v2.10.0

Worker convenience release: worker init can install and start the background service for you, agent reasoning effort is configurable across harnesses, every prepared worktree gets its dependencies installed, and the Sentry watcher stays quiet until it picks something up.

Added

• Guided worker operating policy: worker init now asks when new tasks may be picked up, whether failing CI should be repaired automatically, and whether pull-request conflicts should be handled immediately, on a schedule, or manually. Existing values become the prompt defaults when the wizard is rerun
• Guided Sentry setup for workers: worker init can now validate and add an optional repo-bound Sentry auto-fix project, while worker connect sentry adds projects to existing workspaces. Both store the API token in a source-specific owner-only env file and avoid persisting an enabled monitor when validation fails. Sentry remains an addition to the normal tracker/query setup
• One-step worker service install (DEV-115): worker init now offers to install and start the background service automatically — a user-level systemd unit on Linux or a launchd agent on macOS — so setup finishes with the worker already running and auto-restarting. The unit uses the user manager’s default.target, records the current executable PATH, invokes Bun directly, and enables user lingering so it starts after reboot without waiting for login; macOS health is only reported when launchd shows a running process. Declining keeps it manual (definition plus exact install commands written into the workspace home), --no-service skips service setup entirely, re-running detects an existing DevIntern-managed service and offers an in-place update, and a hand-written definition is never overwritten. Any failed step restores the previous state and prints the manual install commands so nothing is left half-installed
• Configurable reasoning effort (DEV-116): AGENT_EFFORT in .devintern-code/.env (low | medium | high) tunes how deeply the agent reasons per run, mirroring AGENT_MODEL and applying to every agent spawn. Harnesses with a documented per-run mechanism emit it (Claude Code, Grok, Antigravity, and Reasonix via --effort; Cline via --thinking; OpenCode and Kilo Code via --variant; Codex via its model_reasoning_effort config override; pi composes it into the model string), harnesses without a headless mechanism (Cursor, Goose, Qwen, Kimi) ignore it with a one-line warning, and invalid values fail with a clear error. With the setting unset, no extra flags are emitted and behavior is unchanged

Changed

• Safer worker workspace bootstrap: worker setup and repository imports now enforce owner-only permissions on the shared credential file and repair existing Sentry credential-file permissions when setup is rerun. Workspace repositories follow origin/HEAD unless an explicit per-repository override is configured; the fleet-wide [defaults].default_branch setting has been removed
• Every prepared worktree gets its dependencies installed (DEV-119): dependency installation no longer happens only for review worktrees — task, automation, and conflict-resolution worktrees are prepared with the package manager auto-detected from the lockfile (bun/pnpm/yarn/npm, uv/poetry/pip, bundle, go, cargo, composer, maven, gradle) after the final branch checkout, reusing the worker’s environment and registry auth. The base worktree installs once at creation and reuses skip reinstalling; a missing lockfile or failed install continues the run with a warning so the agent can still set dependencies up itself
• Quieter Sentry poller logging (DEV-118): the Sentry error watcher stays quiet while issues sit below min_occurrences — skipped issues are only explained with --verbose output, and a single line (project, issue id, occurrence count) is logged when an issue is actually picked up