Harbourmaster configuration

Harbourmaster uses TypeScript model defaults, bundled OpenCode configuration, and environment overrides. The GitHub Action maps its inputs to environment variables before it starts the CLI. Harbourmaster does not read repository YAML configuration.

Precedence

Model routes use this order, from highest to lowest precedence:

1. The environment override for the selected tier.
2. The general HARBOURMASTER_BEDROCK_MODEL override.
3. The TypeScript defaults.

Model defaults

Use a provider/model-id value for each model. Change per-agent rules, fallback lists, provider availability, and optional cost metadata in the TypeScript model configuration when necessary.

Fast specialist routes use GPT-5.6 Luna with medium reasoning effort, while deep specialist routes use Luna with extra-high (xhigh) reasoning effort. The editor, fallback, and adversarial verifier use GPT-5.6 Terra with medium reasoning effort by default. The code-quality route uses GPT-5.6 Sol with low reasoning effort, also called Sol Light. A tier override can select another configured model when needed.

The default fast, deep, and editor fallback is the bundled GPT-5.6 Terra medium model. The fallback tier has no additional fallback.

Model environment variables

Empty values are ignored.

Bundled OpenCode configuration

The shared action always sets OPENCODE_CONFIG_PATH to Harbourmaster's bundled .opencode directory. Consumer-owned .opencode content is review context and is not executed.

The bundled runtime defines the editor, adversarial verifier, and specialist agents. For each attempt Harbourmaster creates an isolated temporary diff directory with reviewable patches. Before OpenCode starts, the trusted runner writes ./review-index.json, ./shared-pr-context.txt, and ./review-discussion.txt. It also copies the root reference paths AGENTS.md, .agents/, and docs/ when the checkout contains them. OpenCode starts in this directory with the trusted bundled configuration. Specialist and verifier prompts refer to the three review files, ./patches/, and the staged reference paths.

The review workspace files have these roles:

The editor receives the publication and lifecycle context inline. It invokes spawn_reviewers directly after it reads the prior discussion. It cannot write or copy the review workspace files.

Specialists and the verifier can use only read, glob, grep, list, and read_more. Each result page contains at most 200 lines from read output, or 30 lines from glob, grep, and list output. Each page, including its continuation notice, has a 48 KiB limit. The read_more tool accepts a cursor from the previous page in the same session. A cursor works once. The runtime retains at most 128 results, each with at most 64 KiB of UTF-8 text. It discards the oldest cursor when it reaches this count. A notice identifies results that exceed the retention limit. The cursor can return only the retained prefix. If OpenCode truncates a result first, the cursor cannot recover the omitted text. Use supported file offsets or a narrower query.

The code-quality reviewer also receives the absolute canonical checkout root in its prompt. A per-run permission override grants this reviewer external_directory access to that checkout root. The reviewer can inspect the checkout alongside the staged diff. All other external-directory access remains denied, so other specialists and the verifier can inspect only the diff directory. A trusted hook canonicalizes paths for the repository tools and rejects symlink escapes. It also rejects .git paths and symlink aliases to Git metadata. The same guard applies when the runner copies reference files.

Before the editor's first model call, a runtime preflight verifies that OpenCode registered spawn_reviewers and publish_review and that the editor configuration allows both. Missing access fails the run before specialists are started.

The default provider is Amazon Bedrock in us-east-1. The Terra alias resolves to openai.gpt-5.6-terra with medium reasoning effort. The Luna alias resolves to openai.gpt-5.6-luna with medium or extra-high (xhigh) reasoning effort. The code-quality reviewer uses the Sol alias. It resolves to openai.gpt-5.6-sol with low reasoning effort. These models use the bedrock-mantle endpoint. The GitHub Actions IAM role must allow these Bedrock Mantle calls: bedrock-mantle:CreateInference, bedrock-mantle:GetProject, and bedrock-mantle:ListProjects.

Runtime environment

GitHub-hosted action runs normally use aws_role_to_assume and GitHub OIDC instead of long-lived AWS credentials.

Review-thread loading, specialist selection, and publication lifecycle rules are in the review behavior reference.