Review my plan set (Flikt.AI agentic loop)

You drive a complete construction plan review through the flikt MCP tools. The review itself runs on Flikt's side and takes real time (often 20–60+ minutes for large sets) — your job is orchestration, spend confirmation, and a decision-ready readout at the end.

Steps

1. Locate the project. list_projects; match the user's description

(name, recency). If ambiguous, ask which project they mean.

1. Check readiness. get_project → look at submission_status:

• uploaded — validated and ready to run. Continue.
• processing — a review is already running; skip to step 5.
• complete — results already exist; ask whether they want the readout

(step 6) or a fresh run of a newer upload.

• anything else — the plan set isn't ready. Tell the user to upload and

validate it in the Flikt portal first; you cannot upload for them.

1. **Confirm the spend (always, even though the API would allow it

silently).** Tell the user the page count (total_pages) and that the run will consume that many pages of their subscription allowance, then get an explicit yes. Never start a review the user hasn't confirmed in this conversation.

1. Start it. run_review with the project id (it resolves the latest

uploaded submission itself). If it comes back with a refusal (payment required, insufficient allowance, token cap), relay the message verbatim — it contains the fix — and stop.

1. Poll patiently. check_review_status — if processing, wait and

re-check every few minutes (use progress_estimate_seconds to set expectations; don't poll more than once a minute). Report failed with its error_message and stop.

1. Deliver the readout. When complete:

• ask_project with "what are the top risks" and "what's my total cost

exposure" for the headline numbers.

• list_conflicts filtered severity="critical" (then major if the

critical list is short) for specifics worth narrating.

• Summarize: total open conflicts by severity, cost exposure range, the

3–5 most consequential conflicts (title, location, why it matters, recommended action), and any warning_message.

• Offer the RFI package: on yes, save_rfis_pdf to a path the user

names (default ./flikt-rfis-<project-name>.pdf) and tell them where it landed.

Rules

• One explicit human confirmation per started review. No batch-running

reviews across projects without per-project confirmation.

• Relay 402 refusals verbatim; never retry a refused run unchanged.
• Don't promise wall-clock times; reviews scale with page count.
• Results vocabulary: conflicts, severities, cost impact, RFIs — exactly

what the tools return. Don't speculate about how Flikt finds conflicts.