Historical version
This page records design, operation, or validation facts from an earlier release. It is not current V7.17.3 guidance. Start from the documentation hub for current instructions.
V7.7.0 Release Notes¶
This release restores controlled apply_patch writes for explicitly enabled projects while keeping ordinary conversations and projects without an enabled policy independent from gate state.
- Add repository-external Operation v2: the first attempt A only creates an origin and is denied; after preparation, a different attempt B must atomically claim permission.
- Add the canonical PreToolUse/PostToolUse adapter. It trusts only
tool_name=apply_patch,tool_input.command, and the realtool_use_id; Edit/Write remain matcher aliases. - Add
--operation-reftocapability-task-prepare/finish/check/cancel. The CLI cannot create host origins or supply dispatch IDs, while legacy session/turn arguments remain compatible with GateTask v1. - Completion requires B's matching PostTool receipt, a real in-scope change, complete reuse decisions, index maintenance, and readable external receipts. Policy changes, cancellation after permission, missing receipts, and stale evidence remain
OUTCOME_UNKNOWN. - Patch bodies are never persisted in Operations, receipts, or observation logs. Command, target, path, and file-read limits are fixed and tested.
- Bind all eleven frozen Codex stable versions to official PreToolUse/PostToolUse schema plus the successful ApplyPatchToolOutput, PostTool payload, and string-response implementations by tag, commit, path, and SHA-256. Hosts missing any evidence cannot install the static enforcing Hook.
Hook permission and the actual file write are not one cross-process transaction. Shell, MCP, and unknown writers are outside this protocol. Workflow evidence neither approves business semantics nor grants commit, push, publication, deployment, or data-write authority.