every repo into one shared key. The fix resolves the namespace from the project identity (CLI override → board's project: header → git root basename → cwd):

def parse_project(md: str) -> str | None:
    m = re.search(r"^project\s*:\s*(\S+)\s*$", md, flags=re.MULTILINE)
    return m.group(1) if m else None

def project_namespace(project: str | None) -> str:
    if project:
        return project
    git_root = subprocess.run(["git", "rev-parse", "--show-toplevel"],
                              capture_output=True, text=True)
    if git_root.returncode == 0 and git_root.stdout.strip():
        return Path(git_root.stdout.strip()).name
    return Path.cwd().name

2. cooldown_s aligned with the live scheduler. The markdown board says 60; the live scheduler (live_state.json) says 300. The migration overrides stale board values with live ground truth:

def apply_live_overrides(rows, live):
    sched = live.get("scheduler", {})
    default_cooldown = sched.get("cooldown_s")
    for row in rows:
        state = sched.get("tasks", {}).get(row["task_id"], {})
        if default_cooldown is not None:
            row["cooldown_s"] = int(default_cooldown)   # live wins over markdown
        if "status" in state and state["status"] != row["status"]:
            row["status"] = state["status"]             # 🟡 in_progress -> 🟢 complete
    return rows

3. 🟡 in_progress → complete. T-002 ("migrate board to duckdb") is x under 🟡 in the matrix, but the scheduler reports it complete → migration marks it complete.

4. Two-commit cutover. Commit 1 ships migration + board/*.parquet + .gitignore (board.db, *.db); board.db is a gitignored thin DuckDB file with CREATE VIEW tasks/events AS SELECT * FROM read_parquet(...). After commit 1's hash is known, --backfill-commit <hash> stamps it into both parquet files → commit 2:

def backfill_commit(commit_hash):
    for path in (TASKS_PARQUET, EVENTS_PARQUET):
        frame = pq.read_table(path).set_column(
            frame.schema.get_field_index("commit_hash"),
            pa.field("commit_hash", pa.string()),
            pa.array([commit_hash] * frame.num_rows, type=pa.string()))
        write_parquet_if_changed(frame, path)

5. Determinism / idempotency guards (needed so a stray re-run can't dirty the tree between the two commits): timestamps come from the board's updated: header (not wall clock), and write_parquet_if_changed skips writes when existing.equals(frame); re-migration preserves an already-backfilled commit_hash.