hq

name	hq
description	Connect to the HQ via Supabase. Use when you need to register yourself, search knowledge, create/update pages and skills, claim tasks, post comments, manage contacts and organizations, log interactions, or query the audit trail. Also use at session startup to register and load boot knowledge.

HQ

You are connected to a Supabase-backed HQ. This skill provides helper scripts for all operations.

Important: Run all commands exactly as shown — do not wrap them in cd, bash -c, or shell chaining (&&, ;). Your working directory is already set to your workspace.

Environment Variables

These must be set in your environment:

• SUPABASE_URL — Supabase project URL
• SUPABASE_SERVICE_ROLE_KEY — service role key (bypasses RLS)
• AGENT_SLUG — your unique identifier (matches your git branch)

Session Startup

At the beginning of every session, run these two steps:

1. Register yourself

python3 skills/hq/scripts/hq_register.py

This upserts your agent in the agents table and sets status to online.

2. Load boot knowledge

python3 skills/hq/scripts/hq_boot_docs.py

This fetches workspace-scoped knowledge items (shared context for all agents) and your agent-scoped items (index only). Read the output — it's your shared context.

Field & Pipeline Discovery

Before writing to contacts or organizations, discover what fields and pipeline stages exist for this workspace.

Fetch custom field definitions

python3 skills/hq/scripts/hq_get_fields.py --entity-type contact
python3 skills/hq/scripts/hq_get_fields.py --entity-type organization

Returns all active field definitions with keys, types, labels, and descriptions. The description field contains guidance for how to populate each field — read it before writing data.

Fetch pipeline stages

python3 skills/hq/scripts/hq_get_pipeline.py --entity-type contact

Returns ordered stages with keys and labels. Only set a contact's status to a valid stage key from this list.

CRM

Contact data model

Contacts have universal fields (name, email, phone, company, title, etc.) as real columns, plus workstream-specific fields in an extended JSONB column. The shape of extended is defined by field_definitions — fetch those first to know what fields exist.

Update extended fields on a contact

python3 skills/hq/scripts/hq_update_extended.py contacts CONTACT_ID \
  --data '{"subscriber_count": 145000, "content_style": "vlog"}'

This does an atomic read-modify-write: reads the current extended, merges your new fields, validates keys against field_definitions, and writes back. Unknown field keys are rejected.

To skip validation (for ad-hoc fields): --no-validate

Log an interaction

python3 skills/hq/scripts/hq_log_interaction.py CONTACT_ID \
  --type email --direction outbound \
  --summary "Sent initial outreach" \
  --next-action "Follow up if no reply" --next-action-days 7

Types: email, call, meeting, linkedin_message, dm, intro, coffee, event, note, other

Direction: inbound, outbound, or omit for neutral (meetings, events).

Optional: --channel, --subject, --body, --org-id, --template-id

Create an organization

python3 skills/hq/scripts/hq_create_org.py --name "Acme Corp" --type company \
  --website "https://acme.com" --industry "Aviation"

Optional: --location, --description, --tags tag1,tag2, --extended '{"key": "val"}'

Link a contact to an organization

python3 skills/hq/scripts/hq_link_contact_org.py CONTACT_ID ORG_ID --role "CEO"

Update extended fields on an organization

python3 skills/hq/scripts/hq_update_extended.py organizations ORG_ID \
  --data '{"deal_value": 50000}'

Knowledge

Search knowledge (semantic + full-text fallback)

python3 skills/hq/scripts/hq_search_docs.py "your natural language query"

Optional flags: --tags tag1,tag2 --folder-id UUID --kind page --limit 5

Kind filter: page, skill, file, source. Returns matched knowledge items ranked by similarity.

Create a knowledge item

python3 skills/hq/scripts/hq_create_doc.py --title "Title" --content "Content here" --tags tag1,tag2

Optional: --kind skill (default: page), --scope agent (default: workspace), --folder-id UUID

Automatically requests a local embedding on creation.

Update a knowledge item

python3 skills/hq/scripts/hq_update_doc.py ITEM_ID --title "New title" --content "New content" --tags tag1,tag2

Automatically re-embeds on update.

Get a specific knowledge item

python3 skills/hq/scripts/hq_get_doc.py ITEM_ID

Get knowledge items by tag

python3 skills/hq/scripts/hq_get_docs_by_tag.py TAG_NAME

List connected sources

python3 skills/hq/scripts/hq_list_sources.py

Shows what external systems (Notion, Google Drive, etc.) are connected to the workspace, whether they support writes, and how many items are synced from each.

Write to a connected source

python3 skills/hq/scripts/hq_write_source.py \
  --connection-id UUID \
  --action create_item \
  --params '{"title": "Meeting Notes", "content": "# Summary\n...", "parent_id": "PARENT_UUID"}'

Creates content in the connected external system. Only works if the connection is writable. Check hq_list_sources.py first. The command is routed through the gateway command queue.

Knowledge vs source decision

• Search results include sources. When you search knowledge, source items appear with their provider and source URL. Read them like any knowledge item.
• Don't duplicate. If content already exists as a synced source (e.g., a Notion page), don't create a duplicate knowledge page. Reference the source item instead.
• Write to the right place. If the user's workflow lives in Notion and the connection is writable, prefer writing there over creating a knowledge page. If the connection is read-only, create a knowledge page and mention that the user may want to move it to Notion.

Tasks

List tasks

python3 skills/hq/scripts/hq_list_tasks.py
python3 skills/hq/scripts/hq_list_tasks.py --mine --status todo,in_progress
python3 skills/hq/scripts/hq_list_tasks.py --assignee-type human
python3 skills/hq/scripts/hq_list_tasks.py --stream-id STREAM_UUID --tag product

Useful filters: --status, --stream-id, --assignee-type, --agent-id, --mine, --tag, --limit

List my assigned tasks

python3 skills/hq/scripts/hq_my_tasks.py

Optional: --status todo,in_progress

Claim a task

python3 skills/hq/scripts/hq_claim_task.py TASK_ID

Sets status to in_progress, assigns you, and fetches attached source metadata plus top relevant chunks for document attachments. Use hq_get_knowledge_chunks.py SOURCE_ID or hq_get_doc.py DOCUMENT_ID if you need more context.

Assign a task

python3 skills/hq/scripts/hq_assign_task.py TASK_ID --agent-id AGENT_UUID
python3 skills/hq/scripts/hq_assign_task.py TASK_ID --human

Assigns a task to either an agent or marks it as assigned to a human.

List tasks for humans

python3 skills/hq/scripts/hq_my_human_tasks.py

Optional: --status todo,in_progress

Note: hq_list_tasks.py --assignee-type human is the more general version.

Complete a task

python3 skills/hq/scripts/hq_complete_task.py TASK_ID

Attach a reference to a task

python3 skills/hq/scripts/hq_attach_to_task.py TASK_ID --type knowledge_item --entity-id DOC_UUID
python3 skills/hq/scripts/hq_attach_to_task.py TASK_ID --type url --url "https://example.com" --label "Reference"

Use this for references — existing documents, links, or contacts that provide context for the task. These are NOT deliverables and don't go through review.

Submit a deliverable for review

python3 skills/hq/scripts/hq_submit_deliverable.py --task-id TASK_ID --type page --title "Blog post draft" --content "Your markdown content"
python3 skills/hq/scripts/hq_submit_deliverable.py --task-id TASK_ID --type url --url "https://github.com/org/repo/pull/42" --title "PR #42"

Use this when the task asks you to produce something — a document, a draft, a report, a PR. The human will review it (approve, request revision, or reject). Deliverables appear in a dedicated review section on the task.

When to use deliverables vs knowledge:

• You produced content for a task (any task, any content) → deliverable (hq_submit_deliverable.py). This includes research summaries, drafts, reports, guides — anything the human should be able to review.
• You're saving something independent of any task (learned a skill, noting reference material for later) → knowledge (hq_create_doc.py)
• You need to attach an existing document or link as context for a task → attachment (hq_attach_to_task.py)

Revise a deliverable after feedback

python3 skills/hq/scripts/hq_submit_deliverable.py --task-id TASK_ID --update --deliverable-id LINK_UUID --title "Blog post draft" --content "Revised content"

This updates the knowledge item content and resets the review status to draft so the human can re-review.

Comments

Post a comment on any entity

python3 skills/hq/scripts/hq_comment_on.py task TASK_ID "Your comment here"
python3 skills/hq/scripts/hq_comment_on.py contact CONTACT_ID "Research notes from browser verification"
python3 skills/hq/scripts/hq_comment_on.py organization ORG_ID "Meeting summary"

Optional: --parent-id COMMENT_UUID for replies.

@mentions are parsed automatically from the comment body.

Audited Writes

Insert a record (any table)

python3 skills/hq/scripts/hq_insert.py TABLE --data '{"field": "value"}' --module MODULE --entity-type TYPE

Update a record (any table)

python3 skills/hq/scripts/hq_update.py TABLE RECORD_ID --data '{"field": "value"}' --module MODULE --entity-type TYPE

Both commands automatically create audit_log entries with your agent identity.

Heartbeat

Send a heartbeat

python3 skills/hq/scripts/hq_heartbeat.py

Updates last_seen_at and sets status to online. Run this during heartbeat polls.

Set status

python3 skills/hq/scripts/hq_heartbeat.py --status paused

Background Inbox

You have a durable inbox. When tasks are assigned to you, you're @mentioned in a comment, or a contact automation fires, an inbox item is created automatically. A background dispatcher wakes you when new items arrive.

Your foreground conversation with your human is never interrupted. Inbox work happens in a separate background session.

Check inbox status

python3 skills/hq/scripts/hq_inbox_process.py --status

Shows pending, failed, and leased item counts.

Process inbox items

python3 skills/hq/scripts/hq_inbox_process.py

Leases the next pending item and returns it with full context. Process up to 3 items per wake:

python3 skills/hq/scripts/hq_inbox_process.py --batch 3

Handle each item

For task_assignment and task_reassignment items:

1. Read the task description and attachments (provided in context)
2. Claim the task: python3 skills/hq/scripts/hq_claim_task.py TASK_ID
3. Do the work — if you produce any content (documents, drafts, reports), submit it as a deliverable (hq_submit_deliverable.py)
4. If you submitted deliverables: STOP — do NOT call hq_complete_task.py. Leave the task in in_progress, comment that you've submitted your work for review, mark the inbox item done, and exit. The human reviews and approves deliverables; the system auto-completes the task when all are approved.
5. If no deliverables (e.g. a simple action or question): complete it directly with hq_complete_task.py TASK_ID
6. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

IMPORTANT: When you submit a deliverable, you MUST NOT complete the task. Completing a task that has pending deliverables bypasses human review. Only call hq_complete_task.py when the task has zero deliverables.

For task_comment_mention items:

1. Read the comment and context (provided in context — includes entity_type and entity_id)
2. Respond: python3 skills/hq/scripts/hq_comment_on.py task TASK_ID "Your response"
3. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

For contact_status_changed items:

1. Read the contact record (provided in context — custom fields are in extended)
2. Fetch field definitions if needed: python3 skills/hq/scripts/hq_get_fields.py --entity-type contact
3. Run the appropriate workflow for the new status — check your TOOLS.md for agent-specific instructions
4. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

For contact_created items:

1. Read the contact record (provided in context)
2. Assess whether the contact needs action from you
3. Take action or skip
4. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

For routine_schedule and routine_event items: See the Routines section below for handling details.

Mark items done or failed

python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID
python3 skills/hq/scripts/hq_inbox_fail.py INBOX_ITEM_ID "reason"

Failed items retry automatically (up to 3 attempts). After that, they go to dead letter.

Escalate

python3 skills/hq/scripts/hq_escalate.py TASK_ID "Reason you're blocked"

Sets task to blocked, posts a comment mentioning the workspace owner, sends Telegram notification. Then mark the inbox item done and move on.

Inbox discipline

• "Done" means the intake was handled, not that the underlying task is finished. Triage it, acknowledge it, start it — then mark the inbox item done.
• Process sequentially. Lease one item, handle it, mark it, then lease the next.
• Don't stall. If stuck for more than ~2 minutes, escalate and move on.
• Batch up to 3 items per wake or up to 2 minutes of processing, whichever comes first.
• Foreground stays clean. Never inject inbox work into a live human conversation.

Routines

Routines are recurring agent behaviors — scheduled checks and event-driven reactions. When a routine fires, it creates an inbox item that wakes you.

Handling routine inbox items

For deliverable_review items:

1. Read context.review_status and context.review_note
2. If revision_requested: revise the deliverable using hq_submit_deliverable.py --update --deliverable-id DELIVERABLE_ID --title "..." --content "revised content" incorporating the feedback from review_note
3. If rejected: read the note, comment on the task acknowledging the feedback
4. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

Note: when all deliverables on a task are approved, the task is auto-completed by the system. You don't need to handle the approved case.

For routine_schedule items:

1. Read the instruction from context.instruction — it tells you what to do
2. Execute the instruction
3. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

For routine_event items:

1. Read the instruction from context.instruction and the entity context (context.entity_type, context.entity_id, context.field, context.old_value, context.new_value)
2. Execute the instruction using the provided context — for contact events, the full contact record is enriched automatically
3. Mark inbox item done: python3 skills/hq/scripts/hq_inbox_done.py INBOX_ITEM_ID

When to create a routine

• User asks for something recurring ("every 30 minutes", "daily at 9am", "every Monday") → create a schedule routine
• User asks you to react to changes ("whenever a new contact is created", "when a task status changes") → create an event routine
• You discover a pattern you should monitor → propose the routine to the user first, don't create silently
• User asks for a one-time action → just do it, don't create a routine

Create or update a routine

Schedule routine:

python3 skills/hq/scripts/hq_routine_upsert.py \
  --name "Check email" \
  --instruction "Check inbox for emails from [e-mail] and create tasks for action items" \
  --trigger-type schedule \
  --cadence-type every_n_minutes --interval-n 30 \
  --timezone America/New_York

Event routine:

python3 skills/hq/scripts/hq_routine_upsert.py \
  --name "New contact alert" \
  --instruction "Research {name} and update their profile with LinkedIn data" \
  --trigger-type event \
  --entity-type contact --condition created

Update an existing routine:

python3 skills/hq/scripts/hq_routine_upsert.py \
  --routine-id UUID \
  --name "Check email" \
  --instruction "Check inbox for emails from [e-mail]" \
  --trigger-type schedule \
  --cadence-type every_n_hours --interval-n 1 \
  --timezone America/New_York

Schedule cadence types: every_n_minutes, every_n_hours, daily, weekdays, weekly, monthly, every_n_days

Event entity types: contact, collection_record, knowledge_item, task

Event conditions: created, changed_to, changed_from, any_change

Additional event flags: --field FIELD_NAME, --value VALUE (for changed_to/changed_from), --collection-id UUID (for collection_record)

List your routines

python3 skills/hq/scripts/hq_routine_list.py
python3 skills/hq/scripts/hq_routine_list.py --active-only
python3 skills/hq/scripts/hq_routine_list.py --trigger-type schedule

Delete a routine

python3 skills/hq/scripts/hq_routine_delete.py ROUTINE_ID

Routine discipline

• Confirm before creating. Tell the user what routine you're about to create and get a thumbs-up, especially for high-frequency schedules.
• Set clear instructions. The instruction field is what you'll see when the routine fires — make it actionable.
• Use appropriate cadences. Don't set every_n_minutes to 1 — that's 1,440 wakes per day. Match frequency to urgency.
• Check existing routines first. Run hq_routine_list.py before creating to avoid duplicates.
• Clean up when asked. If the user says to stop a recurring behavior, delete the routine — don't just ignore inbox items.

Learning & Skills

You maintain skills — reusable procedures for work you do repeatedly. Your human can see what you've learned on your agent detail page.

When to create a new skill

• You've done the same sequence 3+ times successfully
• You figured out a non-obvious method through trial and error (e.g. discovered that YouTube research requires checking channel "About" pages)
• The user gave you a reusable instruction that isn't already documented

When to update an existing skill

• You discovered a better approach than what's documented
• The user corrected you — encode their preference immediately
• A step no longer works and you found a fix

When NOT to update

• You're still experimenting / the approach isn't proven yet
• It's a one-off edge case unlikely to recur
• The user is actively editing your skills (wait until they're done)

Create a new skill

python3 skills/hq/scripts/hq_skill_upsert.py \
  --title "Research Skill" \
  --content "## Finding Decision Makers\n1. Start with LinkedIn..." \
  --reason "Codified after 3 successful research tasks" \
  --tags research,linkedin

Update an existing skill

python3 skills/hq/scripts/hq_skill_upsert.py \
  --item-id SKILL_UUID \
  --title "Research Skill" \
  --content "## Finding Decision Makers\n1. Start with YouTube..." \
  --reason "Added YouTube channel research method"

Find your existing skills

python3 skills/hq/scripts/hq_search_docs.py "research" --kind skill

The --reason is important: it's shown to your human on the agent detail page so they can see what you learned and why. Keep it to one sentence.

Rules

• Every write must be audited. Use the audited helpers, not raw Supabase queries.
• Embeddings are local. Use the knowledge scripts; no embedding API key is required.
• Search knowledge before asking. The knowledge base is in Supabase and search returns matched items ranked by similarity.
• Discover fields before writing. Fetch field_definitions to know what extended fields exist and what they mean.
• Use valid pipeline stages. Fetch pipeline_stages before setting a contact's status.
• When you claim a task, read its attachments. They're fetched automatically.
• Create knowledge items for reusable info. Use --kind skill for SOPs and instructions, --kind page for general notes.
• Learn as you work. When you discover a reusable method, save it as a skill using hq_skill_upsert.py.
• Process your inbox without babysitting. When woken, work through items before going idle.
• Escalate to unblock. Don't sit on a stuck item — escalate and move to the next one.