An AI coding tool can load too much context, run out of capacity halfway through a task, or leave you wondering what its subagents consumed. Limitcast helps make those things visible across Claude Code, Codex and Grok, with optional controls and reversible context trimming.
This week I focused on the decisions behind those controls: when there is enough evidence to trim, when a forecast is still current, and how an upgrade can preserve the rest of your setup.
Five patches shipped during November 16–20: 0.20.1 through 0.20.5. Here is the full update.
Require complete evidence before trimming
“I didn't find a tool call” and “this tool was unused” are different findings.
Version 0.20.1 fixes activity parsing that could miss calls because of JSON spacing. It also recognizes slash commands in both plain messages and text-block arrays. Cached history carries its completeness information forward, so reading from a cache does not turn incomplete evidence into permission to cut.
When activity history is missing or incomplete, Claude trimming refuses to apply changes. Unsupported nested settings and malformed undo records are rejected before writing.
Codex and Grok MCP proposals also need evidence about the specific server: five distinct completed sessions that observed it as available at both session start and finish. Observing its removal resets that history; re-adding it starts a new evidence window. Duplicate completion events do not count as extra sessions.
Claude transcript ingestion now uses the shared 400-tool catalog bound. Incomplete reads and partial catalogs cannot qualify as complete pruning observations.
A forecast should age when the activity does
A usage percentage that has not changed is still a new observation.
Previously, dropping repeated readings could leave an earlier burst looking current and keep holding new subagents after an idle period. Limitcast now retains the latest observation for each series without filling history with duplicate readings.
Short-window burn rates need recent observations. A predicted runout time already in the past cannot authorize a forecast hold. Fresh activity can still trigger protection when the evidence supports it.
See what the optimizer actually observed
limitcast optimize status now reports verification-check observations by host, including eligible, passed, failed, unsupported and ambiguous checks, alongside active facts.
These counters start with 0.20.1. Older observations are not reconstructed, and excluded command text is not retained. The report makes an evidence gap easier to diagnose; a counter does not prove that a task improved.
One shared hook, broader large-result notices
Versions 0.20.3 and 0.20.4 address the work that happens after a tool returns.
When both the repository optimizer and large-result alerts are installed, one PostToolUse process handles both. That shared process now checks every tool result, including writes, edits, searches and directory listings. A qualifying large result gets a notice once per tool per session, with guidance for narrowing a later call.
Codex and Grok also record observed result sizes. An alerts-only installation retains its existing hook matcher; the broader coverage applies to the shared optimizer process.
The optimizer now locks its own state independently of the settings lock and does not read config.json for these updates. Trimming or malformed settings cannot stall or skip that state update. If the state has not changed, the file is left alone.
Combining the hooks removes a second process start where both previously ran. It does not establish an overall speed or token-savings improvement.
An upgrade should preserve your other hooks
Hook groups can contain commands from several tools. Version 0.20.5 makes install, reinstall, notice folding and uninstall preserve unrelated commands and group metadata instead of replacing the whole group.
If Grok's dedicated Limitcast hook file contains another tool's commands, installation refuses to overwrite it.
Installation also stages host settings and the launcher before committing changes. Invalid settings on a later host leave earlier hosts unchanged. Atomic writes and the existing recovery journal support reconciliation after interruption, while conflicting later edits are preserved.
Concurrent Grok receipts keep their place
Two subagents finishing together could previously read the same pending queue and overwrite one another's additions. Atomic writes kept the file readable but could still lose an update.
Queue changes and receipt settlement now share a lock. Native identity aliases are retained, and repeated pending or completed receipts are ignored. Regression coverage exercises concurrent completion and settlement so distinct receipts survive and repeated events do not count twice.
The small patch and the website work
Version 0.20.2 changes package metadata, with a public maintainer contact and a link to my project profile. It changes no runtime behavior.
Website updates this week added release explanations, clearer disclosure of optional Codex-backed model usage, and author and article structured data. Those are documentation and discoverability changes, not new optimizer capabilities.
The 0.20.0 article also appeared November 16. That package predates this week's five patches. Its foundation includes recoverable trim/apply/undo changes, stronger lock ownership, preservation of explicit configuration choices, and a generic limiter that leaves MCP arguments untouched. It also makes optimizer context overhead and optional model usage visible.
Release checks and the next measurement
The 0.20.5 release validation records 332 passing tests and all eight CI jobs. Its macOS Node 24 line coverage was 95.39%, above the unchanged 95% floor. The preceding 0.20.4 release missed that floor; this update adds regression coverage without lowering it.
Publication now requires a clean checkout and completed, successful push CI for the exact release commit. Missing, pending or failed checks block npm publication. limitcast doctor also requires Node 22+, matching the package.
There is a new optimizer measurement protocol: compare three synthetic tasks with the optimizer off and on, three repeats per condition. It separates task success, actual token categories, optimizer overhead, time and interventions.
That is a protocol, not a completed experiment. These releases establish specific correctness fixes and validation results. They do not establish net token savings, lower subscription costs or improved real-world task success.
Try 0.20.5
For an npm installation, use Node.js 22 or later:
node --version
npm install -g limitcast@0.20.5
limitcast --version
limitcast onboard
onboard provides read-only setup guidance. For an existing installation, limitcast doctor --fix refreshes the hook launcher for installed features. Update any installed host plugins and restart open sessions to load refreshed commands. Review any host hook-approval requests.
The npm release is verified at 0.20.5. The GitHub release notes still list Homebrew at 0.20.4 pending a formula update, so check the installed version if you use Homebrew.
Limitcast remains free and open source under the MIT License. Optional model-backed learning and judging use your authenticated Codex CLI and consume that allowance; those features remain off by default.
If you try it, which part of your AI coding setup is hardest to understand: starting context, usage limits or subagent activity?
Website · Full changelog · GitHub release · Release CI
Disclosure: I used AI assistance to prepare this article from Limitcast's changelog, repository history and release documentation.
Top comments (0)