docs(schema): noodle about a schema

Amolith and Crush created

Had a combo of Claude Code/Sonnet 4.5 and Amp/GPT-5 (Oracle) write one
schema each, then used CC and Crush/GPT-5 (high) to polish up CC's
version. Then I presented both schemas to CC and Crush/GPT-5 and asked
which they prefer. Both suggested copying some choice ideas from CC's
schema, but instead adopting Amp's schema overall.

Co-authored-by: Crush <crush@charm.land>

Change summary

docs/Potential schema.md | 322 ++++++++++++++++++++++++++++++++++++++++++
1 file changed, 322 insertions(+)

Detailed changes

docs/Potential schema.md 🔗

@@ -0,0 +1,322 @@
+<!--
+SPDX-FileCopyrightText: Amolith <amolith@secluded.site>
+
+SPDX-License-Identifier: CC0-1.0
+-->
+
+# BadgerDB Schema Design for nasin pali
+
+## TL;DR
+
+- Use a single BadgerDB with a small, hierarchical keyspace: dir/…, idx/…, and s/{sid}/… (per-session).
+- Store session documents (goal, tasks, meta) under s/{sid}/…, append-only events under s/{sid}/evt/{seq}, and small secondary indexes for fast lookups (active session by dir, tasks by status, active-session set).
+- Rely on prefix scans and Badger subscriptions over these prefixes to drive the TUI in real time. Keep values JSON for readability; keep index values empty when possible; small payloads OK for convenience.
+
+## Recommended approach (simple path)
+
+### Keyspace layout (prefixes and values)
+
+#### Schema/version
+- meta/schema_version -> "1"
+
+#### Working directory → active session lookup (with parent-walk)
+- dir/{dir_hash}/active -> {sid}
+  - Path canonicalization:
+    - Start with current working directory
+    - Convert to absolute path
+    - Resolve all symlinks (EvalSymlinks)
+    - Normalize path separators to forward slashes
+    - On Windows, always case-fold to lowercase; on other OSes, do not case-fold
+    - Result: canonical_path
+  - {dir_hash} = full blake3-256(canonical_path), hex-encoded lowercase (64 hex chars)
+  - This prevents issues with symlinks, bind mounts, and cross-platform path differences
+  - Value is the current active {sid} (string)
+  - Parent-walk resolution: from canonical_path, compute {dir_hash} and check dir/{dir_hash}/active; if absent, ascend to the parent directory and repeat until root; first match wins
+  - Symlinks are resolved during canonicalization via EvalSymlinks; cycles are prevented by EvalSymlinks itself and an additional visited set for safety during parent-walk
+
+#### Active sessions set (for fast listing + subscribe)
+- idx/active/{sid} -> {dir_hash} (or tiny JSON summary later if desired)
+
+#### Archived sessions (queryable by time and by dir)
+- idx/archived/{ts_be}/{sid} -> {dir_hash}          // for global archive lists in chronological order
+- dir/{dir_hash}/archived/{ts_be} -> {sid}          // for "archive history" per working directory
+  - {ts_be} = 8-byte big-endian Unix nanos, hex-encoded lowercase, zero-padded to 16 hex chars; sorts chronologically
+
+#### Per-session namespace
+- s/{sid}/meta -> JSON
+  - { sid, dir_path, dir_hash, state: "active"|"archived", created_at, archived_at: null|ts, last_updated_at }
+- s/{sid}/goal -> JSON
+  - { title, description, updated_at }
+- s/{sid}/task/{task_id} -> JSON
+  - { id, title, description, status: "pending"|"in_progress"|"completed"|"failed"|"cancelled", created_at, updated_at, created_seq }
+  - {task_id} = first 6 hex chars of blake3(normalize(title)+'|'+normalize(description)+'|'+sid); scoped per session, so cross-session collisions are irrelevant
+- s/{sid}/idx/status/{status}/{task_id} -> "" (empty value; presence = membership)
+  - Maintained atomically on task create/status update/delete
+- s/{sid}/meta/evt_seq -> 8-byte big-endian counter stored as raw bytes (monotonic per session)
+- s/{sid}/evt/{seq_be} -> JSON event record
+  - {seq_be} = 8-byte big-endian u64 counter, hex-encoded lowercase, zero-padded to 16 hex chars; assures correct chronological iteration
+  - { seq, at, type, reason: null|string, cmd: "np ...", payload: {...} }
+  - Types you'll likely use: session_started, goal_set, goal_updated, task_added, task_updated, task_status_changed, session_archived, note
+
+### Core operations (transactional)
+
+#### Start session (np s)
+1) Canonicalize working directory using the process described in the "Working directory → active session lookup" section.
+2) Compute dir_hash from canonical_path.
+3) Check if dir/{dir_hash}/active exists:
+   - If it exists, read the active {sid} and print: "Session {sid} is already active for this directory. There's already an active session for this directory; ask your operator whether they want to resume or archive it."
+   - Return 0 (idempotent operation).
+4) If no active session exists, begin txn:
+   - Generate new sid = ULID (time-ordered) or blake3(rand) hex; keep as short as practical (e.g., 26 char ULID).
+   - Put s/{sid}/meta (state=active), s/{sid}/meta/evt_seq=0 (do NOT create goal yet; goal is created when first set).
+   - Put dir/{dir_hash}/active -> {sid}.
+   - Put idx/active/{sid} -> {dir_hash}.
+   - Append event s/{sid}/evt/{0000000000000001} type=session_started.
+5) Commit.
+
+#### Set goal (np g s …)
+1) Lookup sid via dir/{dir_hash}/active.
+2) Check if s/{sid}/goal exists:
+   - If it exists: error with message "Goal already set. Use 'np g u' to update it (requires -r/--reason flag)."
+   - If it does not exist: proceed to step 3.
+3) txn:
+   - Create s/{sid}/goal JSON.
+   - Update s/{sid}/meta.last_updated_at.
+   - Increment s/{sid}/meta/evt_seq and write s/{sid}/evt/{seq} with type=goal_set (no reason required).
+
+#### Update goal (np g u …)
+1) Lookup sid via dir/{dir_hash}/active.
+2) Check if s/{sid}/goal exists:
+   - If it does not exist: error with message "No goal set yet. Use 'np g s' to set it first."
+   - If it exists: proceed to step 3.
+3) Require reason flag (-r or --reason) with brief explanation for the update.
+4) txn:
+   - Update s/{sid}/goal JSON.
+   - Update s/{sid}/meta.last_updated_at.
+   - Increment s/{sid}/meta/evt_seq and write s/{sid}/evt/{seq} with type=goal_updated, reason in payload.
+
+#### Add tasks (np t a …)
+1) For each task:
+   - Compute deterministic id = first 6 hex chars of blake3(normalize(title)+"|"+normalize(description)+"|"+sid)
+     - normalize(x) = strings.TrimSpace(x), then case-fold to lowercase, then apply Unicode NFC normalization
+     - Deterministic across runs and processes; stable within session; won't change on edits because id derives from initial content + sid.
+     - Treat adds as idempotent: if the same task (same title+description) is re-added, it will resolve to the same id and be a no-op.
+     - If a user wants to retry a cancelled task with the exact same title and description, they should update the existing task's status rather than adding a new one, or modify the title/description slightly to differentiate it.
+   - txn:
+     - If s/{sid}/task/{id} absent, increment evt_seq and create task with status=pending, created_at=now, created_seq=evt_seq.
+     - Put s/{sid}/idx/status/pending/{id}.
+     - Update meta.last_updated_at.
+     - Append event task_added with task payload.
+   - Note: When adding multiple tasks in a single np t a invocation, use a single transaction and increment evt_seq for each task to preserve the order they were provided on the command line. Tasks should be displayed sorted by created_seq (not created_at) to maintain stable, predictable ordering.
+
+#### Update task status/title/description (np t u …)
+1) Determine if reason is required:
+   - Required (-r or --reason flag) if updating title or description
+   - Required (-r or --reason flag) if changing status to "cancelled" or "failed"
+   - Not required for status changes to "pending", "in_progress", or "completed"
+2) txn:
+   - Read s/{sid}/task/{id}.
+   - If status changes: delete old s/{sid}/idx/status/{old}/{id}, put new s/{sid}/idx/status/{new}/{id}.
+   - Update task JSON.
+   - Update meta.last_updated_at.
+   - Increment evt_seq, append event task_updated or task_status_changed with reason (if provided).
+
+#### Archive session (np a)
+1) Read s/{sid}/meta:
+   - If already archived (state="archived"), return 0 (idempotent operation).
+2) If active, txn:
+   - Delete dir/{dir_hash}/active.
+   - Delete idx/active/{sid}.
+   - Update s/{sid}/meta.state="archived", archived_at=now, last_updated_at.
+   - Put idx/archived/{ts_be}/{sid} -> {dir_hash}.
+   - Put dir/{dir_hash}/archived/{ts_be} -> {sid}.
+   - Increment evt_seq, append session_archived event.
+
+### Query patterns
+
+- **Active session for current dir:**
+  - Get dir/{dir_hash}/active -> sid; if absent, parent-walk by recomputing dir_hash for parents until found or root
+- **List sessions (for TUI picker, not real-time):**
+  - One-time fetch: iterate idx/active/ for active sessions and idx/archived/ (descending) for archived sessions, limited to {terminal_height} total entries
+  - For each sid, read s/{sid}/meta for dir_path/timestamps and s/{sid}/goal for title
+  - Support paging through additional results if needed
+- **List tasks by status quickly:**
+  - Iterate s/{sid}/idx/status/{status}/, collect ids
+  - For each id, Get s/{sid}/task/{id}
+- **List all tasks:**
+  - Iterate s/{sid}/task/ prefix; sort by created_seq ascending to preserve insertion/provided order
+- **Events chronologically for a session:**
+  - Iterate s/{sid}/evt/ prefix ascending; {seq_be} sorts by event time order
+- **Resume after interruption (np r):**
+  - Use dir/{dir_hash}/active -> sid
+  - Read s/{sid}/goal and s/{sid}/task/* to render plan
+  - Optionally stream recent events from s/{sid}/evt/
+
+### Real-time monitoring via Badger subscribe
+
+#### Session view TUI (viewing a single session) subscribes to these prefixes:
+- s/{sid}/goal
+- s/{sid}/task/
+- s/{sid}/evt/
+- s/{sid}/meta (for last_updated_at/state)
+
+On receiving a change event, re-fetch the affected document(s) and redraw. Keep all write ops in a single txn so subscribers see atomic changes.
+
+#### Session list (NOT real-time)
+- Listing sessions is a one-time fetch operation, not real-time.
+- Fetch the most recent {terminal_height} sessions (active and archived combined) and allow paging.
+- Query pattern: iterate idx/active/ for active sessions, iterate idx/archived/ (descending by timestamp) for archived sessions.
+- For each session, read s/{sid}/meta and s/{sid}/goal to display directory, goal title, and timestamps.
+- No subscriptions needed for the list view—refresh only when user explicitly opens/refreshes the picker.
+
+### Value shapes (JSON)
+
+#### s/{sid}/meta:
+```json
+{
+  "sid": "01JP…",
+  "dir_path": "/abs/path",
+  "dir_hash": "…",
+  "state": "active",
+  "created_at": "RFC3339",
+  "archived_at": null,
+  "last_updated_at": "RFC3339"
+}
+```
+
+#### s/{sid}/goal:
+```json
+{
+  "title": "…",
+  "description": "…",
+  "updated_at": "RFC3339"
+}
+```
+
+#### s/{sid}/task/{id}:
+```json
+{
+  "id": "a1b2c3",
+  "title": "…",
+  "description": "…",
+  "status": "pending",
+  "created_at": "RFC3339",
+  "updated_at": "RFC3339",
+  "created_seq": 1
+}
+```
+
+#### s/{sid}/evt/{seq_be}:
+```json
+{
+  "seq": 1,
+  "at": "RFC3339Nano",
+  "type": "task_added",
+  "reason": null,
+  "cmd": "np t a …",
+  "payload": {}
+}
+```
+
+Note on `reason` field:
+- The `reason` field is optional (null or string) in all events
+- Reason is REQUIRED (via -r/--reason flag) only when:
+  - Updating existing goal or task content (title/description) - explain the change (e.g., "clarified scope", "fixed typo")
+  - Changing task status to "cancelled" or "failed" - explain why (e.g., "no longer needed", "blocked by missing dependency")
+- Reason is NOT required (and should be omitted/null) for expected operations:
+  - Setting goal initially
+  - Adding tasks
+  - Changing task status to "pending", "in_progress", or "completed"
+
+### Concurrency and safety
+
+- Always perform multi-key updates in a single Update txn:
+  - Example: task status change must atomically update task doc, move index keys, and append an event.
+  - When adding multiple tasks in a single CLI invocation, use one transaction that appends all events in order by incrementing evt_seq within the transaction.
+- Event sequencing:
+  - Keep s/{sid}/meta/evt_seq as a u64 counter (stored as 8-byte big-endian bytes, not decimal string, to avoid parse overhead). In txn:
+    - Read current value, increment for each event in the transaction, write back the final value, and write each event at s/{sid}/evt/{seq_be}.
+    - If txn conflict occurs (e.g., concurrent processes updating the same session), retry with short exponential backoff.
+    - This guarantees no gaps or overwrites in the event sequence; sequence numbers strictly increase.
+- Concurrent writers across multiple processes:
+  - Use Badger's optimistic concurrency control with retry logic on transaction conflicts.
+  - Keep transactions small and focused to minimize conflict probability.
+  - For batch operations (multiple tasks added by one command), use a single transaction to ensure atomicity and preserve ordering.
+- Enforce one active session per dir:
+  - np s checks if dir/{dir_hash}/active exists before creating a new session (idempotent behavior documented above).
+- Deterministic task IDs:
+  - Use blake3(normalized(title)+"|"+normalized(description)+"|"+sid), take first 8 hex.
+  - Never recompute on updates; the id is "creation id," not a content hash of current state.
+  - Re-adding the same task (same title+description) resolves to the same id and is treated as a no-op.
+
+### Denormalization (minimal, targeted)
+
+- Status index per session (s/{sid}/idx/status/{status}/{task_id}) to filter efficiently by status.
+- Active sessions index (idx/active/{sid}) for quick listing and subscribe.
+- Archives indices for chronological browsing and per-dir history.
+- Everything else is normalized to keep writes simple and atomic.
+- Task ordering: tasks are displayed sorted by created_seq (not created_at timestamp) to preserve the exact order they were provided via command-line flags, avoiding issues with clock precision and race conditions.
+
+### Key naming conventions
+
+- Lowercase ASCII, "/" as namespace delimiter, fixed segment order.
+- Use big-endian (zero-padded hex) counters/timestamps in keys to preserve sort order.
+- Keep values small; JSON for human-inspectable records; empty values for set-like indexes.
+
+### How subscriptions map to UI updates
+
+- **Session pane (single session view):** Subscribe to s/{sid}/task/ and s/{sid}/goal; when a mutation arrives, read the document(s) and refresh.
+- **Event log pane (single session view):** Subscribe to s/{sid}/evt/; stream latest events as they're appended by iterating from the last seen {seq_be}.
+- **Session list picker:** No subscriptions. One-time fetch on open/refresh: scan idx/active/ and idx/archived/ (limited by terminal height), read s/{sid}/meta and s/{sid}/goal for each session to display directory, title, and timestamps. Support paging through results.
+
+### Effort/scope
+
+- Schema + helpers (hashing, key builders, JSON structs): S (≤1–3h).
+- Implement start, goal set/update, task add/update, archive with atomic txns + events: M (1–2d to get right and tested).
+- TUI subscriptions wired to these prefixes: M (1–2d) once data model is in place.
+
+## Rationale and trade-offs
+
+- Prefix-first design matches Badger's strengths: fast iterators, prefix scans, and subscriptions.
+- Using per-session monotonic seq keys for events is simpler and more reliable than time-sorted keys, and you can still store timestamps in the value.
+- Minimal denormalization: we only add what's necessary for speed (status index, active session index). Everything else is derivable.
+- Hashing dir paths keeps keys safe/short and avoids platform-specific path issues while keeping full path in values for display.
+- JSON values keep early development simple and debuggable; you can binary-encode later if needed, without changing keys.
+- GOOS/GOARCH were removed from dir_hash computation to support cross-platform shared home directories (e.g., NFS, Dropbox, synced directories). The same working directory should map to the same session regardless of which OS accesses it.
+
+## Risks and guardrails
+
+- **Concurrent writers:** Use retries on txn conflicts (especially when incrementing evt_seq). Keep transactions small and focused.
+- **Clock weirdness:** Event order is based on seq, not time, so your chronology is stable even if system clock changes; human-readable timestamps live in the value.
+- **Task ID determinism vs edits:** Because id is based on initial content + sid, it won't change during edits. If you attempt to re-add the same task, you'll get the same id; treat it as idempotent add.
+- **Orphaned indexes:** Always mutate doc + index keys in the same txn; on startup or rarely, you can offer a "repair" command to reconcile status indexes, but careful txns should make this unnecessary.
+- **Path hashing collisions:** With blake3 and 8+ bytes hex, collision risk is negligible; still store full dir_path in meta for verification.
+- **Subscription storms:** Subscribe narrowly (per-session) and debounce UI redraws.
+
+## When to consider the advanced path
+
+- **Thousands of tasks/events per session and UI becomes slow scanning JSON:**
+  - Switch task/event values to a compact binary encoding (e.g., flatbuffers/msgpack) to reduce I/O.
+- **Cross-session global event feed (optional):**
+  - Maintain a global index per event: idx/events/{ts_be}/{sid}/{seq_be} -> "" pointing to s/{sid}/evt/{seq_be}.
+  - Pros: simple "recent activity" TUI across sessions, single subscription to idx/events/, efficient time-range scans.
+  - Cons: extra write per event (write amplification), hot global prefix under high throughput, time-ordering relies on event timestamps (minor clock skew acceptable), large histories require pagination/retention.
+- **Need ordering of tasks:**
+  - Add s/{sid}/idx/task_order/{seq_be} -> {task_id} and maintain it on inserts/moves.
+- **Multi-tenant or remote DB:**
+  - Add tenant prefixes or shard DBs; preserve the same key layout.
+
+## Optional advanced path (brief)
+
+- Use ULID for sids to naturally order sessions by creation time (helps chronological archive listings without extra metadata).
+- Maintain s/{sid}/stats -> {pending, in_progress, …, updated_at} and update counts transactionally on task status changes; lets the session list show counts without reading all tasks.
+- Add CAS-like guards using a version field inside task JSON (optimistic concurrency within your app domain), rejecting stale updates gracefully.
+
+## Summary
+
+This schema gives you:
+- O(1) active-session lookup by working directory.
+- O(#tasks with status) retrieval for status-filtered views via a single prefix iteration.
+- Stable, strictly-ordered per-session event logs.
+- Simple, efficient prefixes to subscribe for real-time TUI updates.
+- Clean archival with preserved queryability and chronological browsing.