[Raezil]

CodeMode repository performance improvements

PyLattice now reuses source context and the Monty runtime between coding steps, and can refactor/create multiple files with fewer policy-checked tool calls.

Runtime & context

• Warm sandbox: one Monty worker pool per agent run, isolated execution sessions per program; always closed after the task.
• context_collect: search or explicitly select up to 24 files in one bounded read-only call, with relevance-centered code excerpts, line offsets, SHA-256 revisions and index_hits/disk_reads diagnostics.
• Workspace search index: bounded 8 MiB metadata-validated in-memory index avoids re-reading unchanged files on repeated queries.
• Context Manager integration: bounded LRU of 64 verified full-source snapshots / 4 MiB, automatically selects context relevant to subsequent CodeMode steps, detects staleness and redacts model credentials.
• Security: raw repository snippets are included as explicitly untrusted user-level reference data, never promoted to system instructions.

Repository edits

• apply_patchset: one approval for up to 24 edits and creations, with schema validation, exact unique replacements, optional expected SHA-256, per-file diffs and sandbox policy.
• Preflight all mutations before staging; no-clobber hard links for new files; source invalidation and best-effort rollback on partial errors. Not a cross-file atomic transaction.
• Read-only policy and approval hooks remain enforced through the existing ToolRegistry/ToolPolicy boundary.

Verification

• 294 tests passed; Ruff lint and Python compilation passed: successful CI.
• A reproducible synthetic benchmark checks 240 files / 2.09 MB on the GitHub Actions runner:
  • Cold scan: 47.165 ms, 240 disk reads.
  • Warm median: 25.819 ms, 0 disk reads, 240 index hits.
  • Warm search approximately 1.83× faster under those conditions.
• This measurement is only filesystem context collection on CI, not end-to-end model/tool routing or overall refactoring speed. No unsupported total speedup claims.

Docs

README includes ready-to-run CodeMode examples for context collection and a multi-file patchset, plus the benchmark command.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
py-lattice	Ready	Preview	Oct 9, 2026 8:04pm UTC
pylattice	Canceled		Oct 9, 2026 8:04pm UTC

[vercel]

Deployment failed for project py-lattice with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/protocol-lattice?upgradeToPro=build-rate-limit

[vercel]

Deployment failed for project pylattice with the following error:

Resource is limited - try again in 24 hours (more than 100, code: "api-deployments-free-per-day").

Learn More: https://vercel.com/protocol-lattice?upgradeToPro=build-rate-limit