[dangrondahl]

💊 Blue pill. One of two alternative PRs that remove the shared Anthropic API key from the release flow. Only one should merge; the other is the red pill, #1300.

bin/suggest-version-ai.sh now calls the Claude Code CLI (claude -p) under the releaser's own Claude login, instead of calling the Messages API with a key from 1Password (or ANTHROPIC_API_KEY). make release behaves as before: same prompt, same parsing, same $EDITOR step and confirmation.

• The call runs from an empty temp dir with all tools disabled, so it stays a plain diff-in, text-out request and doesn't pick up the repo's CLAUDE.md or MCP servers.
• The curl checks are gone from the Makefile; the release guide now lists jq and a logged-in Claude Code CLI as prerequisites.
• After merge, the op://Shared/Anthropic API Key item can be deleted if nothing else uses it.

Choose this if you want make release to stay a fixed, scripted flow.

Tested locally (writing to a scratch dir, no tag created): since v2.48.2 it suggested a patch with "No user-facing CLI changes" (correct, only CI changed); since v2.48.0 it suggested a minor with a changelog under three headings. claude missing and an invalid model both exit 1 and leave no stale output files.

Checklist

• Docs are autogenerated from CLI help — make sure the help text reflects your changes (n/a: no CLI changes)
• Helm chart (charts/k8s-reporter/) updated, if needed (n/a)
• Terraform provider and related changes updated, if needed (n/a)

[claude]

Claude finished @dangrondahl's task in 1m 22s —— View job

---

PR review: claude -p instead of the shared API key

• Static checks (go build, go vet, go mod tidy)
• Review the diff (Makefile, bin/suggest-version-ai.sh, release-guide.md)
• Post inline comments

Static checks: go build ./..., go vet ./... and go mod tidy all pass, and go.mod/go.sum are unchanged. The PR has no Go changes.

Findings

1. The temp dir isn't deleted after a successful run (bin/suggest-version-ai.sh:92). The trap - EXIT at the end of the script clears the trap that removes $WORK_DIR. Each successful make release leaves /tmp/tmp.XXXX/input behind, holding the prompt and up to 50 KB of the diff. The old DIFF_FILE leaked the same way, so this isn't new. The fix is one line: rm -rf "$WORK_DIR" right after the claude -p call.
2. Nit: isolation is partial (bin/suggest-version-ai.sh:97-104). Running from an empty cwd keeps out the repo's CLAUDE.md and MCP servers. The releaser's user-level ~/.claude/CLAUDE.md, hooks and plugins still load, and they could change the output format. --setting-sources "", or --bare if the CLI version supports it, would make the call fully deterministic. Not a blocker: the parser falls back to patch, and the $EDITOR step lets the releaser fix odd output.

What's good

• Removing the shared secret is a real security gain: no 1Password item to rotate, and usage is tied to each releaser's own account.
• --tools "" plus an empty cwd keeps the call a plain diff-in, text-out request, and the comment explains why.
• Errors are handled more robustly. select(.is_error == false) and the || true guards make a missing CLI, an auth failure or a bad model end in a clean exit 1, and the trap still clears stale output files.
• The claude existence check gives an actionable message that points to the make release tag=vX.Y.Z fallback.
• The Makefile, script header and release guide are updated together; dropping the curl check is correct.

Summary

The PR is small and clean, and it's ready to merge once the temp dir cleanup (finding 1) is fixed. Finding 2 is optional hardening.

[claude]

Nit on isolation: the empty cwd keeps out the repo's CLAUDE.md and .mcp.json. The releaser's user-level config still loads, though: ~/.claude/CLAUDE.md, user hooks such as SessionStart/Stop, and plugins. A personal "always answer in X style" memory, or a hook that prints to stdout, could bend the output away from the BUMP: / ---CHANGELOG--- format. If you want the call fully deterministic, look at --setting-sources "", or --bare if your Claude Code version supports it. Also consider --append-system-prompt/--system-prompt to pin the format. Not a blocker: the parser already falls back to patch, and the $EDITOR step catches any odd output.

[claude]

Temp dir leaks on success. The trap - EXIT near the end of the script (line 158) clears this trap, so rm -rf "$WORK_DIR" only runs on failure. Every successful make release leaves a /tmp/tmp.XXXX/input holding the prompt and up to 50 KB of the diff. The old DIFF_FILE leaked the same way, so this isn't new, but it's a one-line fix. Add rm -rf "$WORK_DIR" just before the final trap - EXIT, or clean it up right after the claude -p call since nothing reads it afterwards:

RESPONSE=$(cd "$WORK_DIR" && claude -p ... < input) || true
rm -rf "$WORK_DIR"