fix: order same-second human feedback rows newest first in the Python store - #1458

Summary

• get_human_feedback and get_calibration_examples in autocontext/src/autocontext/storage/sqlite_store_human_feedback.py now order by created_at DESC, id DESC instead of created_at DESC alone.
• This is the Python half of fix: human feedback queries returned the oldest rows, not the newest #1360. That PR fixed the same two queries in the TypeScript store and recorded that "the Python side has no counterpart to this SQLite store". The Python store does have both queries, with the same ORDER BY created_at DESC LIMIT ? and the same one-second created_at default (autocontext/migrations/006_human_feedback.sql:8, and the pip bootstrap schema in storage/bootstrap_schema.py).

Reproduction on main: 12 rows inserted through insert_human_feedback in one second, all sharing one created_at:

get_human_feedback("s", limit=10)       -> ids [1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
get_calibration_examples("s", limit=5)  -> ids [1, 2, 3, 4, 5]

With the tiebreak:

get_human_feedback("s", limit=10)       -> ids [12, 11, 10, 9, 8, 7, 6, 5, 4, 3]
get_calibration_examples("s", limit=5)  -> ids [12, 11, 10, 9, 8]

SQLite returns the tied rows in insert order, so the newest rows of a same-second burst come last, and once the burst reaches the LIMIT they are cut off: a caller asking for the most recent rows gets the oldest ones. In Python that affects:

• the judge calibration anchors (limit=5) passed to evaluate_output and run_improvement_loop (MCP tools in mcp/agent_task_tools.py and mcp/knowledge_tools.py) and to live rubric calibration in execution/agent_task_completion.py
• get_feedback (the MCP autocontext_get_feedback tool), and the lessons that export_agent_task_skill copies into an exported skill package (limit=10)

id is the AUTOINCREMENT rowid, so it is monotonic in insert order and breaks the tie the way DESC already intended.

Surfaces Touched

• Parity decision: this brings Python in line with the TypeScript fix in fix: human feedback queries returned the oldest rows, not the newest #1360; no TypeScript change is needed
• Python package
• TypeScript package
• TUI
• Docs or examples
• CI or release metadata

Verification

• cd autocontext && uv run ruff check src tests: clean
• cd autocontext && uv run mypy src: no issues in 950 source files
• cd autocontext && uv run pytest: two new regression tests in tests/test_human_feedback.py. On main they fail with assert [1, 2, 3, 4, 5, 6, ...] == [12, 11, 10, 9, 8, 7, ...] and assert [1, 2, 3, 4, 5] == [8, 7, 6, 5, 4]; with the fix they pass. The tests pin every row's created_at to one value after inserting, so they reproduce a same-second burst without depending on timing. The storage, MCP and calibration test files also pass on Python 3.11, the CI version (227 passed, 1 skipped). Full suite with -m "not live" on Python 3.12: 11,339 passed, two more than on main (the new tests). The same 166 tests fail on main and on this branch in my sandboxed environment, mostly with IsolationUnavailableError: local child isolation is unavailable; none touch human feedback.
• cd ts && npm run lint: not applicable, no TypeScript changes
• cd ts && npm test: not applicable, no TypeScript changes
• uv run python ../scripts/check_markdown_links.py: passed

Docs And Release Impact

• no user-facing docs changes needed
• updated relevant README/docs/examples
• updated CHANGELOG.md (Unreleased, Fixed)
• updated version metadata if this is part of a release

Notes

• No API, schema or migration change, and no backfill: only the read order changes, and no stored data is wrong.
• As fix: human feedback queries returned the oldest rows, not the newest #1360 notes, giving created_at millisecond precision would make ties rarer without removing them, so the tiebreak is the fix either way. The migration is left alone.

🤖 Generated with Claude Code