SecondeyeSAFE

CocoPlus is an Agentic Operating System for Snowflake Coco. It brings structured, multi-agent workflows to data engineering projects — covering everything from project initialization through spec, plan, build, test, review, and ship phases.

The instruction file, verbatim from the audited commit — this is the text the model reads, and the surface the audit's instruction layer examines. Quoted here so you can judge it without cloning anything.

---
name: "secondeye"
description: "Multi-model parallel plan critique. Spawns five SecondEye Critic instances in parallel (Efficiency, Completeness, Risk, Devil's Advocate, Edge Case Hunter), aggregates findings with HITL/AFK, BLOCKING/MINOR, and six-severity labels, and writes a structured report. Critical findings create a soft gate on $build. Usage: $secondeye [--artifact spec|plan|review] [--model haiku|sonnet|opus]."
version: "1.2.0"
author: "CocoPlus"
tags:
  - cocoplus
  - secondeye
---

Your objective is to run a multi-model adversarial critique of a lifecycle artifact.

Before proceeding, verify that `.cocoplus/` exists.
If not: output "CocoPlus not initialized in this directory. Run `$pod init` to begin." Then stop.

## Shadow Rule Commands

Shadow rules are candidate SecondEye rules evaluated silently alongside active rules — their findings never affect the verdict, but their activation rate and developer-acceptance rate accumulate evidence for promotion.

### `$secondeye shadow-add`

Interactively collect a rule definition (lens, trigger pattern, description) and write it to `.cocoplus/secondeye/shadow-rules.json`. Do not add the rule to the active critic pool.

### `$secondeye shadow-report`

Invoke `secondeye/shadow-report`. Reads `.cocoplus/secondeye/shadow-findings.json`, computes per-rule activation rate (how often the rule fires across reviews) and developer-acceptance rate (how often a surfaced shadow finding matches a finding the developer subsequently acted on), and reports promotion readiness against configured thresholds.

### `$secondeye shadow-promote <rule-id>`

Move a shadow rule from `.cocoplus/secondeye/shadow-rules.json` into the active critic rule pool once `shadow-report` shows it meets promotion thresholds. Removing it from the shadow pool after promotion.

## Parse Arguments

- `--artifact [spec|plan|review]`: target artifact (default: plan)
- `--model [haiku|sonnet|opus]`: single-model mode (default: five-lens parallel)

Resolve artifact path:
- spec → `.cocoplus/lifecycle/spec.md`
- plan → `.cocoplus/lifecycle/plan.md`
- review → `.cocoplus/lifecycle/review.md`

Verify the artifact file exists. If not: output "Artifact [file] not found. Run the appropriate lifecycle phase first." Then stop.

## Create Staging Directory

Generate timestamp: `YYYYMMDD-HHMMSS`
Create: `.cocoplus/lifecycle/.secondeye-staging/[timestamp]/`

## Spawn Critics

If single-model mode (`--model` flag provided): spawn only one critic with that model and the Completeness lens.

Otherwise: spawn five critics IN PARALLEL:

**Critic 1 — Efficiency Lens (Haiku)**
Write task prompt to `.cocoplus/lifecycle/.secondeye-staging/[timestamp]/haiku-task.md`:
```
You are a SecondEye Critic. Your lens is: EFFICIENCY.
Artifact to review: [artifact path]
Staging file: .cocoplus/lifecycle/.secondeye-staging/[timestamp]/haiku-findings.md

Efficiency lens: redundant steps, over-engineered solutions, token-expensive approaches, too many passes.

Assign each finding a `severity` label: blocking | important | nit | suggestion | learning | praise
At least one `praise` finding is required if any well-constructed pattern is present.
```

**Critic 2 — Completeness Lens (Sonnet)**
Write task prompt to `.cocoplus/lifecycle/.secondeye-staging/[timestamp]/sonnet-task.md`:
```
You are a SecondEye Critic. Your lens is: COMPLETENESS.
Artifact to review: [artifact path]
Staging file: .cocoplus/lifecycle/.secondeye-staging/[timestamp]/sonnet-findings.md

Completeness lens: missing requirements, unresolved dependencies, gaps between spec and plan, unhandled edge cases.

Assign each finding a `severity` label: blocking | important | nit | suggestion | learning | praise
At least one `praise` finding is required if any well-constructed pattern is present.
```

**Critic 3 — Risk Lens (Opus)**
Write task prompt to `.cocoplus/lifecycle/.secondeye-staging/[timestamp]/opus-task.md`:
```
You are a SecondEye Critic. Your lens is: RISK.
Artifact to review: [artifact path]
Staging file: .cocoplus/lifecycle/.secondeye-staging/[timestamp]/opus-findings.md

Risk lens: hard-to-reverse decisions, external dependencies that could fail, data loss scenarios, security gaps, unvalidated assumptions.

Assign each finding a `severity` label: blocking | important | nit | suggestion | learning | praise
At least one `praise` finding is required if any well-constructed pattern is present.
```

**Critic 4 — Devil's Advocate Lens (Sonnet)**
Write task prompt to `.cocoplus/lifecycle/.secondeye-staging/[timestamp]/da-task.md`:
```
You are a SecondEye Critic. Your lens is: DEVIL'S ADVOCATE.
Artifact to review: [artifact path]
Staging file: .cocoplus/lifecycle/.secondeye-staging/[timestamp]/da-findings.md

Devil's Advocate lens: Find the strongest possible argument that this plan should NOT proceed as written.
Identify ONE of: the single most damaging assumption, the most likely architectural failure, or the most serious underspecified requirement.
Do not seek balance. Your mandate is adversarial rigor — construct the most compelling case for why this plan will fail.

All your findings are classified BLOCKING by default.
Assign severity `blocking` to all DA findings by default.
For each finding, include a required "rebuttal_score" field (1–5) when the developer responds:
- Score 4–5: concede (response is sufficient)
- Score 1–3: re-assert the concern and explain why the response is insufficient
```

**Critic 5 — Edge Case Hunter Lens (Haiku)**
Write task prompt to `.cocoplus/lifecycle/.secondeye-staging/[timestamp]/edge-task.md`:
```
You are a SecondEye Critic. Your lens is: EDGE CASE HUNTER.
Artifact to review: [artifact path]
Staging file: .cocoplus/lifecycle/.secondeye-staging/[timestamp]/edge-findings.md

Edge Case Hunter lens: identify evaluation blind spots, rare input shapes, data distribution shifts, failure cases not represented in the plan, and acceptance criteria that would miss bad behavior.

Assign severity `important` by default; use `

SAFEgrade B · trust 89/100 Nothing in the source contradicts what it says it does. Grade A is reserved for packages that have also passed the behavioural sandbox.

What the source does

Filesystem

none-observed

Network

none-observed

Shell

none-observed

Dependencies

pinned

Secrets in source

none-found

Findings (0)

No findings outside the package's declared scope.

Gates applied: no_behavioural_pass.