Childhood

Stay with a release after it ships. Verify what users receive and what production serves, then fix verified regressions within the user's authorized scope. A successful merge or publish command is the start of this work.

Establish the release

• Identify the repository, release version, immutable tag commit, merge commit, production deployment, and affected delivery surfaces. Use the release receipt and actual provider/registry state; do not guess from the current branch or today's latest build.
• Record the release, observation start/end, checks, findings, and next action with work_state_write when available. Keep announcements and unrelated infrastructure failures distinguishable from product regressions.
• Carry forward existing authorization. Monitoring does not itself authorize another release, rollback, production migration, or external announcement. Prepare and validate a concrete fix before requesting any authorization still missing.

Verify the delivered release

Choose the checks relevant to its delivery surfaces; independent checks can run in parallel.

For a claude-mem search release, exercise Smart search as three separate calls: search for an index, timeline for context, and get_observations for only the selected records. Exercise optional dumb_search explicitly where available, and readable invalid-input guidance through fresh local and hosted MCP connections when those surfaces shipped. Verify that Smart is the default, no auto or guided mode or continuation cursor is exposed, and tool results are concise text without pure JSON or duplicated structured metadata. Keep probes read-only unless writes are part of the explicit test scope. Never print credentials or private result bodies.

Publication can lag behind a successful command. An exact-version 404 or stale distribution tag remains pending until the registry agrees; use bounded retries and investigate the actual publish target/status before republishing. A warning alone is not a regression: verify the final artifact. When the repository publishes through CI, a failed publish job is a finding to fix through the normal release path, not to cover with a manual publish; for claude-mem, also confirm the published version carries a provenance attestation.

Observe and repair

• Honor the requested observation window. Without one, watch through 30 minutes after the release first becomes publicly available or production becomes ready, whichever affected surface is later. For an older release, do a fresh sweep and inspect the completed late CI and reports covering that interval.
• Poll pending or changing checks every 30–60 seconds; check steady signals every 2–5 minutes. Keep individual waits at most 60 seconds so user steering and progress updates remain responsive. An unavailable API is a coverage gap, not evidence that the release is healthy.
• Verify each finding before changing code. Distinguish a release regression, a pre-existing issue, an operational failure, and a resolved/transient signal; give each its own evidence and next action.
• Make focused fixes, run checks appropriate to the change, and use the repository's normal PR/release/deployment process. Use babysit for a repair PR and version-bump when a new claude-mem release is authorized. Published versions and existing tags stay immutable.
• After a repair ships, restart observation for the changed delivery surface and rerun its affected user path. Preserve unrelated worktrees, workers, registrations, user settings, and inference/provider policy.
• A repeated infrastructure failure or deleted notification webhook needs a concrete follow-up. Keep it pending with its cause and owner action; do not repeatedly retry it or announce success to another destination.

Finish with evidence

Finish when the relevant CI is settled, delivered artifacts and affected user paths pass, no actionable release regression remains, and the observation window is satisfied. Do one final fresh sweep before reporting.

If a required check cannot be completed, report the release as partially verified with that coverage gap and next action. Separate an externally blocked operational follow-up from verified product health; do not silently mark either as resolved.

Return concise purpose-specific text: release/tag/deployment identity, checks and observation times, verified findings and repairs, and anything still pending. Keep raw API envelopes, internal metadata, and complete logs out of model-facing output. Link to a compact receipt when useful.