Skip to content
Agents tracked: 483 Releases (7d): 766 Agent status: 1 with issuesSince last limit reset: Claude 42d · Codex 3d Updated Nov 3, 2026
AgentGid
✦ Ask

Loki Mode changelog: what's new each month

Every stable Loki Mode release summarised by month: the highlights, new features, improvements, fixes and anything you need to act on. 2 months covered; the current month updates daily.

November 2026 · so far

24 releases: 11.101.0 → 11.120.0

November brought major language and platform support expansions, including full .NET coverage and improved Java testing, alongside significant performance improvements and a streamlined codebase. The month also introduced learning from past runs, refined verify logic, and made provider and CLI setup more flexible.

Action needed
  • Factory-operations and release-cadence skill modules removed; projects should not reference these.
  • Old-loop-only flags --openspec, --bmad-project and --issue are refused; use loki start owner/repo#N instead.
  • Lessons and worked examples now opt-in: set LOKIPRLESSONS=1 and LOKIWORKEDEXAMPLE=1 to enable; default runs are unchanged.
  • Plans pinned to a model (e.g. Haiku) now apply that model to all sessions including conflict referee.

Highlights

  • Added .NET support with detection of .csproj, .sln and global.json files, running scoped dotnet test and writing acceptance tests for xUnit, NUnit and MSTest.
  • Improved Java testing with test file mapping for Maven and Gradle, detection of changed classes by name and reference, and automatic use of repo-local ./mvnw or ./gradlew.
  • Removed the separate Bun and SDK run loops, consolidating to one default engine while keeping legacy support for large PRDs.
  • Loki now records lessons from each run to brief future work on the same codebase, capturing failing checks, test edits and spec conflicts.
  • Made provider setup optional: runs now work with an API key alone using the bundled Claude SDK, and auto-detect the first supported CLI if Claude CLI is not installed.

New

  • .NET language support with xUnit, NUnit and MSTest test detection and acceptance test generation.
  • Issue slicing: large issues with multiple deliverables or acceptance criteria automatically split into smaller runnable slices.
  • Lesson system that records failures and fixes from each run and applies relevant lessons to future tasks on the same files.
  • Worked examples: tasks similar to past verified runs receive the previous run as a short example in their brief.
  • loki quickstart command that builds a complete demo repo with git identity and sealed receipt.
  • Safety check acceleration: shell command safety now decided in-process in 0.2 ms instead of launching a bash script.
  • Per-stage diff summaries and context recall metrics showing how much of changed code was covered by the brief.
  • Test-pinning: tasks that already have a relevant test skip Loki's pre-written acceptance tests and go straight to implementation.

Improved

  • Verify no longer has a fixed 120 second time limit; it reserves time after checking Wall and task-named tests.
  • Java test selection now includes test file mapping, so changed classes select their tests by name and reference instead of running all tests.
  • Gradle and Maven test results counted reliably across all versions; Gradle results via init script without modifying repo build files.
  • Authentication and credit failures now show the provider's own message and the fix command in the run Reason line.
  • Test file handling: existing tests become read-only unless the task names them, and edits outside the task scope are restored after the session.
  • Fix loop logic: a fix round leaving the same failing tests ends the loop instead of starting another round.
  • Wall test runners chosen from the package that owns the tests, so nested fixture packages no longer decide the root package runner.
  • Model selection: when no provider named and Claude CLI not installed, Loki uses the first supported CLI on PATH and says which one it picked.

Fixed

  • Per-stage diff summaries were empty on git 2.55 due to a pathspec issue with ignored paths.
  • Split issues no longer falsely detect HTTP 429 rate limits from bare "429" strings in hashes, run IDs or cost figures.
  • Docker image now includes quality-gate scripts so test runners can execute them.
  • Competitor survey and roadmap sections removed from shipped reference docs; release checks now fail if planning or competitor docs are committed.
  • npm package reduced from 913 to 789 files by removing old test suites, Python test folders and internal planning docs.
  • loki bench now says plainly that the benchmark harness was removed instead of pointing to a missing script.

October 2026

56 releases: 11.68.0 → 11.100.0

October brought major improvements to Loki's verification, resume capabilities, and test handling across languages. The month focused on stronger proof verification, live monitoring, acceptance criteria extraction, and better support for Java, Go, and Rust projects.

Action needed
  • Removed loki trigger install, loki trigger uninstall and loki trigger daemon --loop.
  • loki setup is now optional on machines already logged in; claude check follows the same condition as preflight.

Highlights

  • Runs can now be resumed mid-way with loki start, keeping finished stages and running only the rest; Ctrl-C or SIGTERM stops and offers resume on the next start.
  • Live monitoring with loki status --follow shows one line per stage, test and cost event until completion, with alerts for stuck runs and budget milestones.
  • Acceptance criteria are now extracted from task specifications and validated in tests; Wall tests are written separately and can be arbitrated when they conflict with the task.
  • Java projects now get independent acceptance tests and verify runs instead of being skipped; Go and Rust can finish VERIFIED with better test result parsing.
  • Receipt seals now carry measured accuracy data showing wrong-pass rates from seeded-defect testing, with 95% confidence bounds for each language and test runner.

New

  • loki try runs on public repository URLs in temporary clones, keeping only the receipt.
  • loki github protect requires the loki/receipt status check on the default branch.
  • loki next suggests three small tasks from issues, TODO comments and failing tests with cost estimates.
  • loki cost tracks spending by month with --csv export showing repo, user, USD, verdict and run ID.
  • loki notify desktop sends native desktop notifications on macOS and Linux.
  • loki answer resumes a BLOCKED run with a text answer, keeping prior stages.
  • loki memory search now available with safeguards against leaking private keys and rejecting hardlinked files.
  • Earned auto-merge with LOKIEARNEDMERGE enforces minimum verified receipts per path before merging.

Improved

  • Wall test authoring now happens in a separate walled provider call before the main loop, with schema-checked JSON test cases on Python when LOKIE10WALLCASES=1.
  • Test verification now checks imported modules affected by changes through relative paths and tsconfig/jsconfig settings, catching downstream breakage.
  • Engine turn limits derived from time budgets prevent infinite loops; resumed implement sessions no longer start second paid sessions.
  • Pull request bodies now fill the target repository's own PR template with verified plan details, with pre-ticked boxes reset.
  • Multi-part tasks route to group runner only for github.com owner/repo origins; other sources run as single local runs.
  • Run cost is now unified across all surfaces: receipt, console, status, dashboard, Slack and handoff, including child sessions.
  • Project model detection now reads Cursor, Copilot, Windsurf and Gemini rule files alongside AGENTS.md and CLAUDE.md.
  • Intake recognizes runtime evidence like stack frames, request-response pairs and tracker URLs in task text, requiring Wall tests that reproduce them.

Fixed

  • Deleted sealed acceptance-check copies now report as tampered instead of passing verification.
  • Air-gap mode now skips npm update checks on both loki start and loki version.
  • Ablation runs now restore the caller's SIGINT handling instead of resetting to default.
  • Go test results with regex characters in -run arguments are now correctly counted as passes.
  • A run that crashed after signing its checkpoint now resumes on events up to that checkpoint.
  • Wall tests created in subdirectories are now collected and sealed instead of skipped.

Summaries are written automatically from the official release notes (full changelog ↗); check the original notes before relying on a detail. Loki Mode: pricing, features and alternatives · All changelogs