Vc Predict

vc-predict — Multi-Persona Pre-Analysis

Output style: Follow process/development-protocols/communication-standards.md — answer-first, plain language, no unexplained jargon, TL;DR on long responses.

Five expert personas independently analyze a proposed change, then debate conflicts to produce a consensus verdict before a single line of code is written.

When to Use

• Before implementing a major or high-risk feature
• Before a significant refactor or architecture change
• Evaluating competing technical approaches
• Stress-testing assumptions in a proposed design

When NOT to Use

• Trivial or low-risk changes (use debugger for bugs, generate-plan / plan-agent for already-decided tasks)
• Already-approved work with no open design questions
• Pure dependency upgrades with no API changes

---

Mode Selection

vc-predict runs in Simple or Deep mode. Choose based on the conditions below.

Deep Mode — Trigger Conditions (any one is sufficient)

• The approach involves a pattern previously attempted in this codebase (git history may show prior attempts)
• The approach touches a surface with known failure history: auth, billing, container lifecycle, streaming, or WebSocket reconnect
• Caller explicitly requests deep mode (--deep flag or "use deep predict")
• The plan is COMPLEX shape and this is the pre-checklist predict call

Simple Mode — Trigger Conditions (default)

• Approach is a contained feature with clear scope
• No prior attempts at this surface area are likely
• Plan is SIMPLE shape and the design is not controversial

---

Deep Mode — Research Subagent Protocol

Before the 5-persona debate, spawn a research subagent that performs the following steps in order:

1. Git history scan — run git log --oneline --all -- [relevant files] for each file or directory the approach touches. Flag any commits within the last 6 months that suggest a prior attempt, revert, or known fix.
2. Prior phase reports — search task folders under process/features/{feature}/active/ and process/features/{feature}/completed/ for _REPORT_ files mentioning this surface (pattern: active/{slug}_{date}/{slug}_REPORT_{date}.md). Read any that contain keywords from the approach (e.g. "streaming", "SSE", "auth", "billing", "container lifecycle"). Legacy sibling reports/ dir may also exist — scan it too if present.
3. Test failure history — grep test output files and CI logs for FAIL or Error patterns on relevant module names. Note recurring failures.
4. Return a Historical Context block containing:
   • What was tried before (commit refs, dates, brief description)
   • What failed and why
   • Current state of the surface (is it stable, under active churn, recently refactored?)
   • Known landmines (specific lines, patterns, or edge cases that broke before)

The 5 personas then receive this Historical Context block before their independent analysis phase. Each persona incorporates it when relevant.

Output Quality Difference

Simple predict (no research):

"Senior dev: this streaming approach could cause a memory leak in the SSE proxy."

Deep predict (with historical research):

"Senior dev: we tried this exact streaming approach in commit a3f921 (2026-03-15) — it caused a memory leak in the SSE proxy because the Bun response body was never released on client disconnect. The current proposal has the same pattern in packages/api/src/routes/gateway-proxy.ts line 84. The fix at the time was adding an AbortController listener; verify that is still present or re-apply."

---