task-inbox-fill

# 📥 Todoist Inbox Fill — ⚠️ DEPRECATED ⚠️ > [!warning] **DEPRECATION NOTICE: Transitioned to Obsidian TaskNotes (June 23, 2026)** > Justin has stopped using Todoist entirely and transitioned his task management to **Obsidian TaskNotes**. > - This skill is legacy/deprecated and should not be run for Todoist. > - Any future automated task ingestion or "inbox fill" sweeps should target the Obsidian TaskNotes system (creating new task notes in `TaskNotes/Tasks` or checklists in daily notes for the EIIRP sweep), following the conventions in `/home/justin.guest/Developer/obsidian-vault/TaskNotes/Setup.md`. Scan external sources for open actions Justin owns and needs to do. Deduplicate against what already exists in Todoist. Surface candidates to Justin for confirm/edit. Batch-add confirmed tasks to Inbox. **This is not a sync.** It's a one-way capture pass. It does not complete, reschedule, or remove tasks. It only adds what's genuinely missing. --- ## Step 1 — Pre-flight Determine the lookback window. If Justin specified one (e.g. "last 24 hours", "go back 3 days"), use it. Otherwise default to **48 hours**. Compute dates and snapshot the current Todoist state. Do this in-context (not in a subagent). ```bash TODAY=$(date +%F) # e.g. 2026-05-21 TOMORROW=$(date -d "$TODAY + 1 day" +%F) TODAY_SLASH=$(date +%Y/%m/%d) TOMORROW_SLASH=$(date -d "$TODAY + 1 day" +%Y/%m/%d) WEEK_FROM_NOW=$(date -d "$TODAY + 7 days" +%F) # Lookback window (default 48h = 2 days) LOOKBACK_HOURS=48 # override if Justin specified something else LOOKBACK_START=$(date -d "$TODAY - ${LOOKBACK_HOURS} hours" +%F) LOOKBACK_START_SLASH=$(date -d "$TODAY - ${LOOKBACK_HOURS} hours" +%Y/%m/%d) ``` Pass `LOOKBACK_HOURS`, `LOOKBACK_START`, `LOOKBACK_START_SLASH`, and `LOOKBACK_ISO` (ISO 8601 datetime for Linear, e.g. `2026-05-21T00:00:00.000Z`) to subagents so they scope their searches to the right window. Compute `LOOKBACK_ISO`: ```bash LOOKBACK_ISO=$(date -u -d "${LOOKBACK_HOURS} hours ago" +%Y-%m-%dT%H:%M:%S.000Z) ``` Then snapshot **all open tasks** in Todoist so you can deduplicate later. Call both: 1. `find-tasks-by-date(startDate="today", daysCount=7, overdueOption="include-overdue", limit=100)` 2. `find-tasks(filter="no date", limit=100)` Merge results and keep them in-context. You'll diff against this snapshot in Step 3. This two-call method is robust and completely avoids HTTP 400 errors caused by complex Todoist search query syntax. --- ## Step 2 — Gather candidates (parallel subagents) Spawn one subagent per source in a **single batch**. Each subagent returns a compact bullet list of *open actions* — things Justin needs to do. Not FYIs. Not things others owe him (unless he needs to follow up). Not decisions already made. Pass `TODAY`, `TOMORROW`, `TODAY_SLASH`, `TOMORROW_SLASH`, `WEEK_FROM_NOW`, `LOOKBACK_HOURS`, `LOOKBACK_START`, and `LOOKBACK_START_SLASH` into each subagent as verbatim substituted strings. **Budget for every subagent: ≤8 tool calls. Return partial results and stop if budget is exhausted.** Sources to cover in two/three batches (concurrency cap is 3): - **Batch 1:** Slack, Gmail, Obsidian daily notes - **Batch 2:** Calendar, Linear, Granola meeting notes Sources: **Slack, Gmail, Obsidian daily notes, Calendar, Linear, Granola.** **Do not pipe command output into a language interpreter** (no `... | python3 -c "..."`, no `... | bash`, no `... | node -e "..."`). The security scanner flags `cmd | python3` etc. as `pipe_to_interpreter` (HIGH) regardless of intent and will halt your run for approval. If you need to post-process JSON, use `jq` (installed). If you need real Python, write a short script to a tempfile and run it as `python3 /tmp/foo.py` — the file boundary is what satisfies the scanner. --- ### Subagent A — Slack (SignLab) - **Toolsets:** `["terminal"]` - **Skill:** `slack` - **Goal:** Find open actions for Justin in Slack today. Budget: 8 tool calls. - **Context:** > Extract open actions from Slack for Justin. **Only surface messages where Justin has saved the message or set a Slack reminder** — these are the clearest signal of intent to act. > > Run one search: > 1. `python3 ${HERMES_HOME:-$HOME/.hermes}/skills/social-media/slack/scripts/slack.py search '(has:reminder OR is:saved) after:<LOOKBACK_START>' --limit 50` > > `is:saved` matches messages Justin has explicitly added to his "Saved items" (formerly starred messages) list. `has:reminder` matches reminders. Together, these represent "saved Slack reminders/messages" Justin means to act on. > > Note: the word "reminder" appearing in message *text* (e.g. "Reminder: meeting at 3pm") can occasionally leak in via `has:reminder`. Drop these if they're clearly not reminder-flagged or saved — use context and channel to judge. > > **Command safety:** Do NOT pipe `slack.py` output into a language interpreter (`| python3 -c`, `| bash`, `| node -e`). The security scanner blocks these as `pipe_to_interpreter` (HIGH) and your run will halt for approval. If you need to inspect or reshape the JSON, use `jq`. Example: > ```bash > python3 ${HERMES_HOME:-$HOME/.hermes}/skills/social-media/slack/scripts/slack.py search '(has:reminder OR is:saved) after:<LOOKBACK_START>' --limit 50 \ > | jq -r '.[] | "\(.channel_name) | \(.username) | \(.permalink) | \(.text[:200])"' > ``` > | jq -r '.[] | "\\(.channel_name) | \\(.username) | \\(.permalink) | \\(.text[:200])"' > ``` > For non-trivial Python, write to a tempfile and run as `python3 /tmp/foo.py` — the file boundary is what satisfies the scanner. > > Format each as a candidate task: > `- [Slack] <concise action> | context: <#channel or @person, brief what/why> | url: <permalink>` > > End with `Total: N candidates`. > > Budget: 8 tool calls. Return what you have and stop if budget exhausted. --- ### Subagent B — Gmail (work + personal-main) - **Toolsets:** `["terminal"]` - **Skill:** `google-workspace` - **Goal:** Find emails where Justin owes a reply or has an explicit action item. Budget: 8 tool calls. - **Context:** > Use the wrapper at `${HERMES_HOME:-$HOME/.hermes}/skills/productivity/google-workspace/scripts/gws_multi.py`. Read-only — do NOT attempt sends. > > Run exactly two searches (no personal-junk unless Justin explicitly asked): > 1. `gws_multi.py --account work gmail search 'after:<LOOKBACK_START_SLASH> before:<TOMORROW_SLASH> in:inbox -label:sent' --max 50` > 2. `gws_multi.py --account personal-main gmail search 'after:<LOOKBACK_START_SLASH> before:<TOMORROW_SLASH> in:inbox -label:sent' --max 50` > > The `in:inbox` filter is critical — it excludes archived threads. **Do not surface tasks from archived emails.** If a thread has been archived, Justin has already processed it and it is not an open action. > > From results (subject, from, snippet only — do NOT fetch full bodies), extract: > - Emails from humans that contain a direct question or request for Justin. > - Emails Justin sent that contain a commitment or follow-up he still needs to do. > - Threads where he's last-replied-to but hasn't responded. > > For each result, include the message ID or thread ID if available in the search output (needed to construct a Gmail link). Gmail deep links follow the pattern `https://mail.google.com/mail/u/0/#Inbox/<threadId>` — include this if you can derive it. > > Skip: automated notifications, CI/build alerts, shipping/delivery, marketing, newsletters, receipts, calendar invites, App Store Connect issue alerts, Todoist onboarding, Readwise, Substack. > > **Command safety:** Do NOT pipe `gws_multi.py` output into a language interpreter (`| python3 -c`, `| bash`, `| node -e`). The scanner blocks these. Use `jq` for JSON, or write a Python helper to `/tmp/foo.py` and run as `python3 /tmp/foo.py`. > > Format each candidate task: > `- [Email/<account>] <concise action> | from: <sender> | subject: <subject>` > > End with `Total: N candidates (work: X, personal-main: Y)`. > > Budget: 8 tool calls. Return what you have and stop. --- ### Subagent G — Granola meeting notes (Next Steps for Justin) - **Toolsets:** `["terminal", "file"]` - **Skill:** `obsidian` - **Goal:** Find Next Steps assigned to Justin in recent meeting notes. Budget: 8 tool calls. - **Context:** > Vault path: resolve from env `OBSIDIAN_VAULT_PATH` (fallback: `~/Documents/Obsidian Vault`). Meeting notes live under `<vault>/Inputs/Meetings/`. Each note filename starts with the meeting date (`YYYY-MM-DD`). > > Scan the `<vault>/Inputs/Meetings/` directory for `.md` files whose date prefix falls within the lookback window (`<LOOKBACK_START>` to `<TODAY>`). > > For each such meeting note: > 1. Read the file. > 2. Find any section with "Next Steps" or "Priorities" in its heading (e.g. `### Next Steps` or `### Priorities and Next Steps`). > 3. Extract only lines that begin with `- Justin:` (case-insensitive) under that section. These are the action items assigned to Justin. > 4. Skip all other lines (other people's action items, bullets under other sections). > > **Command safety:** Use `search_files` and `read_file` from the file toolset — do NOT use `grep` piped to `python3 -c` or similar (scanner blocks `pipe_to_interpreter`). Use `jq` if reshaping JSON. > > Format each candidate task: > `- [Granola/<YYYY-MM-DD>] <action text (strip the "Justin: " prefix)> | meeting: <note title> | file: <relative path>` > > End with `Total: N candidates across M meeting notes`. > > Budget: 8 tool calls. Return what you have and stop. --- ### Subagent C — Obsidian daily notes (lookback window) - **Toolsets:** `["terminal", "file"]` - **Skill:** `obsidian` - **Goal:** Find uncaptured open actions in recent daily notes. Budget: 8 tool calls. - **Context:** > Read `OBSIDIAN_VAULT_PATH` from env (fallback: `~/Documents/Obsidian Vault`). Daily note filename format: `YYYY-MM-DD DayName.md`. Current notes live in the vault root; older ones in `Daily Notes/`. > > **Capture convention:** Raw `- [ ]` checkboxes in daily notes are the intended quick-capture format for tasks. Some of these may already have been converted to TaskNotes (in-vault task files) by the vault hygiene EIIRP sweep — converted tasks appear as wikilinks in the form `[[TaskNotes/Tasks/slug-timestamp]]` rather than as raw checkboxes. > > For each of the last N days matching the lookback window (starting today, <TODAY>, going back to <LOOKBACK_START>), find and read the daily note if it exists. > > From each note, extract **only**: > - Unchecked task items: lines matching `- [ ]` that are not marked done. > - Lines containing explicit first-person commitment phrases: "I need to", "I need to remember", "I have to", "I should", "I must", "don't forget", "follow up on", "remind me". > > Do NOT infer tasks from section headings, bullet points under "Open Questions" or "Blockers", or general observations. Only literal `- [ ]` items and lines with the above explicit keywords qualify. > > Skip: > - `- [x]` lines (completed) > - Headings, pure notes, observations, decisions, highlights > - Any line that is already a TaskNote wikilink (contains `[[TaskNotes/Tasks/`) — these are already captured in-vault and do not need a separate Todoist task unless Justin explicitly indicates otherwise > > Raw `- [ ]` lines that remain unconverted are still valid Todoist candidates — surface them as normal. > > **Command safety:** Use the `file` toolset and `search_files` to read notes — do NOT pipe shell output into `python3 -c` / `bash` / `node -e` (scanner blocks `pipe_to_interpreter`). > > Format each candidate task: > `- [Obsidian/<date>] <action text> | note: <YYYY-MM-DD DayName.md>` > > End with `Total: N candidates across M notes`. > > Budget: 8 tool calls. Return what you have and stop. --- ### Subagent H — Obsidian Vault Notes (New & Modified) - **Toolsets:** `["terminal"]` - **Skill:** `obsidian` - **Goal:** Find open actions, commitments, and suggest follow-up/development tasks for new and modified notes in the vault. Budget: 5 tool calls. - **Context:** > Run the specialized scanning script: > ```bash > python3 ~/.hermes/scripts/fetch_vault_notes_candidates.py --lookback-hours <LOOKBACK_HOURS> --json > ``` > (Replace `<LOOKBACK_HOURS>` with the lookback hours value, e.g. `48` or whatever is set for this run). > > This script scans all non-daily, non-Granola notes in the vault that were created or modified since the previous run. It extracts unchecked tasks (`- [ ]`), commitment statements, and suggests follow-up/development tasks for newly created notes. > > **Task naming rules:** > Format each candidate task exactly as returned by the script: > - Tasks/Commitments: `- [Obsidian] <description> | context: <context>` > - Development Suggestions: `- [Obsidian] <description> | context: <context>` > > End with `Total: N candidates`. > > Budget: 5 tool calls. Return what you have and stop. --- ### Subagent — Linear - **Toolsets:** `["terminal"]` - **Skill:** `linear` - **Goal:** Find Linear issues assigned to Justin (To Do / In Progress) and new Triage issues from within the lookback window. Budget: 8 tool calls. - **Context:** > Use `curl` against `https://api.linear.app/graphql` with `Authorization: $LINEAR_API_KEY`. >