Verify Plan

Verify Plan V2

Convergence loop orchestrator for adversarial plan verification. Spawns fresh Task agents per step — no persistent agents, no agent-to-agent messaging. Main thread holds all state and serializes communication.

Pipeline Position

Plan written → /verify-plan (THIS) → [structural validator, if available] → Execution

Invocation

/verify-plan [path] [--team=base|scaled] [--threshold=N]

Path resolution order:

1. Explicit path argument (e.g., /verify-plan .claude/plans/my-plan.md)
2. Most recently modified file in .claude/plans/
3. Ask user for path

Override flags:

• --team=base — force 3-agent team regardless of complexity
• --team=scaled — force 5-agent team regardless of complexity
• --threshold=N — override complexity threshold (default: 7)

Team Composition

Base Team (complexity < 7, unknowns <= 3)

Fallback: If plan-resolver is not available, use general-purpose subagent type with the same prompt.

Scaled Team (complexity >= 7 OR unknowns > 3) — adds:

Execution Flow

Step 1: Discover Plan File

Resolve path using invocation order above. Confirm file exists and is readable. Read plan content.

Step 2: Complexity Assessment & Team Sizing

Assess plan quality (RICH / ADEQUATE / THIN / TRIVIAL) and complexity:

Decision:

• TRIVIAL plan: Skip team entirely. Output abbreviated PROCEED (trivial) with 1-2 sentence note. ~2K tokens. DONE.
• THIN plan: Force BASE team even if complexity score suggests SCALED.
• Score < 7 AND unknowns <= 3: BASE team (3 agents).
• Score >= 7 OR unknowns > 3: SCALED team (5 agents).
• Override: --team=base | --team=scaled | --threshold=N

Show assessment to user: "Complexity: N/14. Team: BASE|SCALED. Starting verification..."

Step 3: CONVERGENCE LOOP (max 3 iterations)

Initialize iteration_state per protocol.md schema. Each iteration runs Steps 3a-3d.

Step 3a: Challenge Phase

Base team: Spawn plan-challenger agent via Task tool:

Read plan at [path]. Execute full analysis (Ingest → Explore → Unknowns Manifest → Challenge Brief).
Iteration state: [serialized state or "first iteration"]
Return challenge brief with all challenges in Challenge Schema format + unknowns manifest.

Scaled team: Use TeamCreate to run 3 agents in parallel:

• plan-challenger — same prompt as above
• plan-domain-expert — "Read plan at [path]. Produce domain-specific challenges. Challenger produced: [paste challenger's challenges to avoid overlap]. Iteration state: [state]"
• plan-devils-advocate — "Read plan at [path]. Hunt for black-swan scenarios. Other agents produced: [paste their challenges]. Iteration state: [state]"

TeamDelete immediately after all 3 return. Merge all challenges into iteration_state.challenges (cap at 8 total).

Step 3b: Research Phase

Mode 1 — RESOLVE (plan-resolver): If unknowns manifest is non-empty, spawn plan-resolver (or general-purpose if unavailable):

Research the following unknowns from plan at [path]:
[paste unknowns manifest]
For each, resolve as CONFIRMED/REFUTED/UNRESOLVABLE/PARTIALLY_RESOLVED.
Format: [VERIFIED: source | finding | impact]

Update iteration_state.unknowns with findings.

Modes 2-3 — SURFACE + PROBE (plan-researcher): On iteration 1 only (unless Synthesizer directs RE-SWEEP or RE-PROBE): Spawn plan-researcher:

Plan at [path]. Goal: [plan goal]. Key technologies: [extracted from plan].
Iteration state: [serialized state including challenges and resolved unknowns]
Execute Mode 2 (SURFACE) then Mode 3 (PROBE).
Return surfaced contexts and probed risks in protocol schema format.

Merge findings into iteration_state. New challenges (capped at 2) added with appropriate origin tags.

Step 3c: Synthesis Phase

Spawn plan-synthesizer with FULL iteration state:

Evaluate all challenges, unknowns, surfaced contexts, and probed risks.
Iteration state: [full serialized state]
Produce: resolution updates, convergence check, directives (or final verdict if converging).
[If final iteration: include quality score per scoring-rubric.md]

Step 3d: Convergence Evaluation (main thread)

Read Synthesizer output. Update iteration_state.convergence.

Between iterations (non-blocking): Update user with progress summary.

Step 4: Final Verdict

If converged, give Challenger ONE audit shot. Spawn plan-challenger (brief):

Audit the Synthesizer's final verdict and quality score for plan at [path].
Synthesizer verdict: [paste verdict with scores]
If you disagree with any dimension score by 2+ points, flag the discrepancy with your evidence.
Otherwise, confirm the verdict.

Present final output to user:

• Verdict: PROCEED / REVISE / RETHINK
• Resolution table — all challenges with final status
• Quality score breakdown — report-only
• Score discrepancies — if any from Challenger audit
• Surfaced context summary — relevant findings nobody asked about
• Recommended next step:
  • PROCEED → "Run structural validation (if available), then execute"
  • REVISE → "Update plan with changes listed, then re-run /verify-plan"
  • RETHINK → "Fundamental issues found. Do NOT proceed to execution. Rework approach first."

Step 5: Cleanup

TeamDelete if scaled team was created during Step 3a.

User Checkpoint Protocol

Team runs autonomously except:

1. BLOCKING challenge unresolvable by agents → pause, present options
2. Convergence stall (no progress between iterations) → pause, confirm continue
3. Max iterations with unresolved BLOCKING → pause, present findings
4. Contradictory research findings → pause, request tiebreak

Between iterations (non-blocking): "Iteration N: X resolved, Y remaining. Continuing..."

Communication Topology

Main thread is the hub. No direct agent-to-agent messaging. Each agent receives full input from main thread and returns output to main thread.

Main Thread (orchestrator + state holder)
    ├──[spawn]──→ Challenger ──[return]──→ challenges + unknowns
    ├──[spawn]──→ Researcher ──[return]──→ findings + surfaced + probed
    ├──[spawn]──→ Synthesizer ──[return]──→ resolution state + convergence
    └──[evaluate]──→ CONVERGED? / PAUSE? / CONTINUE?

Token Budget

Integration Points

• Designed to run before any structural validation tool (if available in your setup)
• RETHINK → skip downstream validation entirely
• REVISE → user updates plan, optionally re-runs /verify-plan
• PROCEED → continue to structural validation or execution

Anti-Patterns — Do NOT

• Check structure — a structural validator's job, not this skill's
• Review code — a dedicated code review tool's job, not this skill's
• Use persistent agents — fresh Task per step
• Allow agent-to-agent messaging — main thread serializes all state
• Use numeric scores for loop control — binary resolution tracking only
• Exceed challenge caps — 5/iteration, 8 total
• Skip complexity assessment — always assess before team sizing
• Run Modes 2-3 on iterations 2-3 — unless Synthesizer directs RE-SWEEP/RE-PROBE
• Research inside plan-challenger — unknowns manifest only, research delegated
• Create TeamCreate for base team — only for scaled team's parallel challenge phase