Pilot Shell 11 updates the engineering harness for current Claude Code and Codex, adds automatic memory across both agents, and retires Pilot Bot. This is a major release: review the changes below before upgrading.
Breaking changes and migration
- Pilot Bot is retired. The
pilot botcommand, background bot runtime, channel tasks, scheduler integration, and five bot skills have been removed. Existing private bot data andJOBS.yamlfiles are preserved, but Pilot no longer runs those jobs. Move any jobs you still need to another scheduler before upgrading. - Memory retrieval is on demand. Pilot no longer injects a history digest at every session start or refreshes generated
CLAUDE.mdmemory sections. Native agent context remains available; agents query Pilot when they need additional project history. Custom integrations that depended on automatic digest injection should use the memory MCP search and retrieval tools. - Manual model switching is the default. New or unset configurations keep model choice with the user. Existing explicit Automated and Off settings retain their meaning, including migration from the older boolean setting.
Automatic memory across Claude Code and Codex
- Capture useful decisions, discoveries, and fixes in the background with either provider. Automatic selection uses available low-cost models and can fall back when authentication, quota, or temporary provider failures prevent capture.
- Preserve pending evidence through retries, cancellation, worker restarts, and invalid responses. Observations and queue acknowledgements commit together, with bounded queues and explicit capture status.
- Accept valid multi-observation responses containing separate XML fences and skip markers, fixing batches that previously remained deferred. Malformed or incomplete responses still leave their evidence pending.
- Search local history and shared knowledge together without requiring a model call or a working vector service. Lexical matches, semantic matches, filters, and pagination share one result path.
- Hide proven local/shared duplicates in lists and search while keeping both historical IDs directly readable.
- Keep automatic shared findings in
.pilot/memories/<author>/<YYYY-MM-DD>.jsonl. Recover unchanged generated Markdown copies into those daily archives before removing the copies; preserve edited knowledge and conflicting records. - Support explicitly maintained Markdown knowledge with source references, revision conflict protection, validation, and local indexing. Backups include personal knowledge, checkout mappings, and pending capture evidence alongside history.
- Keep Team sharing as a single project switch. Shared files travel through the Git work users authorize; local memory remains usable when sharing or provider access is unavailable.
- Filter routine memory-tool feedback out of automatic capture, and show local and shared findings with their sources in the Console.
Workflows, rules, and reviews
- Review and update the rules, skills, and reviewer instructions for GPT-6 Astra, Claude Opus 5, and Claude Fable 5.1. Instructions follow the actual runtime's tools, permissions, and model capabilities.
- Execute clear requests directly, preserve native Plan and Goal choices, and use structured questions and review results where the runtime supports them.
- Connect explicit Pilot planning to Claude's native plan approval lifecycle. Capture the accepted draft after approval, preserve native permission boundaries, and handle the current Claude response format.
- Carry lane identity and durable review handles through phase handoffs and resumed work. Revalidate changed plans and preserve independent review results across calls.
- Use existing behavioral coverage and proportionate verification; remove arbitrary context limits, repetitive instructions, and unsupported tool assumptions.
Hooks and tool output
- Keep diagnostic and routing advice private and nonblocking, with repeated reminders limited per session.
- Prefer Claude's native file-reading tools for inspection and returned background task handles for output collection.
- Preserve Claude's permission decision when RTK rewrites a command. Codex uses its own supported rewrite contract.
- Bound checker subprocesses, run Go diagnostics at package scope, recognize existing integration coverage, and preserve intentional Unicode in source and content.
- Show the reported reasoning effort beside the model in the statusline; omit unknown values.
Installation and upgrades
- Preserve user-owned Codex rules and configuration while refreshing managed assets. Remove retired managed bot assets without deleting unrelated local skills.
- Apply a checksum-verified display patch to supported native Claude Code installations for detailed tool calls, subagent prompts, and thinking summaries. Patch a copy first, retain the original, and leave unsupported binaries unchanged. Patched macOS binaries are ad-hoc signed; the documented restorer can restore an unchanged Pilot-patched binary.
- Reset verbose display once after successful patching, while preserving later user changes and Claude's own updater.
- Keep local
.pilot/state and private working documents out of the public Pilot Shell repository.
Run pilot update to upgrade, then start a fresh Claude Code or Codex session to load the refreshed hooks and instructions. Claude Code and Codex themselves continue to update through their own installers.