Skip to main content

Documentation Index

Fetch the complete documentation index at: https://mintlify.com/goulinkh/code-review-harness/llms.txt

Use this file to discover all available pages before exploring further.

When you run crh review, CRH executes a fixed sequence of phases: it prepares a deterministic workspace, initializes an agent session with a custom tool surface, runs the orchestrator which fans work out to parallel sub-agents, aggregates the resulting findings, and finally emits the review to your chosen sink. Understanding this sequence helps you reason about what the agent can and cannot see at each stage, and what to expect when something goes wrong.
1

Workspace preparation

prepareWorkspace performs a single parallel network burst against the provider:
  • Metadata: PR title, body, author, source/target git paths → metadata.json and description.md
  • CI: status checks → ci.json
  • Diff rounds: list of all preview diff rounds → preview-diffs/index.json
  • Diffs and comments: for each round, raw diff, numbered diff, per-file patches + line maps, general comments, and inline comments are written under preview-diffs/<id>/
  • Agent files: AGENTS.md/CLAUDE.md/.cursorrules, .cursor/rules/*.md, .clinerules/*, and .crh/skills/*.md are read from the git object store and materialized into workspace/agent/
A symlink preview-diffs/latest → <id> points to the most recent round. The entire workspace tree is byte-stable: re-running on the same input produces identical files.
The git clone (noCheckout: true, depth: 50) is cached in gitdir. If the head ref is already present in the object store, the clone is skipped.
2

Session initialization

createReviewSession configures the in-memory runtime:
  1. Settings and resource loader: an in-memory SettingsManager and a DefaultResourceLoader are created. Extensions, skills, prompt templates, themes, and context files are all disabled — only the system prompt is injected.
  2. Tool registration: all custom tools are created and wrapped with timeout enforcement (10,000 ms, exempt: delegate_review, submit_review) and response size limiting (80,000 chars, truncated with a pagination hint).
  3. Session creation: createAgentSession is called with the tool list: built-in read, grep, find, ls plus all custom CRH tools. Mutating built-ins (bash, edit, write) are not included.
  4. Changed file list: the session handle exposes changedFiles — the list of paths from the latest diff round’s meta.json files — for use by callers.
The orchestrator’s cwd is set to the workspace directory. Built-in read/grep/find/ls are therefore scoped to the workspace, not the repository source tree.
3

Orchestrator execution

The orchestrator receives the system prompt and the initial user message (e.g. "Review merge proposal. Submit final review with submit_review.").It follows a strict sequence with no multi-turn research loops:
  1. Call mp_metadata and diff_list_files in parallel to read the PR description and the list of changed files.
  2. Optionally call agent_files_list once if agent rules are likely to exist.
  3. Group files into scopes — related files in the same module or feature go together; each sub-agent has ≥ 250k context, so scopes should be large.
  4. Call delegate_review exactly once with the full scopes array.
  5. Receive the merged Findings from all sub-agents.
  6. Call submit_review with the final findings. No further tool calls are permitted after this.
The orchestrator is banned from calling repo_read, repo_grep, repo_ls, and repo_stat directly, from multi-turn research passes, and from sequential per-file delegate_review calls. All source inspection is the responsibility of sub-agents.
4

Sub-agent delegation

delegate_review spawns one sub-agent per entry in the scopes array. All sub-agents run concurrently via Promise.all.Each sub-agent:
  1. Receives its own in-memory session with a focused system prompt: inspect only the assigned scope.
  2. Reads the relevant diffs using diff_get_file or diff_numbered.
  3. Optionally calls repo_read or repo_grep (at most one extra call per file) to confirm or refute a candidate finding from the diff alone.
  4. Calls mark_file_reviewed for each file in scope once the keep/drop decision is final.
  5. Calls report_findings exactly once as its terminal action with structured JSON:
{
  "findings": [
    {
      "severity": "major",
      "path": "src/auth/session.ts",
      "line": 142,
      "comment": "Session token is appended to the URL query string, leaking it in server logs."
    }
  ],
  "summary": "One major finding: token leak in session.ts."
}
Calling report_findings terminates the sub-agent session (terminate: true). If a sub-agent exits without calling it, delegate_review returns empty findings with a warning in the summary.
Sub-agent tool responses are also subject to the 80,000-character limit with truncation hints. Use start/end pagination on diff_numbered and startLine/endLine on repo_read for large files.
5

Finding aggregation

After all sub-agents complete, delegate_review merges their outputs:
  • findings arrays are concatenated into a single flat list
  • summary strings are joined, prefixed with the slot number and scope label
The orchestrator receives this merged Findings object as the tool response. It then calls submit_review with the consolidated findings. The orchestrator’s system prompt instructs it to drop speculative or unconfirmed findings and emit only what sub-agents reported with evidence.
If a sub-agent flagged a blocker finding that appears to require cross-scope confirmation, the orchestrator may issue a second delegate_review call — this is the only permitted follow-up delegation.
6

Sink emission

submit_review calls sink.emit(params, { provider, workspace }) and sets terminate: true, ending the orchestrator session.stdout sink: writes one JSON line to stdout containing the full review payload.Launchpad sink:
  1. Posts inline comments for each finding where line is a valid numbered-diff line number.
  2. Posts a summary comment with a verdict vote:
    • approve → Approve
    • needs-work → Needs Fixing
    • abstain → Abstain
Inline comment keys must be single numbered-diff line numbers. Range keys and diff header line keys are silently dropped by the Launchpad sink.

​
Error handling

Tool timeouts: if a tool call (other than delegate_review or submit_review) takes longer than 10,000 ms, the call is rejected with a timeout error surfaced to the agent. The agent’s session continues and it can retry or move on. Truncated responses: when a tool response exceeds 80,000 characters, the response is cut and a hint is appended: 
[TRUNCATED: N chars omitted — paginate with start/end params or narrow the request]
The agent is expected to paginate using the tool’s start/end or startLine/endLine parameters rather than working from incomplete data. Missing sub-agent report_findings: if a sub-agent exits without calling report_findings (e.g. due to a model error or context exhaustion), delegate_review returns { findings: [], summary: "sub-agent ended without calling report_findings" } for that slot. Other slots are unaffected. Workspace errors: if prepareWorkspace cannot resolve the git head ref, it throws a WorkspaceError. The session never starts and no sink is called.

Build docs developers (and LLMs) love

Get started for freeTalk to us