Check

All numbers come from the portfolio engine, a plain script (the session context gives its exact node .../pc.js COMMAND line; pc below). Show its text as is; never recompute. Broker data comes from the brokers' own MCP tools: you call their read tools and a hook saves what they return, so you never retype a broker number.

1. Connect (one message)

Kotak Neo reads need the sessionid from its login, so start Kotak with get_login (UCC from the profile), the QR step, then validate_login. Then read with the broker MCPs, no limits or paging: Kite get_holdings, get_gtts, get_margins, get_mf_holdings, get_orders, get_trades; Kotak Neo get_holdings, get_limits (with the sessionid). Then run pc broker_status (pass --folder = the investor's folder, and the hook plugin and version from the session context).

• Live brokers: say nothing more.
• needs_login, or any read that returned an error: Kite: call the Kite MCP login tool, show its link only, wait for "done", read again. Kotak: get_login, the QR step, validate_login, read again. If reads still fail right after a successful login, treat it as an outage; do not ask again.
• outage: say the broker is down and do not ask for a login.
• not_connected: only when that broker's MCP tools are not in this chat at all. Say which, and go on with the others. If the tools exist but a read errored, that is a login, above.
• Never write a login URL yourself.

Profile when the chat is not linked to the PC

If the investor's folder is not reachable, ask them to attach profile.json (kept in the claude.ai Project). Save it into a scratch folder, pass that as --folder, and tell them plainly: "the profile is a copy". After any profile_update or log_append, hand the changed file back for them to re-upload; the copy is stale until they do.

Profile when the chat is linked to the PC

pc.js runs in the container and needs --folder there, so use stage, run, write back:

1. Stage: through the device link, copy profile.json, log.md and the latest snapshots from the investor's folder into a scratch folder in the container.
2. Run pc against the scratch folder.
3. Write back: copy the changed profile.json, log.md, snapshots and feedback.md to the investor's folder on the PC before ending, and say which files were written. If the link drops, say the PC copy is stale. If the device shell can run Node, running pc.js on the PC directly (with the real folder) is simpler and needs no staging.

2. Pull

Run pc refresh. If an IBKR connector is available, call its get_account_summary (primary), then get_account_positions, and pass them to ingest with broker: "ibkr", currency: "USD", and fx_rate + fx_source (a dated USD/INR rate; say where it came from). If summary fails, ingest positions and say the cash figure is last known. For a broker with no connection, "import my holdings": ask for a CSV export and call ingest with csv_path.

3. Macro (morning and mid-day only)

Search: US close, Asian session, GIFT Nifty or Nifty, Brent, USD/INR, today's India events. Before using any level, confirm its date is today's date; otherwise label it "not same-day" and keep it out of the verdict. Give setups and levels, never direction calls. Also search held stock symbols for upcoming bonus, split or demerger ex-dates. If any, run the gtt skill's audit with corporate_actions (mark unconfirmed dates as confirmed: false).

4. Report

1. One-line verdict that agrees with every number shown (re-read before sending; the weaker of global and domestic wins).
2. The refresh text: broker lines, needs attention, changes since last snapshot, brokers, positions, allocation.
3. Macro table (morning) and any corporate-action GTT risk.
4. Close (EOD): add gtt_audit summary line. SIP status is not readable from any broker: call protection_check to show the SIPs the investor has confirmed; if none are saved, ask, don't infer, and save them with profile_update only after they confirm.

5. Log

Call log_append with a 3-5 line entry (what was pulled, totals, what changed, anything flagged). Rewrite start_here only when goals, next actions or decisions owed change. Never edit a verdict; log the contradiction.

Open actions ("work through my to-dos")

1. Read the open actions: the Todoist project for portfolio actions (Todoist tools, find-tasks; ask which project if unclear), else the "Start here" block of the log.
2. Take them one at a time. Map each to the skill or tool that does it: GTT work to gtt, a stock review to analyze, risk, buffer or rebalancing to risk, a number the investor holds to profile_update, SIP, cover and buffer inputs to protection_check.
3. Run it. Ask the investor only for inputs the plugin cannot get (a number, a decision). Never guess one.
4. Complete the Todoist task (complete-tasks) only after the work is done and verified, and log_append it. An action no tool can do stays open: call report_gap for it.

Gaps

When a tool says "not supported", a known limitation, or errors with no fallback, call report_gap (title, what was attempted, what is missing, the error). Never put quantities, prices or P&L in it. If the result says no issue was filed (no token, or GitHub rejected it) and a GitHub MCP is connected in this chat, file it yourself: search open issues labelled gap for the same title and comment on it, otherwise create the issue with label gap in arunendapally/portfolio-copilot, using the same text report_gap logged and no investor numbers. Then tell the investor where it was logged and filed.

Scheduled run (investor away)

Do not wait for logins. Run broker_status once, then check (last snapshot, clearly dated) and the macro read, then one line per broker that needs login. Nothing is presented as live unless the tool says live.