Current behavior
AgentReviewService.reviewPullRequest() fetches the full unified diff via repositoryClient.getPullRequestDiff() and embeds it (up to MAX_DIFF_CHARS_FOR_CONTEXT = 60,000 chars) into the kickoff user message. For large PRs this overwhelms the AI provider's context window, causing HTTP 400 errors. Since the agentic review already runs multiple tool rounds with read-only exploration tools (cat, rg, ctags-signatures`, etc.), the full diff upfront is unnecessary — the agent can load file content on demand.
Expected behavior
Replace the full-diff kickoff message with a compact file-list summary so the agent starts with a lightweight overview and uses its existing tools to load per-file diffs and source content as needed. This keeps token usage low on the first turn and lets the agent focus its context budget on the files it actually needs to reason about.
Current behavior
AgentReviewService.reviewPullRequest()fetches the full unified diff viarepositoryClient.getPullRequestDiff()and embeds it (up to MAX_DIFF_CHARS_FOR_CONTEXT = 60,000chars) into the kickoff user message. For large PRs this overwhelms the AI provider's context window, causing HTTP 400 errors. Since the agentic review already runs multiple tool rounds with read-only exploration tools (cat,rg,ctags-signatures`, etc.), the full diff upfront is unnecessary — the agent can load file content on demand.Expected behavior
Replace the full-diff kickoff message with a compact file-list summary so the agent starts with a lightweight overview and uses its existing tools to load per-file diffs and source content as needed. This keeps token usage low on the first turn and lets the agent focus its context budget on the files it actually needs to reason about.