Minimizing API usage when running Archify with CLI agents

[milos-plavsic]

Minimizing API usage when running Archify with CLI agents

If you feed your entire repository into a high-end model just to generate an Archify diagram, you are paying for more than you need. The efficient approach is to split the work across multiple models, assigning each stage the minimal reasoning depth required.

Below are single commands for Claude Code and Codex CLI that preserve the cost savings by using cheap models for the grunt work and the high-end model only for the final synthesis step.

---

Reasoning depth scale used in this guide

• Low: pattern matching, deterministic extraction. No semantic understanding.
• Medium: syntactic comprehension. Understands imports, exports, signatures, but not overall intent.
• High: abstract synthesis. Groups concepts, detects cycles, infers trust boundaries.
• Max: overkill for Archify. Only useful for esoteric metaprogramming.
• Ultra: complete waste of compute for this use case.

---

The pipeline (same structure for both agents)

The pipeline runs four stages:

1. Discovery: read manifests, list directories (Low depth, cheap model)
2. Dependency extraction: parse imports, exports, decorators (Medium depth, cheap model)
3. Synthesis: group components, detect cycles, define boundaries (High depth, premium model)
4. Compilation: validate and render the HTML (local CLI, no LLM cost)

Stages 1-3 are chained into a single command. Stage 4 runs automatically at the end.

---

One command for Claude Code

This uses Haiku 4.5 for discovery and dependency extraction, then Sonnet 4.6 (default) for the final synthesis. Sonnet delivers Opus-level quality for 60% less cost.

claude --model haiku "List all root directories. Read every package.json, go.mod, Cargo.toml, requirements.txt, pyproject.toml, and .env.example. Extract: 1. The main entry point file paths. 2. All external dependency names. 3. The folder structure tree (max depth 4). Output as plain bullet points. No commentary." > discovery.txt && claude --model haiku "Using the folder structure from discovery.txt, read ALL source files (.ts, .py, .go, .js, .rs, .java). For each file, extract: - Absolute import paths. - Exported classes, functions, or structs. - Decorators, middleware, or route annotations. Do NOT summarize what the code does. Just output the raw import/export map grouped by file path." > deps.txt && claude "You are an architecture expert. Using the attached data below, generate the strict Archify JSON IR. Rules: - Group files into exactly 8-12 runtime components based on their imports/exports. - Identify the single primary request/event path through the system. - Mark any external service as an 'external' node with a clearly defined trust boundary. - Put all supporting details into descriptive 'cards' (node metadata), NOT as extra edges. - Output ONLY valid Archify JSON IR. Do not add markdown formatting. Discovery data: $(cat discovery.txt) Dependency map: $(cat deps.txt)" > diagram.json && node archify/bin/archify.mjs validate diagram.json && node archify/bin/archify.mjs render diagram.json output.html

---

One command for Codex CLI

This uses gpt-5.4-mini for discovery and dependency extraction, then gpt-5.6-terra for the final synthesis. Terra is the latest workhorse optimized for codebase reasoning.

codex --model gpt-5.4-mini "List all root directories. Read every package.json, go.mod, Cargo.toml, requirements.txt, pyproject.toml, and .env.example. Extract: 1. The main entry point file paths. 2. All external dependency names. 3. The folder structure tree (max depth 4). Output as plain bullet points. No commentary." > discovery.txt && codex --model gpt-5.4-mini "Using the folder structure from discovery.txt, read ALL source files (.ts, .py, .go, .js, .rs, .java). For each file, extract: - Absolute import paths. - Exported classes, functions, or structs. - Decorators, middleware, or route annotations. Do NOT summarize what the code does. Just output the raw import/export map grouped by file path." > deps.txt && codex --model gpt-5.6-terra "You are an architecture expert. Using the attached data below, generate the strict Archify JSON IR. Rules: - Group files into exactly 8-12 runtime components based on their imports/exports. - Identify the single primary request/event path through the system. - Mark any external service as an 'external' node with a clearly defined trust boundary. - Put all supporting details into descriptive 'cards' (node metadata), NOT as extra edges. - Output ONLY valid Archify JSON IR. Do not add markdown formatting. Discovery data: $(cat discovery.txt) Dependency map: $(cat deps.txt)" > diagram.json && node archify/bin/archify.mjs validate diagram.json && node archify/bin/archify.mjs render diagram.json output.html

---

If summary files exceed the shell argument limit

Use heredoc for the synthesis step. Replace the third segment in the chain with:

codex --model gpt-5.6-terra <<EOF > diagram.json
You are an architecture expert. Generate the Archify JSON IR. Output ONLY valid JSON.

Discovery:
$(cat discovery.txt)

Dependencies:
$(cat deps.txt)
EOF

For Claude Code:

claude <<EOF > diagram.json
You are an architecture expert. Generate the Archify JSON IR. Output ONLY valid JSON.

Discovery:
$(cat discovery.txt)

Dependencies:
$(cat deps.txt)
EOF

---

Stage summary

Stage	Task	Claude Code	Codex CLI	Depth
1	Discovery (manifests, directory tree)	Haiku 4.5	gpt-5.4-mini	Low
2	Dependency extraction (imports/exports)	Haiku 4.5	gpt-5.4-mini	Medium
3	Architecture synthesis and JSON IR	Sonnet 4.6 (default)	gpt-5.6-terra	High
4	Validation and rendering	local CLI	local CLI	None

---

Why this reduces usage

The Low and Medium models handle roughly 80-90% of the raw token volume (reading all source files). The High model only sees the compressed summaries – about 15,000 tokens instead of 500,000 tokens of raw code. You run the High model once, not multiple times due to retries. Models above High (Max, Ultra, Opus, Fable) add nothing for this task and are explicitly omitted.

For Claude Code, Sonnet 4.6 with default high effort is the proven sweet spot. For Codex CLI, gpt-5.6-terra provides the same level of reasoning at the right price. This gives the same diagram quality at a fraction of the cost.