Command-line options

dakar-review accepts the following options:

• --repo-root <path> selects the Git checkout to review. The default is the current working directory.
• --config <path> selects a CodeRabbit YAML file relative to the repository root.
• --base <ref> selects the base ref for the first review. The default is origin/main.
• --head <ref> selects the head ref to review. The default is HEAD.
• --state-root <path> overrides the review-history root.
• --max-tasks <number> caps the planned review tasks. Under the deterministic-flex-v1 route it composes with --max-luna-calls: the effective finder-pack cap is the smaller of the two, so --max-tasks 1 --max-luna-calls 4 dispatches at most one finder pack. See the review-tuning flags below.
• --max-candidates <number> and --max-findings <number> override the workflow limits described below.
• --synthesis-model <model> and --synthesis-reasoning <level> are accepted for backward compatibility and still appear in the dry-run contract's synthesisModel/synthesisAdapter fields, but under the deterministic-flex-v1 route the audit call always runs on the fixed Terra Flex lane (gpt-5.6-terra, high reasoning); these flags no longer change which model or adapter performs the audit.
• --timeout <seconds> sets the ODW wait timeout. The default is 3600. Operators overriding this should keep it above the review's worstCaseReviewSeconds (2,020 s at default limits; see the dry-run example below), the worst-case wall clock the retry schedule can take.
• --runs-root <path> selects the ODW runs directory used for the run, logs, and result.
• --format <json|markdown> selects the output format. The default is json.
• --odw-bin <path> selects the ODW executable. The default is odw.
• --telemetry streams ODW logs to standard error while preserving the final result on standard output.
• --dry-run returns the workflow contract without launching review agents.
• --help prints command usage, and --version prints the Dakar version.

The following review-tuning flags forward directly to the workflow arguments. The CLI only parses and forwards them; the workflow enforces every bound, so values above a documented maximum clamp to that maximum, while invalid values or values below a minimum use the option's default rather than fail. For budgetGbp, invalid values or values below 0.01 fall back to 0.15, a valid 0.05 remains 0.05, and values above 10 clamp to 10. See "Cost, budget, and the ledger" and "Retries, downgrades, and deferral" below for what each knob controls.

• --budget-gbp <number> sets the hard admission budget in GBP. The default is 0.15, raised from ADR 002's historical 0.1 hard-budget setting because, at pricing table 2026-07-18, the reserve-first audit left less than one finder pack's worst case inside that old benchmark (see the ADR's 2026-08-13 amendment).
• --max-luna-calls <number> caps the Luna Flex finder calls. The default is 4. It composes with --max-tasks: the effective finder-pack cap is the smaller of the two.
• --transaction-max-files <number> sets the maximum files per finder pack. The default is 5.
• --transaction-max-input-tokens <number> and --transaction-max-output-tokens <number> set the per-finder token estimates used for admission. The defaults are 12000 and 2000.
• --terra-max-input-tokens <number> and --terra-max-output-tokens <number> set the audit-call token estimates. The defaults are 48000 and 5000. The output defaults allow for reasoning tokens, which bill as output under the high-reasoning default lanes.
• --adapter-overhead-tokens <number> sets the per-call adapter overhead added to every admission estimate. The default is 13000; raising it towards 28000 better models the cache pi's agentic loop writes on first, uncached calls.
• --max-audit-candidates <number> caps the candidates forwarded to the audit. The default is 30.
• --luna-reasoning <low|medium|high> selects the Luna finder reasoning effort. The default is high (since the 2026-08-13 Flex repricing); medium and low de-escalate to the cheaper adapters.
• --routing-policy <policy> selects the routing policy. The default and sole live value is deterministic-flex-v1.
• --flex-attempts <number> sets the Flex retry attempts per call. The default is 3.
• --per-call-timeout <seconds> sets the per-model-call timeout. The default is 300.

When --config is omitted, Dakar resolves review configuration in this order:

1. Repository-local .coderabbit.yaml.
2. Repository-local .coderabbit.yml.
3. Repository-local coderabbit.yaml.
4. Repository-local coderabbit.yml.
5. User-level $XDG_CONFIG_HOME/dakar/config.yaml, or ~/.config/dakar/config.yaml when XDG_CONFIG_HOME is unset.
6. Dakar's bundled example config.

The user-level config is treated as the current repository's CodeRabbit config only when no repository-local CodeRabbit YAML exists. It is useful for agents that should apply one house review policy across repositories without copying a .coderabbit.yaml into every checkout.

Dakar uses the resolved file at two boundaries. Host code deterministically discovers the path, safely parses the YAML without constructing custom tags, and validates Dakar's supported policy subset before preparation or any ODW, Luna, or Terra dispatch. The normalized serializable policy is then passed to the workflow. Models never parse the YAML file.

The host rejects a missing explicit --config path, malformed YAML, and an invalid shape or value for any supported field. Diagnostics identify both the resolved configuration path and the invalid field. Unsupported keys are reported in the result's ignoredPolicyKeys list and have no routing, gating, budget, retry, or recording effect.

Table: CodeRabbit configuration support levels in the current route.

Executable custom checks use this validated subset:

pre_merge_checks:
  custom_checks:
    - name: Test
      mode: error
      command: make test
    - name: Advisory analysis
      mode: warning
      command: make advisory

Commands run in configuration order from the reviewed repository after range preparation and before ODW starts. A failed mode: error check returns stage: "deterministic-gates" with redacted evidence and launches no Luna or Terra calls; a failed non-error check is retained in SARIF and semantic review continues. Commands must not embed credentials. Output is bounded, common credential forms and sensitive environment values are redacted, and full streams are represented by SHA-256 digests.

For repository-local configuration, executable commands are loaded from the prepared trusted base commit, not the head under review. A pull request cannot grant itself host command execution by adding or changing a custom check. User-level and bundled configurations are outside the reviewed repository and are read from their resolved operator-controlled paths.

Each reviews.path_instructions entry requires non-empty string path and instructions fields. Dakar matches the glob against normalized changed paths and includes an instruction only in finder packs containing a matching path. General language, tone, profile, and natural-language custom-check guidance is included without exposing executable commands or ignored keys. A root AGENTS.md is loaded through its own dedicated path (described below), independently of unsupported knowledge_base.code_guidelines patterns.

ODW normally runs agents in copied workspaces. Those copies may not contain the repository's .git directory, so live review runs should pass repoRoot as an absolute path. Finder and verifier prompts use git -C <repoRoot> for diff evidence, and the prepare step passes the same path to the state helper.

If the reviewed repository has a root AGENTS.md, dakar-review passes it to the workflow as repository-local review context. Workflow schema rules, machine-readable output requirements, and Dakar safety rules still take precedence over repository instructions.

When the operator's mcp CLI is on PATH, finder prompts describe two optional context tools. The CLI attempts CodeGraph warmup before dispatch only when the reviewed head is checked out and the worktree is clean; otherwise it skips warmup with a warning on standard error. Warmup calls share an overall 30-second deadline. When eligible, the CLI calls codegraph_index_directory on the reviewed checkout, then codegraph_index_markdown on the root AGENTS.md, README.md, and any markdown files in the review's changed set (bounded). Warmup is advisory — a missing mcp command or a failed call warns on standard error and never blocks the review — and can be disabled by setting DAKAR_SKIP_CONTEXT_WARMUP in the environment. Finders are directed at the CodeGraph query tools (context, callers, impact, symbol, and documentation search) in preference to broad file reads.

The CLI resolves the repository's GitHub identity from its origin remote and constructs optional context guidance at the host-side adapter boundary. Only this translated guidance crosses into the workflow; the repository identity and MCP command details do not. If no usable GitHub origin is available, the guidance says that DeepWiki is unavailable. Otherwise it describes the DeepWiki lookups (ask_question, read_wiki_structure, read_wiki_contents) with an explicit caveat: DeepWiki is not realtime. It helps a finder understand dependencies and the overall purpose of the codebase, but it may not incorporate changes made over the past week, so it must never be cited as evidence about the head under review.

For a syntax and contract check that does not call review agents, run either:

dakar-review --dry-run --repo-root "$PWD"
npm run odw:dry-run

Dry-run output includes the workflow version, default finder model set, synthesis model and adapter, task kinds, limits, default task graph, and JSON Schemas used for candidate, verifier, and synthesis handoffs.

By default, dakar-review stays quiet until the workflow finishes so standard output is easy to parse. To watch ODW progress while keeping the final result machine-readable, pass --telemetry:

dakar-review --repo-root "$PWD" --base origin/main --telemetry

Telemetry follows odw logs <run-id> --follow and writes the live ODW event stream to standard error. The final JSON or Markdown result still goes to standard output, so callers can redirect the channels independently:

dakar-review --repo-root "$PWD" --telemetry > review.json