agent-sh/agnix v0.52.0 on GitHub

pre-commit hooks install agnix themselves
(#1439). The three hooks in
.pre-commit-hooks.yaml move from language: system to language: python,
so pre-commit and prek create a virtualenv and pull the agnix wheel that
matches the pinned rev instead of requiring a manual cargo install on
PATH. This is what the new PyPI distribution makes possible. A root
pyproject.toml exists solely for that install: pre-commit clones the repo
and pip install .s it, so the clone has to be pip-installable, and this shim
declares the matching agnix==<version> as its only dependency.
scripts/sync-versions.sh moves the shim's version and its pin together, so a
hook pinned at rev: vX.Y.Z always runs agnix X.Y.Z. Platforms without a
wheel can still pin language: system in their own config, which the
configuration guide now documents.

CC-SET-028: invalid feedbackDrafts setting
(#1435). Claude Code v2.1.247
added the SendFeedback tool and the feedbackDrafts setting that governs it.
The setting is a string enum - notify, quiet, or off - so the natural
guess of a boolean silently leaves the default notify in place. Flags any
value outside the documented enum.
CC-SET-029: invalid spinnerTipsOverride shape
(#1435). v2.1.247 extended the
key with tip objects (id, text, cooldownSessions, priority), tipsFile,
and label. Claude Code drops an invalid tip with a debug warning instead of
rejecting the file, so a malformed tip is simply never shown. Validates the
object's fields, each tip entry, the documented id/text limits, the
0-1000 and -10-10 numeric ranges, the 40-character label cap, and
that tipsFile is absolute or ~/-rooted.
CC-PL-016: plugin name not kebab-case. The plugin reference documents
name as a kebab-case identifier with no spaces, and v2.1.247 hardened
marketplace handling to reject names containing control or invisible
characters. Those fail kebab-case by construction, so this catches them before
a plugin reaches a marketplace.

SendFeedback and ListAgents reported as unknown tools
(#1435). KNOWN_AGENT_TOOLS
had not been refreshed since 2026-07-31, so CC-AG-009/010 flagged two
documented built-in tools - ListAgents from Claude Code v2.1.224 and
SendFeedback from v2.1.238 - as invalid names in an agent's tools or
disallowedTools list. Diffed the whole list against the live tools reference,
and these were the only two missing.

Tool baselines refreshed (closes
#1435,
#1436,
#1437). Claude Code
v2.1.246 to v2.1.247, Cline v4.1.15 to v4.1.16, and Codex CLI
rust-v0.149.1 to rust-v0.150.0. Cline's release moves hook workspace
resolution from shared global state to the VS Code window, which changes
runtime behaviour rather than any validated file contract. Codex adds an
Interrupt hook event; CDX-PL-013 validates hook shape without an event-name
allowlist, so nothing false-positives, and the missing allowlist is recorded in
RESEARCH-TRACKING along with the uncovered spinnerVerbs setting.

Tool-release triage produced no summaries. Two separate faults, both of
which the watcher degrades silently to "post the raw changelog": the
GLM_API_KEY secret had expired, and scripts/glm-extract.js ran with an
output budget too small for a thinking model. Reasoning tokens are drawn from
the same max_tokens budget as the answer, so at 4096 the model intermittently
spent the entire budget reasoning and returned finish_reason: "length" with
empty content (reproduced at 1 failure in 3 runs, 4095 of 4096 tokens on
reasoning). MAX_TOKENS is now 32768, roughly six times the 1.9k-5.2k
reasoning observed on real triage prompts, and thinking stays enabled because
it measurably improves the "does this touch a validated config surface?"
judgement. The default model is now named glm-5.3 explicitly rather than
relying on the endpoint's glm-5 alias, so an alias move cannot change triage
behaviour silently. The empty-content error now also reports finish_reason,
reasoning size, and usage instead of just "empty content", and
scripts/glm-extract.test.js - which no workflow ran - is now part of CI.