[Yifan Yang (Yif-Yang)]

Summary

Follow-up to #354. That PR normalized transcript_source (case, surrounding whitespace) only inside harvest_for_config(). The per-(project, source) harvest checkpoint from #322/#312 and the tasks document still used the raw config value. As a result:

• "Codex" harvested Codex but read and advanced a <project>::Codex cursor. That cursor may hold a time written while the misspelling was silently harvesting Claude, so older Codex sessions could be skipped (the skillopt-sleep: harvest checkpoint is keyed by project, so a second --source in the same project skips its unseen sessions #312 class of bug).
• A non-string value raised AttributeError instead of a clear error.

This adds normalize_transcript_source() in skillopt_sleep/harvest_sources.py and uses that one canonical value everywhere:

• the harvest itself;
• the cursor read and writes in run_sleep_cycle and in the handoff pin path;
• the tasks document written by the handoff pin path and by cmd_harvest.

run_sleep_cycle now validates the source before it creates the night's staging directory, so an invalid value no longer leaves an orphan staging dir. Canonical names keep byte-identical cursor keys, and the legacy bare-<project> cursor is still read as the fallback for claude. get_backend() and state.py are unchanged.

CHANGELOG.md [Unreleased] → Changed now also tells users that a misspelled backend or transcript_source stops the run instead of silently using the mock backend or Claude transcripts (#350, #354, thanks Hrushikesh Ramilla (@Hrushikesh-ramilla)).

Verification

• 13 new tests in tests/test_harvest_checkpoint_source.py. 12 fail on current main for the intended reasons: the stale ::Codex cursor is read, the tasks document records " Codex ", a list value raises AttributeError, and an orphan staging dir is left behind. All 13 pass here. Canonical keys and the legacy bare-project cursor are pinned unchanged.
• Full isolated suite: 2027 passed, 13 skipped, 4 xfailed, 8 warnings, 381 subtests passed in 47.78s. mkdocs build --strict passes. New code is ruff-clean; the one remaining finding is an existing I001 in cycle.py's import block, also on main.

Compatibility notes:

• Users with a non-canonical spelling switch to the canonical cursor, falling back to the legacy cursor or the lookback_hours window. The stale ::Codex entry is left in state.json but is no longer used.
• Runs with seed_tasks or pinned digests now also reject an invalid transcript_source.

[Yifan Yang (Yif-Yang)]

Merging this maintainer follow-up to #354. Background: an independent re-check of #354 (Claude Code, cross-checked here) found that transcript_source was normalized only inside harvest_for_config(), while the per-source harvest checkpoint from #322/#312 still used the raw spelling. So "Codex" harvested Codex but read a ::Codex cursor that could hold a stale time. The change was drafted with Claude Code and reviewed and re-run here before merging:

• 12 of the 13 new tests fail on main and all pass here;
• canonical cursor keys and the legacy bare-project cursor are unchanged;
• full isolated suite: 2027 passed, 13 skipped, 4 xfailed, 8 warnings, 381 subtests passed in 47.78s;
• official CI is green on this head.

The CHANGELOG entry credits Hrushikesh Ramilla (@Hrushikesh-ramilla) for #350/#354.