psquared-development ProcessFlow — Work one process (param: process name)
Input: one process name (the arg). Goal: bring it to guardrail-completion. Load
processflow-review/references/guardrails.md (Definition of Done + Fehlermuster-Register) and
processflow-app (Chrome-MCP + copilot mechanics). If a TodoWrite helps, create one todo per
required guard.
0. Locate + load
Find the process by name in data/<company>/processes.json (get its UUID). Pull the current
canvas + report + score fresh from Supabase via the Chrome session token (localStorage
sb-…-auth-token → access_token; anon apikey) — a colleague may have edited it. Read the relevant
transcript passages for THIS process.
1. Transcript-check first (ground truth)
Before reworking, verify the process against the transcript (canvas_wb, transcript,
scenario_correct, industry_correct, no_invent): Is it real? Is the SCENARIO right (what actually
happens)? Correct industry/products? Any invented tools/brands/devices/ERP names (grep them)? This
sets what needs fixing.
2. Rework via Chrome-MCP (writes hit REAL customer data)
Use the Konzept-Copilot (floating button, "Vom Copilot überarbeiten lassen") for the
Lösungskonzept: it edits report text + diagrams + solutions formatting-preserving in one run. Use the
Canvas-Copilot (Canvas tab) for canvas fields — remember to click Übernehmen. Give precise,
scoped instructions. After each run: reload + verify (don't trust "done" before completion). Apply the
substance guardrails, not just surface: use_existing_capabilities (does the ERP/a product already do
it? build only the delta — Build-vs-Buy in Schritt 0), right_ai_technique (research the correct
approach — embeddings vs VLM, rules vs ML), smallest_fit/ki_solution_sense, discovery_first for
unclear/large/sensitive, os_module_option for frontends, pricing_verified (web-check any
price/"included" claim), hours_real (AI under-estimates), analysis_per_phase, roles_not_initials,
diagram guards (diag_simple, diag_caption_clean, diag_storyline — E2E starts at the user action,
top-down, ≤~10 nodes in architecture, no jargon in captions).
3. Re-score the Prozessanalyse
If the rework changed effort/approach/hours, re-run "Durch KI bewerten" (Prozessanalyse tab) so the score matches the concept (else quadrant/effort stay stale). Sanity-check the result; if the AI still low-balls effort, flag it (score is a human call).
4. Keep the helper app synced (always)
After each meaningful change, POST /api/review (helper app, localhost:8765) to set the guard flags +
notes for this process. Keep skill and app in lockstep. If the app server is down, restart it
(python3 app.py).
5. Done criteria
Every required guard is true, or set to n.a. with a documented reason (e.g. arch_diag n.a. for a single-tool process; compliance_synced n.a. = read-only field → flag). Fresh token when the ~1h JWT expires (pull again from Chrome localStorage).
Parallel-Betrieb (du bist evtl. einer von mehreren)
Andere Agenten bearbeiten parallel andere Prozesse derselben Firma. Vor dem Rework den Team-Update-Feed
lesen (GET /api/updates?company=<slug>) und eigene Querschnitts-Findings teilen
(POST /api/update {company,agent,process,tags,text}): erfundene/echte Tools, Branche, geteilte
Module/Eingänge (InboxMate, Artikel-Matching), OS-Konsolidierung, Preis-/Lizenz-Fakten, System-of-Record.
Prozess-Details → review.json; nur Übergreifendes → Feed. Nur DEINEN Prozess anfassen.
Hard rules
- NEVER set
approved/HI-freigegeben (human) or name the presentation date (Dominik). - Every write is live customer data — outside an explicitly-authorized batch run, confirm each write. In a user-authorized multi-terminal batch, proceed but stay precise and verify.
- Output: per-guard status + notes for this one process (hand back to the wrapper / user).