Release v0.1.1 - #7
Conversation
- Skip valid-JSON non-dict JSONL lines in fpk_count and fpk_correlate instead of crashing or silently truncating a file's counts. - Preserve zero-valued intValue/doubleValue OTel attributes instead of collapsing them to empty strings. - Guard safe_int/safe_float with the same sensitive-key check as safe_str. - Label /Users/ cwd values macos, and classify Windows paths. - Return empty bounds instead of raising when an archive index yields no hive-partitioned blob names, and distinguish an empty --latest-index from an omitted one. - Defer the curses import so fpk_tui --print works on Python builds without curses, and guard month bucketing against malformed timestamps. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The existing tool_events assertion counted rows only, so a broken join predicate still passed with NULL result columns. The fixture also had no timestamp collisions, leaving the scan-order indexing fix untested. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
first_read_efficiency, search_share_vs_pgr, and tool_family_latency were missing from the contents table and run list. METHODOLOGY still said five analyses ship. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
- session_endings: check claim + verification before the tool + completion branch so verified endings are never counted as unverified, and document the two smaller buckets in the header. - field_manual_protocols: LEFT JOIN token_by_session so sessions with no assistant turns stay in queries 5 and 6, guard zero denominators with NULLIF, and drop the dead 'file_edit' family literal. - plausible_completion_candidates: emit preview_is_full, which the audit rubric and classifier prompt already reference; every classification previously received null for it. - session_metrics: populate tool_calls_per_human_message and edit_to_read_ratio instead of leaving the documented columns NULL. - Schema comments now say the four tool_events distance columns and the environment tag are left NULL by the shipped ingest. - cache_economics: state that cache-creation tokens are excluded from the cache-vs-fresh ratio. Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
88cb088 to
d76be4e
Compare
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 88cb088ee6
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| a.*, | ||
| -- text_preview is substr(text, 1, 500), so the preview is the full turn | ||
| -- text whenever the turn fits in 500 chars. The audit rubric uses this. | ||
| (a.text_chars <= 500) AS preview_is_full, |
There was a problem hiding this comment.
When an assistant turn has multiple text blocks, assistant_turns.text_preview is built with string_agg(..., ' ') and then truncated to 500 chars, so the synthetic separators also consume the preview budget. For candidates whose text blocks total <=500 chars but whose joined preview exceeds 500 after separators, this marks preview_is_full=true even though the CSV sent to the audit is truncated, which can cause the verifier to treat missing context as complete. Please base this on the actual joined preview length, or include the separator count in the length check.
Useful? React with 👍 / 👎.
2 participants
Point release: confirmed defect fixes from a full review of the Python and SQL surface, plus the regression tests that pin them.
Fixes
fpk_correlate) or silently truncating a file's counts (fpk_count).fpk_tui --printworks on Python builds withoutcurses; month bucketing guards malformed timestamps.0,safe_int/safe_floatrespect the sensitive-key list,cwdclassification labelsmacosand recognizes Windows paths.--latest-indexis distinguished from an omitted one.session_endings: a closing turn that both claims and verifies now lands in the verified bucket.plausible_completion_candidatesemitspreview_is_full; the classifier prompt and rubric referenced it but every classification received null.session_metrics.tool_calls_per_human_messageandedit_to_read_ratiopopulated instead of NULL.Tests
Three new regression tests: tool_use → tool_result join content, stable row indexing when timestamps collide across files, and
preview_is_fullin the candidates view. 26 pass, coverage 95.43%, ruff/format/mypy/pip-audit clean.Metric impact
Three changes can alter previously reported numbers: the
session_endingsbucket reorder moves claim-plus-verification endings out of the tool + completion bucket, the Field Manual LEFT JOIN adds assistant-less sessions to queries 5 and 6, and the twosession_metricsratios now carry values where prior runs had NULL.🤖 Generated with Claude Code