agent-model-audit

name	agent-model-audit
description	Use when auditing agent models, reviewing provider offerings, optimizing model assignments by cost and speed, or updating agent descriptions with cost/speed indicators. Trigger keywords: "audit agents", "audit models", "update agent models", "review models", "optimize agents".

Agent Model Audit

Audit connected providers, analyze model offerings (cost, speed, usage limits), and optimize agent-to-model assignments. Produces a dated audit report and updates agent configurations.

Important: The current model assignments in opencode.jsonc are the output of this process, not the input. Discard them at the start. All assignments will be determined fresh from the data gathered in steps 1–5.

Priority Mode

The optimization priority is determined by an optional argument passed to the command:

Argument (EN/ES)	Priority Order
skill / ability / habilidad (default)	Role fit → Cost → Speed
price / cost / precio / barato	Cost → Role fit → Speed
speed / fast / velocidad / rapido	Speed → Role fit → Cost

If no argument was provided, use skill mode (role fit first).

Workflow

1. Discover Available Models

Run the following command to get the authoritative list of models available in opencode:

opencode models

This returns the full list of provider/model identifiers actually configured and accessible. This is the only source of truth — do not use general knowledge or assumptions about what models exist.

From the output:

• Note every provider/model identifier
• Group by provider
• Identify which providers are active (not in disabled_providers in opencode.jsonc)

Critical: Only models that appear in opencode models output are candidates for agent assignment.

⚠️ Model identifiers must match EXACTLY — character for character. Pay close attention to punctuation. Model names use hyphens, not dots:

• ✅ opencode/claude-sonnet-4-6 (correct — hyphens)
• ❌ opencode/claude-sonnet-4.6 (incorrect — dot)
• ✅ opencode/gpt-5.4 (correct — dot is part of the version)
• ❌ opencode/gpt-5-4 (incorrect — hyphen instead of dot)

Never guess or normalize model names. Always copy-paste the exact identifier from the opencode models output. This applies to all steps that reference model names (steps 2, 3, 5, 6, 7, 8, and 9).

Also read opencode.jsonc to collect:

• Agent names and their roles (from descriptions and agents/*.md files)
• Current variant values — note which agents already have a variant field set. The variant differentiates configurations of the same base model (e.g., max, high, low). Ignore current model assignments — those will be replaced.

Filter Out Models

Check if the optional filter file exists at skills/agent-model-audit/filter-out.md. If it exists, read its content and apply it as strict exclusion rules before proceeding:

test -f skills/agent-model-audit/filter-out.md && cat skills/agent-model-audit/filter-out.md

If the file does not exist, skip this step — all models from opencode models remain candidates.

The filter-out.md file is a git-ignored markdown file that may contain:

• Exact model identifiers to exclude (e.g., opencode-go/some-model)
• Patterns (e.g., *experimental*, opencode-go/legacy-*)
• Conditional rules (e.g., exclude models with cost > $X, exclude models slower than Y t/s)
• Provider-level exclusions (e.g., exclude all models from a specific provider)

⚠️ CRITICAL: The contents of filter-out.md are strict rules, not suggestions. Any model matching a filter-out rule MUST be excluded from ALL subsequent steps. This takes precedence over any other selection criteria. Apply the rules as literal as possible — if a rule is ambiguous, err on the side of exclusion.

Add excluded model identifiers (if any) to a temporary note for reference in the audit report.

Filter Out Models

Check if the optional filter file exists at skills/agent-model-audit/filter-out.md. If it exists, read its content and apply it as strict exclusion rules before proceeding:

test -f skills/agent-model-audit/filter-out.md && cat skills/agent-model-audit/filter-out.md

If the file does not exist, skip this step — all models from opencode models remain candidates.

The filter-out.md file is a git-ignored markdown file that may contain:

• Exact model identifiers to exclude (e.g., opencode-go/some-model)
• Patterns (e.g., *experimental*, opencode-go/legacy-*)
• Conditional rules (e.g., exclude models with cost > $X, exclude models slower than Y t/s)
• Provider-level exclusions (e.g., exclude all models from a specific provider)

⚠️ CRITICAL: The contents of filter-out.md are strict rules, not suggestions. Any model matching a filter-out rule MUST be excluded from ALL subsequent steps. This takes precedence over any other selection criteria. Apply the rules as literal as possible — if a rule is ambiguous, err on the side of exclusion.

Add excluded model identifiers (if any) to a temporary note for reference in the audit report.

2. Research Model Characteristics and Validate URLs

For each model listed in step 1, gather from provider documentation:

• Cost per request — from provider pricing pages
• Requests per time window — from rate limit docs
• Speed/latency — fast / normal / slow
• Context window — maximum input tokens
• Best use case — reasoning, coding, analysis, quick tasks, etc.

Key sources per provider:

• opencode-go: https://opencode.ai/docs/go/#usage-limits
• opencode-zen: https://opencode.ai/docs/zen/#pricing
• GitHub Copilot (pricing): https://docs.github.com/en/copilot/reference/copilot-billing/models-and-pricing
• GitHub Copilot (multipliers): https://docs.github.com/en/copilot/reference/copilot-billing/model-multipliers-for-annual-plans
• OpenRouter / Zen: https://openrouter.ai/models
• Other providers: consult their official documentation

URL Validation — REQUIRED:

For every URL listed above that you navigate to:

1. If the page loads successfully → use it as-is.
2. If the page returns 404 or redirects to an unrelated page → search for the correct URL, then update this SKILL.md with the correct URL before continuing. This ensures future executions use the correct URL.

3. Collect Concrete Model Data

For each confirmed-available model, extract the following numeric values from your step 2 research:

• Intelligence Score (0–100): A numeric capability score based on reasoning and coding benchmarks (MMLU-Pro, GPQA, HumanEval, SWE-bench, etc.) and/or community consensus. The goal is a single comparable number per model. If multiple benchmarks are available, use a weighted average.
• Cost per Million Tokens (USD): The provider's listed cost per 1M input tokens. Use the input token price as the standard comparator.
• Speed (tokens/second): Measured output tokens per second from provider docs or community benchmarks. If only latency is available, convert to a comparable throughput estimate.

Important: These must be actual numeric values from research, not subjective categories. If exact numbers aren't available, estimate conservatively and note the source.

Record the data in a structured JSON file at docs/agent-audits/model-data.json for later use with the scale calculator script. Example:

{
  "models": {
    "provider/model-name": {
      "intelligence": 82,
      "cost": 3.50,
      "speed": 120
    }
  }
}

4. Identify Agent Roles

Read each agents/*.md file and opencode.jsonc agent descriptions. For each agent, determine its role:

Role archetype	Needs
Reasoning/architecture	Deep thinking, extended context
Implementation	Precision, balanced speed
Quick tasks	Fast, low latency
Large context analysis	High context window
Execution/terminal	Reliable, fast, good at structured output
General purpose	Balanced across all dimensions

5. Propose Optimized Assignments

Assign each agent a model from the step-1 list. Apply the Priority Mode from the top of this skill to determine the ordering of trade-offs.

Default (skill mode) — Role fit → Cost → Speed:

1. Find the model whose capabilities best match the agent's role archetype (step 4)
2. Among models equally suited to the role, prefer the cheaper one
3. Among equally cheap models, prefer the faster one

price mode — Cost → Role fit → Speed:

1. Start with the cheapest models and work up
2. Among equally cheap models, pick the one that best fits the role
3. Among equally good fits at the same price, prefer the faster one

speed mode — Speed → Role fit → Cost:

1. Start with the fastest models
2. Among equally fast models, pick the one that best fits the role
3. Among equally good fits at the same speed, prefer the cheaper one

All modes — non-negotiable constraints:

• No duplicate model+variant — each agent must use a unique model+variant combination. Same model with different variant values IS allowed (e.g., deepseek-v4-flash with variant: max and variant: high). If two agents share a model, they MUST have different variant values.
• Only models from step 1 — never assign a model not confirmed available
• No assignment left blank — every agent must have a model
• Exact identifier match — model names must be copied verbatim from step 1 output, including all hyphens, dots, and version numbers. Never modify, normalize, or guess identifiers.
• Respect filter-out rules — models excluded by skills/agent-model-audit/filter-out.md (if the file exists) must never be assigned to any agent.
• Respect filter-out rules — models excluded by skills/agent-model-audit/filter-out.md (if the file exists) must never be assigned to any agent.

Save the resulting assignments to docs/agent-audits/assignments.json for use in step 6. If any agent uses a model shared with another agent, include the variant field to differentiate them:

{
  "build": "opencode-go/deepseek-v4-flash",
  "x-teach": "opencode-go/deepseek-v4-flash",
  "z-spark": "opencode-go/deepseek-v4-free"
}

The assignments dict maps agent names to model identifiers (for scale calculator compatibility). Variant info is recorded in the audit report and opencode.jsonc only.

6. Calculate Scale-Based Indicators

After assigning models to all agents (step 5), compute the quality, cost, and speed indicators relative to the selected set of models only.

Run the scale calculator script with your collected data and assignments:

python skills/agent-model-audit/scripts/scale-calculator.py \
  --data docs/agent-audits/model-data.json \
  --assignments docs/agent-audits/assignments.json \
  --format json \
  --output docs/agent-audits/scale-results.json

The JSON output contains value, tier, icon, and label for each dimension of each selected model.

Algorithm:

For each dimension:

1. Collect the numeric values for all assigned models
2. Apply the dimension-specific rule for ---
3. Find the relevant minimum and maximum values
4. Calculate the range: range = max - min
5. Divide the relevant range into equal segments
6. Assign icons based on which segment each model falls into:

Dimension Rules

• Intelligence: --- means Basic. Use all intelligence values and split the full range into 4 equal segments.
• Cost: --- means Free. Use --- only when cost is exactly 0. Split the positive-cost range into 3 equal segments for the remaining tiers.
• Speed: --- means Unknown. Use --- only when the model has no speed value. Split the known-speed range into 3 equal segments for the remaining tiers.

Dimension	Segment 1	Segment 2	Segment 3	Segment 4
Intelligence	--- (0 — Basic)	󰫣 (1 — Good)	󰫣󰫣 (2 — Very Good)	󰫣󰫣󰫣 (3 — Excellent)
Cost	--- (0 — Free)	 (1 — Economic)	 (2 — Normal)	 (3 — Expensive)
Speed	--- (0 — Unknown)	󱐋 (1 — Slow)	󱐋󱐋 (2 — Normal)	󱐋󱐋󱐋 (3 — Fast)

Edge cases:

• If intelligence has no spread (range = 0), assign the second segment (1 icon) to all models.
• If all selected costs are 0, assign --- to all cost values.
• If the positive cost range has no spread, assign the second non-zero segment (2 icons) to all positive-cost models.
• If no speed values are known, assign --- to all speed values.
• If the known speed range has no spread, assign the second non-zero segment (2 icons) to all known-speed models.

Note: For intelligence and speed, more icons = better. For cost, fewer icons = cheaper (better).

7. Create Audit Report

Write to docs/agent-audits/YYYY-MM-DD.md:

# Agent Model Audit — YYYY-MM-DD

## Priority Mode

`skill` (role fit first)

## Providers Connected

| Provider    | Models Available | Notes                  |
| ----------- | ---------------- | ---------------------- |
| opencode-go | 12               | Go subscription active |

## Model Analysis

Numeric data collected in step 3:

| Model             | Intelligence | Cost ($/1M tok) | Speed (t/s) |
| ----------------- | ------------ | ---------------- | ----------- |
| deepseek-v4-flash | 65           | 0.15             | 200         |
| minimax-m2.5      | 60           | 0.20             | 180         |
| qwen3.5-plus      | 72           | 0.35             | 150         |
| qwen3.6-plus      | 78           | 1.50             | 100         |
| deepseek-v4-pro   | 85           | 2.00             | 90          |
| minimax-m2.7      | 80           | 1.80             | 170         |
| mimo-v2.5-pro     | 88           | 8.00             | 45          |
| kimi-k2.6         | 82           | 6.00             | 50          |
| glm-5.1           | 90           | 10.00            | 40          |
| claude-sonnet-4-6 | 95           | 15.00            | 80          |

## Scale Calculation

Range divided per dimension (computed in step 6):

| Dimension    | Min / Scope      | Max / Scope      | Range | Q1 (< )     | Q2 (< )     | Q3 (< )     | Q4 (>=)     |
| ------------ | ---------------- | ---------------- | ----- | ----------- | ----------- | ----------- | ----------- |
| Intelligence | 60               | 95               | 35    | 60–69       | 69–78       | 78–87       | 87–95       |
| Cost         | 0.15 (positive)  | 15.0 (positive)  | 14.85 | 0.15–3.86   | 3.86–7.58   | 7.58–11.29  | 11.29–15.00 |
| Speed        | 40 (known)       | 200 (known)      | 160   | 40–80       | 80–120      | 120–160     | 160–200     |

## Proposed Assignments

| Agent   | Variant         | Model             | Intelligence | Cost | Speed | Role Fit  | Rationale           |
| ------- | --------------- | ----------------- | ------------ | ---- | ----- | --------- | ------------------- |
| build   | `max`           | deepseek-v4-flash | 󰫣󰫣          |     | 󱐋󱐋    | Good      | Balanced for coding |
| x-teach | `high`          | deepseek-v4-flash | 󰫣󰫣          |     | 󱐋󱐋    | Good      | Teaching companion  |
| z-spark | —               | deepseek-v4-flash | 󰫣           |     | 󱐋󱐋󱐋   | Excellent | Fast & cheap        |
| ...     | ...             | ...               | ...          | ...  | ...   | ...       | ...                 |

## Changes from Previous Config

- Assigned `z-forge` → `mimo-v2.5-pro` (precision coding, best role fit)
- Assigned `explore` → `minimax-m2.5` (fast exploration, lowest cost)
- Updated all agent descriptions with `[quality | cost | speed]` indicators

## Uniqueness Check

All agents use unique models: ✅

8. Update Configurations

All model assignments must come exclusively from models confirmed in step 1.

⚠️ CRITICAL: Before writing any model identifier to opencode.jsonc, verify it character-for-character against the opencode models output from step 1. Common mistakes: replacing hyphens with dots (claude-sonnet-4-6 → claude-sonnet-4.6) or vice versa. Always copy-paste, never type from memory.

opencode.jsonc:

• Set agent.<name>.model for every agent from the proposed assignments in step 7
• Set agent.<name>.description with [quality | cost | speed] prefix
• If an agent shares a model with another agent, set agent.<name>.variant to the variant value from step 7. Agents using a unique model do NOT need a variant field.

agents/*.md:

• Do NOT add or modify description in agent .md frontmatter.
• The description field with [quality | cost | speed] indicators must ONLY exist in opencode.jsonc.
• Agent .md files contain the agent's behavioral instructions (mode, tools, system prompt).
• Duplicating descriptions in both places creates drift and confusion.

Format: [quality | cost | speed] Description text

• Quality padded to 4 chars: --- (Basic), 󰫣 (Good), 󰫣󰫣 (Very Good), 󰫣󰫣󰫣 (Excellent)
• Cost padded to 4 chars: --- (Free only when cost is 0),  (Economic),  (Normal),  (Expensive)
• Speed padded to 4 chars: --- (Unknown only when speed is missing), 󱐋 (Slow), 󱐋󱐋 (Normal), 󱐋󱐋󱐋 (Fast)
• Example: [󰫣󰫣 |  | 󱐋󱐋 ] Reasoning architect for decomposition and design decisions

9. Verify

• No duplicate model+variant across agents (same model + same variant never appears twice) ✅
• All agent descriptions have [quality | cost | speed] indicators ✅
• Scale calculation ran on selected models only ✅
• Icons assigned by range quarters (Q1 → ---, Q2 → 1 icon, Q3 → 2 icons, Q4 → 3 icons) ✅
• All assigned models came from opencode models output ✅
• All model identifiers match opencode models output character-for-character ✅
• Models excluded by filter-out.md are NOT assigned to any agent ✅
• Audit report saved to docs/agent-audits/YYYY-MM-DD.md ✅
• opencode.jsonc is valid JSONC ✅

Notes

• docs/agent-audits/ should be gitignored (working documents, not committed)
• Config changes (opencode.jsonc) should be committed
• Agent .md files should only be committed if their behavioral content (mode, tools, system prompt) was modified — not for description changes
• Remind user to restart opencode after config changes