feat: /pause and /resume skills for cross-session long-task continuation

.claude/plugins/onebrain/.claude-plugin/plugin.json

@@ -1,6 +1,6 @@
 {
   "name": "onebrain",
-  "version": "2.4.9",
+  "version": "2.4.10",
   "description": "OneBrain — Where human and AI thinking become one. A powerful thinking partner powered by AI synergy.",
   "author": {
     "name": "OneBrain Contributors"

.claude/plugins/onebrain/INSTRUCTIONS.md

@@ -58,6 +58,7 @@ If `[agent_folder]/MEMORY.md` has no `## Identity & Personality` section, onboar
 07-logs/           Logs split by type:
                      session/YYYY/MM/  Session logs (YYYY-MM-DD-session-NN.md)
                      checkpoint/       Stop-hook scratch (flat; consumed by /wrapup)
+                     pause/            Pause-thread snapshots (flat; _active.md pointer + YYYY-MM-DD-{slug}-pause-NN.md)
                      update/           Release history (flat, version-named)
                      log/YYYY/MM/      Skill log entries (recap, doctor, memory, ...)
 attachments/       Copied files from /import --attach (pdf/, images/, video/)
@@ -73,6 +74,8 @@ TASKS.md           Live task dashboard (created by /tasks, read-only query block
 - Archive items: `[archive_folder]/YYYY/MM/filename.md` (organized by date archived)
 - Session logs: `[logs_folder]/session/YYYY/MM/YYYY-MM-DD-session-NN.md`
 - Checkpoints: `[logs_folder]/checkpoint/YYYY-MM-DD-{session_token}-checkpoint-NN.md` (flat, auto-generated by hooks, not manual)
+- Pause snapshots: `[logs_folder]/pause/YYYY-MM-DD-{slug}-pause-NN.md` (flat, NN scoped to slug across all dates)
+- Pause pointer: `[logs_folder]/pause/_active.md` (single line containing the active slug)
 - Update logs: `[logs_folder]/update/YYYY-MM-DD-update-vX.Y.Z.md` (flat, written by /update)
 - Skill logs: `[logs_folder]/log/YYYY/MM/YYYY-MM-DD-{skill}.md` (append per day; some have discriminator e.g. `distill-{slug}`, `qmd-{subcommand}`)
 - Inbox items: `[inbox_folder]/YYYY-MM-DD-topic.md` (flat, no subfolders)
@@ -149,6 +152,8 @@ These workflows are documented in `.claude/plugins/onebrain/skills/`:
 | `/tasks` | `tasks/SKILL.md` | Create or update live task dashboard (TASKS.md) and open in Obsidian | user asks to view the task dashboard, regenerate TASKS.md, or open it in Obsidian |
 | `/moc` | `moc/SKILL.md` | Create or update vault portal (MOC.md) and open in Obsidian | user asks to update the vault map |
 | `/wrapup` | `wrapup/SKILL.md` | Wrap up session → session log | explicit `/wrapup` command only — end-of-session signals are handled silently by Auto Session Summary |
+| `/pause` | `pause/SKILL.md` | Save snapshot of long-running work to resume later | user signals pause of long task ("pause", "หยุดก่อน", "step away", "ทิ้งงานไว้") |
+| `/resume` | `resume/SKILL.md` | Load active pause thread state | user signals resume of paused work ("resume", "กลับมาทำต่อ", "pick up", "ที่ค้างไว้") |
 | `/learn` | `learn/SKILL.md` | Teach the agent : facts or behavioral preferences | user tells the agent to remember or learn something |
 | `/memory-review` | `memory-review/SKILL.md` | Interactive memory pruning | (manual only) |
 | `/clone` | `clone/SKILL.md` | Package agent context for vault transfer | (manual only) |
@@ -178,6 +183,8 @@ These agents live in `.claude/plugins/onebrain/agents/` and are dispatched autom
 
 When a user message clearly maps to a skill, invoke it directly — no `/command` needed. If intent is ambiguous, use AskUserQuestion to confirm before invoking. When trigger conditions overlap, prefer the lighter-weight skill (e.g. `/capture` over `/braindump`, `/bookmark` over `/summarize`). Skills marked "manual only" require explicit `/command` always.
 
+**/pause vs /capture vs /wrapup:** Use `/pause` only when work in progress will continue across multiple sessions. `/capture` is for a single idea unrelated to ongoing work. `/wrapup` is terminal — when the work is done. If unsure between `/pause` and `/capture`, prefer `/capture` (lighter weight). Routing keywords for `/pause`: "pause", "step away", "ทิ้งงานไว้", "ค่อยมาทำต่อ", "long task". Routing keywords for `/resume`: "resume", "continue", "pick up", "ทำต่อ", "ที่ค้างไว้".
+
 ## Search Strategy
 
 If qmd MCP tools are available (`mcp__plugin_onebrain_qmd__query` in tool list):
@@ -240,18 +247,20 @@ On weekends: lighter, less task-focused tone. **No-repeat rule:** don't ask abou
 - Glob `[inbox_folder]/*.md` → count files as `inbox_count`
 - Run the Grep tool **twice** — once with `path: "[projects_folder]"` and once with `path: "[inbox_folder]"`, both with the pattern `- \[ \] .*📅 [0-9]{4}-[0-9]{2}-[0-9]{2}` and `output_mode: "content"`. Combine the two result sets. (Grep handles UTF-8 correctly on every platform; the previous Bash shell-out relied on `LC_ALL=en_US.UTF-8` and GNU grep BRE semantics, both Bash-only.) Keep only tasks where date ≤ today; group overdue first, then due today
 - Run `onebrain orphan-scan "[logs_folder]" "[session_token]"` (from vault root) → parse JSON output; read `orphan_count` field. JSON shape: `{"orphan_count":N}`. If the command fails or is unavailable, fall back to a structure-aware glob: if `[logs_folder]/checkpoint/` exists, glob `[logs_folder]/checkpoint/*-checkpoint-*.md` (post-v2.4.0 flat layout); else glob `[logs_folder]/**/*-checkpoint-*.md` (pre-v2.4.0 nested layout — multi-vault user on an unmigrated vault). Then discard files whose date has a non-auto-saved session log (look in `[logs_folder]/session/YYYY/MM/` for post-v2.4.0, or `[logs_folder]/YYYY/MM/` for pre-v2.4.0), and count distinct session tokens among remaining files.
+- Read `[logs_folder]/pause/_active.md` if present → parse single-line content as `active_pause_slug`. If absent, set `active_pause_slug = null`. Then if non-null: glob `[logs_folder]/pause/*-{active_pause_slug}-pause-*.md` and count them as `active_pause_count`; read the latest file's `date` frontmatter as `active_pause_last_date`.
 - **Legacy structure detection (post-v2.4.0):** Check whether `[logs_folder]/session/` exists (any of the new top-level subfolders works as a sentinel; `session/` is the most representative). If it does NOT exist AND `[logs_folder]/YYYY/` does exist (legacy structure pre-v2.4.0), set `vault_structure_legacy = true`. If both `session/` and a legacy `YYYY/` exist (partial migration), still treat `vault_structure_legacy = false` — `/update` will resume cleanup on next run. If neither exists (fresh vault), `vault_structure_legacy = false`.
 
 **Step 4 — Send startup status (after Step 3 completes):**
 
-If inbox_count = 0 and orphan_count = 0 and qmd_unembedded = 0 and vault_structure_legacy = false and no tasks found: show nothing after the greeting.
+If inbox_count = 0 and orphan_count = 0 and qmd_unembedded = 0 and active_pause_slug is null and vault_structure_legacy = false and no tasks found: show nothing after the greeting.
 
 Otherwise, append after the greeting:
 
 ```
 📥 inbox [N]                               ← omit if inbox_count = 0
 📋 [N] orphan session(s) — /wrapup?        ← omit if orphan_count = 0
 ⚠️ qmd: [N] doc(s) need embedding — /qmd embed   ← omit if qmd_unembedded = 0
+📂 active pause: {active_pause_slug} ({active_pause_count} snapshots, last {active_pause_last_date}) — /resume?   ← omit if active_pause_slug is null
 ⚠️ vault structure outdated — run /update to migrate   ← omit if vault_structure_legacy = false
 
 Pending tasks:
@@ -335,12 +344,12 @@ Different commands have different verbosity expectations. Match output to the pr
 
 | Profile | Commands | Behavior |
 |---------|----------|----------|
-| **Capture** | `/capture`, `/braindump`, `/bookmark`, `/learn` | Write the note, confirm done in 1 line. No elaboration. |
+| **Capture** | `/capture`, `/braindump`, `/bookmark`, `/learn`, `/pause` | Write the note/snapshot, confirm done. /pause adds the active-thread line. No elaboration. |
 | **Automated** | cron jobs, Auto Session Summary, `/wrapup` | Structured output only (bullets/sections). No commentary. Under 300 words. |
-| **Interactive** | `/research`, `/connect`, `/consolidate`, `/reading-notes`, `/weekly`, `/distill`, `/recap` | Normal verbosity : depth matches task complexity. |
+| **Interactive** | `/research`, `/connect`, `/consolidate`, `/reading-notes`, `/weekly`, `/distill`, `/recap`, `/resume` | Normal verbosity : depth matches task complexity. |
 | **Diagnostic** | `/doctor` | Structured report output. No meta-commentary. Lead with findings. |
 | **Config/Setup** | `/onboarding`, `/tasks`, `/moc`, `/qmd` | Confirm actions taken. No verbose explanation unless asked. |
-> **Terminal rendering:** The Obsidian terminal plugin renders markdown natively (tables, bold, headers, code blocks). Interactive and Diagnostic profiles should use full markdown. Capture profile: 1-line plain-text confirm only. Automated profile: no headers — output is read async via Telegram where markdown rendering varies.
+> **Terminal rendering:** The Obsidian terminal plugin renders markdown natively (tables, bold, headers, code blocks). Interactive and Diagnostic profiles should use full markdown. Capture profile: 1-line plain-text confirm only (exception: `/pause` emits a 2-line confirm + 2-line user instruction block per its SKILL.md Step 6). Automated profile: no headers — output is read async via Telegram where markdown rendering varies.
 
 
 For cron/automated agents specifically: output is read by the user async (often via Telegram) : lead with the content, skip all meta-commentary about what you're doing.

.claude/plugins/onebrain/skills/doctor/SKILL.md

@@ -130,6 +130,52 @@ Only run when vault.yml contains a `schedule:` block. Skip entirely otherwise.
 - **Schedule drift** — Read `vault.yml` `schedule:` block. For each entry, check that the corresponding launchd plist exists at `~/Library/LaunchAgents/com.onebrain.<labelSafe>.plist` where `labelSafe` strips leading `/` from `entry.skill` and replaces non-`[a-zA-Z0-9-]` chars with `-`. If any entry's plist is missing → 🟡 drift — suggest `onebrain register-schedule`. If any installed plist no longer matches a vault.yml entry (stale orphan) → 🟡 stale plist — suggest `onebrain register-schedule --remove` then re-register.
 - **One-shot reachability** — For each entry with `at:` (one-shot), verify the timestamp has not already passed. If passed and the plist still exists → 🟡 expired one-shot not cleaned up — suggest `onebrain register-schedule --remove` to clear the stale plist (the self-delete shell may have failed to run).
 
+### Pause: Orphan Pointer
+
+Check: `[logs_folder]/pause/_active.md` exists but no pause file matches its slug.
+
+How to detect:
+1. Read `_active.md`. If absent → skip.
+2. Parse slug. Glob `[logs_folder]/pause/*-{slug}-pause-*.md`. If empty match → orphan pointer.
+
+Report (if orphan):
+```
+⚠️ Pause pointer references `{slug}` but no snapshot files exist.
+   Fix: rm 07-logs/pause/_active.md (or create initial /pause)
+```
+
+### Pause: Missing Pointer
+
+Check: pause files exist but `_active.md` is missing.
+
+How to detect:
+1. Glob `[logs_folder]/pause/*-pause-*.md`. If empty → skip.
+2. If `_active.md` exists → skip.
+3. Otherwise: missing pointer. Identify slug(s) from filenames.
+
+Report (if missing):
+```
+⚠️ Pause files exist but no active pointer:
+   - {slug-1} (N snapshots, latest YYYY-MM-DD)
+   - {slug-2} (M snapshots, latest YYYY-MM-DD)
+   Fix: echo "{chosen-slug}" > 07-logs/pause/_active.md
+```
+
+### Pause: Idle Thread
+
+Check: active pause thread has had no new activity (no new pause file, no `/resume`) for > 14 days.
+
+How to detect:
+1. Read `_active.md`. If absent → skip.
+2. Glob pause files for the slug. Get max date prefix from filenames.
+3. If `(today - max_date).days > 14` → idle.
+
+Report (if idle):
+```
+⚠️ Pause thread `{slug}` idle for {N} days (last snapshot YYYY-MM-DD).
+   Fix: /wrapup to consolidate, or /pause to refresh, or /resume to continue
+```
+
 ---
 
 ## Step 3: Report Findings

.claude/plugins/onebrain/skills/pause/SKILL.md

@@ -0,0 +1,118 @@
+---
+name: pause
+description: "Save a snapshot of long-running work mid-flight to resume later in a future session. Use when user signals they need to pause a long task and continue another time — 'pause this work', 'หยุดก่อน', 'พักงานนี้ก่อน', 'พักไว้ก่อน', 'step away', 'ทิ้งงานไว้ก่อน', 'ค่อยมาทำต่อ'. Writes a snapshot to 07-logs/pause/ but does NOT end the session or clear context. Do NOT use for: ending a session (use /wrapup), capturing a single idea (use /capture), saving a memory (use /learn)."
+schedulable: false
+---
+
+# /pause — Save Long-Task Snapshot (TL;DR)
+
+Writes a per-thread snapshot of the current work to `[logs_folder]/pause/` so the next session can `/resume` and pick up seamlessly. Multiple pause files accumulate across days for the same task; `/wrapup` later consolidates them all into one session log.
+
+**Mental model:** `/pause` = SAVE snapshot. It does NOT end the session and does NOT clear context. User runs `/clear` manually when ready to free context.
+
+---
+
+## Scope
+
+- /pause writes pause snapshot files only. It does NOT touch checkpoints, session logs, MEMORY.md, or memory/ files.
+- It DOES update the pointer file `[logs_folder]/pause/_active.md` to set/confirm the active thread slug.
+
+---
+
+## Step 1: Determine Active Thread Slug
+
+1. Read `[logs_folder]/pause/_active.md` if present. The file is plain text — one line — containing a kebab-case slug.
+2. If the file exists and is non-empty: use that slug as `active_slug`.
+3. If absent / empty:
+   - Agent reviews session context → propose a kebab-case slug from the dominant topic (e.g. `cli-refactor`, `oracle-borrows`)
+   - Use `AskUserQuestion` to confirm. Question: "Active thread name? Suggested: `<proposed-slug>`" with options: (a) accept suggestion, (b) type a different slug
+   - On confirmation, set `active_slug = <chosen>` and continue
+4. If `/pause --task=<slug>` was invoked with an explicit slug:
+   - If `_active.md` is empty or contains the same slug → set `active_slug = <slug>`, skip warning
+   - If `_active.md` contains a different slug → use `AskUserQuestion`: "⚠️ Active thread: `<existing>` (N unmerged snapshots). Switch to `<new>`? (y/N)". On `y` → set `active_slug = <new>`. On `n` → abort silently with message "Pause aborted — active thread unchanged."
+
+---
+
+## Step 2: Compute Next NN
+
+1. Glob `[logs_folder]/pause/*-{active_slug}-pause-*.md` (no date filter — NN is scoped to slug across all dates).
+2. Parse the `NN` segment (the two-digit number before `.md`) from each filename.
+3. `next_nn` = `max(NN parsed) + 1`. If no files → `next_nn = 01`. Always zero-pad to 2 digits.
+
+---
+
+## Step 3: Generate Snapshot Content
+
+Review the current conversation since the last pause file of `active_slug` (or since session start if none). Fill the body sections following `skills/startup/references/session-formats.md` → Pause File Format.
+
+**Required sections (in order):**
+
+- `## Where I Stopped` — 1–2 sentences: file being edited, decision pending, last test run, current focus
+- `## Resume With` — one concrete next action: specific file path, command, decision point. Must be actionable as the FIRST thing on resume
+- `## What We Worked On` — 2–3 sentences describing the pause-interval's focus
+- `## Key Decisions` — bullet list (every unique decision since last pause, full detail)
+- `## Insights & Learnings` — omit if none
+- `## Action Items` — `- [ ] task 📅 YYYY-MM-DD` format
+- `## Open Questions` — omit if none
+
+**Word cap:** 350 words total. Preservation rule: every unique decision/action/question since last pause MUST appear — deduplication only, no summarization.
+
+---
+
+## Step 4: Write the Pause File
+
+1. Today's date as `YYYY-MM-DD`.
+2. Ensure directory exists: `mkdir -p [logs_folder]/pause/`
+3. Get `session_token` from agent context (run `onebrain session-init` to recover if missing). **If `session-init` fails or returns no token:** abort the write. Do NOT proceed to Step 5. Output: `⚠️ Could not determine session token. Snapshot not saved — try again or run /doctor.`
+4. Write to `[logs_folder]/pause/YYYY-MM-DD-{active_slug}-pause-{next_nn}.md`:
+
+```yaml
+---
+tags: [pause, session-log]
+date: YYYY-MM-DD
+session_token: <token>
+task_slug: <active_slug>
+pause: <NN>
+trigger: manual
+---
+```
+
+Followed by the body sections from Step 3.
+
+**If the file write fails** (disk full, permission denied, etc.): abort the skill. Do NOT proceed to Step 5 (do not update `_active.md` — that would create an instant orphan pointer). Output: `⚠️ Snapshot failed to save — check disk space or file permissions. Active thread unchanged.`
+
+---
+
+## Step 5: Update Active Pointer
+
+Write `[logs_folder]/pause/_active.md` with a single line: the `active_slug` value (no trailing newline issues — overwrite is idempotent if slug already matches).
+
+---
+
+## Step 6: Confirm
+
+Output exactly this format (Capture profile — 1 confirmation block, no elaboration):
+
+```
+💾 Snapshot saved: 07-logs/pause/YYYY-MM-DD-{slug}-pause-NN.md
+📂 Active thread: {slug} (N snapshots)
+
+คุยต่อได้เลย — ยังไม่ตัด context.
+พิมพ์ /clear เมื่อพร้อมพักงาน (snapshot จะรอ session ถัดไป)
+```
+
+Where `N` = `next_nn` from Step 2 (the total snapshot count for this thread including the one just written).
+
+---
+
+## Auto-Finalize (Called by AUTO-SUMMARY and /wrapup "n" branch)
+
+When AUTO-SUMMARY or `/wrapup` "n" branch needs to finalize an active thread before writing a daily session log, it calls into this skill's auto-finalize path. Three skip conditions — if ANY is true, skip:
+
+1. **No-activity:** Glob `[logs_folder]/checkpoint/*-{current_session_token}-checkpoint-*.md` — if empty, the session produced no work; skip.
+2. **Already-captured-this-session:** Find the latest pause file of `active_slug`. If its frontmatter `session_token` equals current session token AND no checkpoint file mtime > pause file mtime, the current session's work is already captured by an earlier `/pause`; skip.
+3. **No-pause-files-and-untouched:** If no pause file exists yet for `active_slug` AND the newest checkpoint mtime is older than `_active.md` mtime, the slug was set but no thread activity occurred; skip.
+
+If not skipped: run Steps 2–5 above, but in Step 4 frontmatter use `trigger: auto-finalize` instead of `manual`, and in Step 3's `## Where I Stopped` prepend "Auto-finalized at session end. " to the human-readable text.
+
+No Step 6 confirm — auto-finalize is silent.