Skip to content

installation.md

Latest commit

 

History

History
172 lines (117 loc) · 6.46 KB

installation.md

File metadata and controls

172 lines (117 loc) · 6.46 KB
Raw

Installation

Tribunal ships as a single Go binary plus a set of markdown skills and agent definitions. Installation has three layers:

  1. The CLI (tribunal) — installed via go install.
  2. The skills and agents — installed into your editor / harness host via tribunal init.
  3. Per-project state — created in .tribunal/ in the project's working directory.

Prerequisites

Component Required Notes
Go 1.23+ yes For installing the CLI.
Git yes Tribunal records review reports as commits.
A supported harness yes One of: Claude Code, OpenCode, or Cursor.
Anthropic API key optional (v0.2+) Needed for multi-model adversarial review via Anthropic.
OpenAI / Google API keys optional (v0.2+) Needed for multi-model dispatch beyond Claude.
LM Studio optional (v0.2+) Local-model adversary floor.
Burnt XION testnet account optional (v0.3+) On-chain settlement.

Install the CLI

go install github.com/dpdanpittman/tribunal/cmd/tribunal@latest

Verify:

tribunal --version

Initialize a host

Pick your harness:

Claude Code

tribunal init --target claude-code

This copies:

  • All skills/tribunal-* to ~/.claude/skills/
  • All agents/tribunal-*.md to ~/.claude/agents/

After init, the skills are discoverable in any Claude Code session.

OpenCode

tribunal init --target opencode

This generates a plugin entry in your opencode.json plugin list and copies the skills/agents into the resolved plugin directory.

Cursor

tribunal init --target cursor

Tribunal will install at ~/.cursor/plugins/local/tribunal/ and add the appropriate rules. Reload the Cursor window after install.

Auto-detect

If you don't specify --target, Tribunal will detect from the current environment:

  • If CLAUDE_PROJECT_DIR is set or ~/.claude/ exists → Claude Code
  • Else if ~/.opencode/ or an opencode.json is nearby → OpenCode
  • Else if ~/.cursor/ exists → Cursor

Register your agents

Each Tribunal agent (each model + role combination) needs an ed25519 keypair.

tribunal agents add claude-adversary --model claude-opus-4-7 --role adversary
tribunal agents add claude-reviewer-arch --model claude-opus-4-7 --role reviewer-arch
tribunal agents add claude-reviewer-sec --model claude-opus-4-7 --role reviewer-sec
tribunal agents add claude-reviewer-perf --model claude-opus-4-7 --role reviewer-perf
tribunal agents add claude-pm --model claude-opus-4-7 --role project-manager
tribunal agents add claude-implementer --model claude-opus-4-7 --role implementer
tribunal agents add claude-classifier --model claude-opus-4-7 --role classifier
tribunal agents add claude-qa --model claude-opus-4-7 --role qa

This generates ~/.tribunal/agents/<name>.key (private, 0600) and ~/.tribunal/agents/<name>.pub (public + metadata) for each.

For multi-model dispatch later (v0.2+), add agents from other model families:

tribunal agents add gpt-adversary --model gpt-5 --role adversary
tribunal agents add gemini-adversary --model gemini-2.5-pro --role adversary
tribunal agents add local-qwen-adversary --model qwen-3-32b --role adversary

Per-project setup

In any project where you want Tribunal to operate:

cd <your-project>
git init   # if not already
echo ".tribunal/" >> .gitignore  # actually, keep tracked — see below

Tribunal creates .tribunal/ containing:

  • status.json — plan registry, current state machine state, residual findings.
  • ledger.jsonl — append-only signed findings/resolutions.
  • findings/F-*.md — full text of each finding.
  • resolutions/F-*.md — full text of each resolution.
  • reports/<plan-id>/ — QC reports per plan, per reviewer.
  • plans/<plan-id>/ — intent doc, plan, tasks, status per plan.

Decision: track .tribunal/ in git, or not?

Recommended: yes, track it. The review history is valuable context for future reviewers and a public audit trail. The local ledger is signed and append-only, so collaborators can't quietly rewrite it. The chain layer (v0.3+) ultimately settles disputes anyway.

If you have specific reasons to keep it untracked (private bug content, etc.), add .tribunal/ to .gitignore and use tribunal export (planned for v0.4) to generate a shareable summary.

Verify install

cd <your-project>
tribunal agents list
tribunal ledger summary

If tribunal agents list shows your registered agents and ledger summary reports a fresh empty state, you're ready.

Optional: API keys for multi-model dispatch (v0.2+)

For multi-model adversarial review, set the API keys for each provider you want to use:

export ANTHROPIC_API_KEY=...
export OPENAI_API_KEY=...
export GOOGLE_API_KEY=...
export LM_STUDIO_URL=http://localhost:1234    # default; only set if different

Add them to your shell rc file or use a secret manager appropriate to your environment.

Optional: Burnt XION chain config (v0.3+)

For on-chain settlement, configure ~/.tribunal/chain.toml:

[chain]
rpc = "https://rpc.xion-testnet-2.burnt.com:443"
chain_id = "xion-testnet-2"
contract_address = "xion1..."

[operator]
mnemonic_path = "~/.tribunal/operator.mnemonic"  # 0600
gas_price = "0.025uxion"

The operator account is the local human-or-org account that signs the on-chain transactions. It's separate from the per-agent ed25519 keys (those sign findings; the operator pays gas).

Troubleshooting

tribunal init says it can't detect a host. Pass --target explicitly.

tribunal agents add fails with label already taken. Each agent label must be unique in your local registry. Pick a different label.

tribunal ledger summary shows nothing after several reviews. Verify the agent signatures are valid: tribunal ledger verify. If verification fails, the ledger is corrupted or an agent's key was replaced without rotation — see the agent rotation guide.

Multi-model dispatch is silently slow. Verify all required API keys are exported and reachable: tribunal dispatch test --providers claude,openai,google,local.