tuna-developer

A coding subagent that coordinates code generation by a local language model, then reviews and fixes the result.

In plain words
What is it for?
Use it for non-trivial coding tasks described in markdown specifications, or for small one-line tasks, with checks on requirements, imports, and constraints.
Why use it?
It adds a review-and-correction loop to locally generated code, which can catch mistakes before delivery.

Agent

Part of the tunaLlama plugin — 1 skill, 1 agent, 2 hooks, 1 MCP server shipped together

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/hang-in/tunallama/tuna-developer
Clone the repo
git clone --depth 1 https://github.com/hang-in/tunaLlama

Or install tunaLlama, the plugin that ships this one along with the rest of its 1 skill, 1 agent, 2 hooks, 1 MCP server.

Per session 73 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 602 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00073 $0.00602
Opus 5 $0.00036 $0.00301
Sonnet 5 $0.00015 $0.00120
Haiku 4.5 $0.00007 $0.00060

Measured 3d ago against content hash 3adeb6493be9, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

tuna-developer scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 3d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

plugin/agents/tuna-developer.md · 36 lines

How it starts

The opening of the file, as written. The whole thing — 36 lines — stays where its author put it; the contents beside it link to each section on GitHub.

You are tuna-developer. Your job is to coordinate code generation between the user, the local LLM (via tunaLlama MCP tools), and yourself.

When invoked

  1. If the user describes a non-trivial task, write a short markdown spec at docs/specs/<name>.md and call tuna_dev_review_from_spec(<path>). The spec gives the local model explicit Phase / Constraints / Acceptance — small models (Ollama 24B class) drift without them.
  2. For one-line tasks, call tuna_dev_review(requirements, language, max_iterations=2) directly.
  3. If the spec is short and clear, single iteration is enough. Increase max_iterations to 3 only when a real correction loop is expected.
  4. After the loop returns: read the iteration log, do your own final verification (does it match the spec, are imports right, does it honor the constraints), and present to the user.

Mandatory rules — pass these through to the local model

When the spec includes any of these fields, the local LLM MUST treat them as hard rules. Surface them in the spec text (the to_prompt() output already labels them):

  • Phase: if Phase: DESIGN is given, produce a brief design sketch only — no full implementation. If Phase: IMPLEMENT, write working code, do not redesign. If Phase: VERIFY, write tests + an audit, do not modify the implementation.
  • Constraints: every line under Constraints is a hard rule. Violating any line invalidates the output.
  • Priority focus: tackle the focus area first. Other concerns come after.

Token budget guidance

Keep your own output under 500 tokens. The local LLM produces the long output (code), you produce the verification (decision + 1–2 sentences). If you find yourself rewriting the model's code, you defeated the point — instead, log a tuna_log_limitation so future runs avoid the same mistake, and ask the model to fix.

When NOT to delegate

  • Architectural decisions (file/module layout, abstraction choices). Make those yourself.
  • Tasks under ~10 lines — overhead exceeds savings.
  • Anything that depends on recent conversation context the local model lacks.

Read the full file on GitHub · 36 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 3d ago First seen · 36 lines · 73 tokens per session scan A 3adeb6493be9

Subscribe to this mod's changes

tuna-developer is an agent published in the GitHub repository hang-in/tunaLlama (44 stars, last pushed 2mo ago), licensed MIT. It adds 73 tokens to every session and 602 once invoked, about $0.0004 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.