zeikar cc-changelog — Claude Code changelog watch (hyperclaude relevance)
Repo-local dev tool (NOT a plugin skill). Compare a stored "last-checked" Claude Code version against the current latest, mine the official changelog for everything in between, and report only what actually touches this repo's surfaces — then optionally advance the stored version so the next run diffs from here.
The value is not "summarize the changelog" — it is mapping generic Claude Code changes onto the specific surfaces of the hyperclaude plugin and verifying impact against the real repo files, exactly as a human maintainer would.
State file
.claude/skills/cc-changelog/.last-checked-version (sibling of this file, gitignored):
a single line holding the last version the maintainer reviewed, e.g. 2.1.201. It is
the diff baseline; it advances only in Step 6.
Procedure
Step 1 — Baseline
The diff baseline is .claude/skills/cc-changelog/.last-checked-version (one X.Y.Z
line) — call it STORED. Step 2's helper reads it automatically; you don't Read it here.
If the file is missing/empty the helper exits 2 — then ask the user which version to
treat as the baseline (do not invent one) and pass it as the script argument.
Step 2 — Fetch and slice the changelog
The full CHANGELOG.md is ~5k lines; do not pull it whole into context. Run the
local helper — it fetches the changelog and prints to stdout only the entries strictly
newer than the baseline (it version-compares, so skipped numbers like 2.1.191→2.1.193 are
handled; no temp file, no cleanup):
node .claude/skills/cc-changelog/get-changelog.mjs # reads .last-checked-version
# or with a user-supplied baseline: node .claude/skills/cc-changelog/get-changelog.mjs 2.1.201
Read the Bash output directly (no Read, no temp file):
- line 1
LATEST=<x.y.z>— the topmost changelog version → this isLATEST. - the rest — the sliced entries, verbatim, newest first (or
(up to date — …)if none).
If the fetch fails (exit 1), fall back to WebFetch on
https://raw.githubusercontent.com/anthropics/claude-code/main/CHANGELOG.md asking for
entries strictly greater than STORED.
Step 3 — Up-to-date short-circuit
If the output is (up to date — nothing newer than <STORED>) (i.e. LATEST == STORED,
nothing newer), report "already current at STORED, nothing new" and STOP. Do not
rewrite the state file.
Step 4 — Map each entry to a hyperclaude surface
For every entry newer than STORED, decide which surface (if any) it touches, then
classify into one of four buckets. Report a change only if it lands on a surface;
drop pure end-user-UI/terminal/voice items unless they affect this repo.
Buckets:
- 🔧 Actionable — needs a change in this repo (a skill/agent/hook/doc edit).
- ✅ Favorable — the platform now does natively what the plugin did manually; no action, but note it as a future-simplification candidate.
- 🟢 Safe / no-op — plausibly relevant but verified harmless for this repo.
- ℹ️ Informational — worth knowing, behavior-neutral.
Surface checklist (what to grep/read to confirm impact — never assert impact without checking the actual file):
| Surface | What to check | Repo anchors |
|---|---|---|
| Hooks | hooks.json matcher semantics (hyphen/comma/regex), SessionStart & PostToolUse events, hook stderr/exit-code behavior |
hooks/hooks.json, hooks/*.mjs |
| Loop spawn/reply transport | background-agent spawn semantics for a no-name: Agent call, whether the final text still arrives as the task-notification <result>, SendMessage routing by returned agentId, subagent prompt-cache TTL, and any change to plugin-agent definition resolution (upstream #78234 / #81746) |
references/loop-protocol.md, skills/hyper-*-loop/** |
| Agent dispatch (non-loop) | run_in_background default, subagent depth cap, Agent(type) deny/allow rules, subagent model inheritance |
skills/hyper-implement/SKILL.md, skills/hyper-plan, hyper-tdd, hyper-debug, hyper-research, hyper-docs-sync |
| Skill/agent/command frontmatter | name/description/metadata.* parsing, kebab/snake/camel, malformed-YAML handling, slash-command loading |
skills/**/SKILL.md, agents/*.md, commands/*.md |
| Codex bridge / Bash spawn | anything changing how Bash spawns or sandboxes child processes (Codex is external, but the bridge is spawned via Bash) |
scripts/codex-bridge.mjs, scripts/codex/*.mjs |
| Plugin manifest / validate / release | claude plugin validate, plugin.json, marketplace/rename behavior |
.claude-plugin/plugin.json, docs/development.md |
| Overlapping built-ins | built-in /code-review, /agents wizard, default model, /review — usually informational, but flag stale doc references |
docs/**, README.md |
For any flagged-but-uncertain item, deepen before concluding: consult
https://docs.claude.com/en/docs/claude-code (WebFetch) or WebSearch, and grep/read
the anchor files to confirm the actual impact.
Step 5 — Report
Lead with a one-line TL;DR answering "does anything need action?". Then the four
buckets (omit empty ones or say "none"). Cite repo files as clickable
[path](path) links. Keep it tight — this mirrors a maintainer triage, not a full
changelog reprint.
Step 6 — Advance the baseline
After reporting, offer to update .last-checked-version to LATEST (default: yes).
On confirmation, overwrite the file with LATEST + newline. Skip the bump if the user
wants to keep the current baseline for re-review. Never bump before reporting.
Notes
- Primary source is the CHANGELOG (authoritative per-version); official docs are a secondary deepening source for a specific flagged feature.
- This skill only reads and reports — it never edits plugin code as a side effect. Any actionable finding is handed back to the user to act on (e.g. via hyper-plan).