lifecycle: let a draft declare its gate (Closes-when: / Blocked-by:) - #168
Merged
Conversation
…part Leg 1 of draft/feature/pyautomind/draft_staleness_detection_signals.md, filed from the draft/ sweep. Five prompts in that sweep had a stated gate that had since closed, and nobody noticed -- one of them (the test-mode umbrella) had its exit condition met the SAME DAY the status line was written. The blocker was never detection, it was ambiguity. A draft citing an issue can mean two opposite things: "epic closes when #70 ships its recipe leg" (closed => this prompt is DONE) or "BLOCKED until PyAutoArray#431 merges" (closed => this prompt is READY). Prose cannot be graded, so `issues --drafts` had to report both as one question, "shipped, or newly unblocked?", and an advisory that cannot say which is not actionable. Two optional header keys let the author say which they mean: Closes-when: autolens_profiling#70 Blocked-by: PyAutoArray#431, PyAutoGalaxy#486 `issues --drafts` now reports GATE MET (verify, then retire) and UNBLOCKED (ready to start) as separate bands, and prompts that declare a gate drop out of the ambiguous list -- so that list becomes the genuinely-undeclared remainder rather than everything. Design decisions worth recording: Aggregated PER PROMPT, not per reference. A prompt blocked on three PRs is unblocked when the LAST one lands; reporting each ref separately would have claimed "ready to start" three times while it was still blocked. Partly-closed gates get their own weaker band, which is the real state of ep_analytic_updates -- its WP1 gate merged, its WP3/WP4 gates are open. Repo#123 shorthand, not just URLs. Measured on the real backlog: a URL-only extractor found 2 refs, the shorthand form found 8. Prompts do not write URLs. Fenced blocks are documentation. The prompt that PROPOSED these keys shows them in a ```markdown example; parsing that as a declaration would have invented a finding out of the proposal itself. Caught while backfilling, and covered by a test that fails without the fence check. Still advisory, never a gate on the exit code -- retiring a prompt writes to complete/ and stays human, the same contract intake reconcile keeps. Backfilled the five prompts whose gates this sweep confirmed, giving a live ready-to-start list: sampler_cli_output_workspace_sweep (PyAutoFit#1436), preloads_advanced_workspace_guide (PyAutoLens#565), unpark_imaging_scaling_relation_slam (PyAutoArray#431), oversampled_psf_dataset_adoption (three k x s merges), and ep_analytic_updates (partial). Verified against the states this session confirmed via the API: 4 fully unblocked, 1 partial, 0 falsely flagged. 9 new tests, each proving the leg fails without the code -- including the two opposite readings, the per-prompt aggregation, the fenced-example case, and that a declared gate is not also asked as an ambiguous advisory. 118 -> 127 passing. Co-Authored-By: Claude Opus 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01E3MuurHXi3xo9TLRpMLJA6
…I got wrong Legs 1 and 2 of draft_staleness_detection_signals.md are built. Leg 3 (the --repo upstream mode) is not, and is now better motivated. The "all five confirmed findings in the top band" criterion was WRONG WHEN WRITTEN, and contradicts this prompt's own Hard limit section three paragraphs above it -- that section already says one finding left no Mind-side signal at all, so a Mind-local ranker cannot rank it by construction. Chasing the criterion anyway made the tool worse, which is the useful part. A loose <work-type>/<target>/ series match pulled one more finding in, but also FALSELY flagged test_mode_bypass_ordered_assertion_ties -- off references to four unrelated sibling prompts in the same folder -- and that prompt is one the sweep confirmed is NOT shipped. It is precisely the mis-grade the tool must never make, and it was already the criterion immediately below. So the two criteria were in direct conflict. The real finding is that the five need three different tools, not one better ranker: rare-token fan-out and shared identifiers (reconcile) catch two, the Closes-when: header key catches one, and the last two are only reachable by reading the target repo. Recorded as a table so the next session does not re-derive it, along with the corrected criteria and the measured outcome. Co-Authored-By: Claude Opus 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01E3MuurHXi3xo9TLRpMLJA6
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Leg 1 of
draft/feature/pyautomind/draft_staleness_detection_signals.md, filed from the draft/ sweep (#167). Pairs with PyAutoBrain#215 (leg 2, the reconcile re-rank) — independent, mergeable in either order.The problem was ambiguity, not detection
Five prompts in the sweep had a stated gate that had since closed and nobody noticed — one had its exit condition met the same day the status line was written.
A draft citing an issue can mean two opposite things:
Prose can't be graded, so
issues --draftshad to report both as one question — "shipped, or newly unblocked?" — and an advisory that can't say which isn't actionable.Two optional header keys
issues --draftsnow reports GATE MET (verify, then retire) and UNBLOCKED (ready to start) as separate bands. Prompts that declare a gate drop out of the ambiguous list, so that list becomes the genuinely-undeclared remainder rather than everything.Design decisions worth reviewing
Aggregated per prompt, not per reference. A prompt blocked on three PRs is unblocked when the last one lands; reporting each ref separately would have claimed "ready to start" three times while it was still blocked. Partly-closed gates get their own weaker band — the real state of
ep_analytic_updates, whose WP1 gate merged while WP3/WP4 remain open.Repo#123shorthand, not just URLs. Measured on the real backlog: a URL-only extractor found 2 refs, the shorthand form found 8. Prompts don't write URLs.Fenced blocks are documentation. The prompt that proposed these keys shows them in a
```markdownexample — parsing that as a declaration would have invented a finding out of the proposal itself. Caught while backfilling; covered by a test that fails without the fence check.Still advisory, never gates the exit code. Retiring a prompt writes to
complete/and stays human — the same contractintake reconcilekeeps.Backfilled, giving a live ready-to-start list
Five prompts whose gates this sweep confirmed. Against the states verified via the API: 4 fully unblocked, 1 partial, 0 falsely flagged.
sampler_cli_output_workspace_sweeppreloads_advanced_workspace_guideunpark_imaging_scaling_relation_slamoversampled_psf_dataset_adoptionep_analytic_updatesTests
9 new, each proving the leg fails without the code: both opposite readings, per-prompt aggregation, the fenced-example case, and that a declared gate isn't also asked as an ambiguous advisory. 118 → 127 passing.
lifecycle check/orphans/index --checkall OK.Also corrects an acceptance criterion in the source prompt that was wrong when I wrote it — see the second commit; the same note appears on PyAutoBrain#215.
Generated by Claude Code