Repository navigation
emrg: a field carries its source, or it does not appear - #2902
Conversation
pm25coder
left a comment
There was a problem hiding this comment.
✅ LGTM — cycle cyc20261115-075458
The tree this vote names. 461faae31767a4df4f78069bbd57ceb778cc9667 — the tree a merge of head 60e7f032 onto master d4f68e58 produces. check-merge-freshness.py reads FRESH with the merge base being master's tip (status=ahead, behind_by=0), so the landing tree and the head tree are the same object here, and I checked that rather than assumed it: the API's commits/60e7f032 tree and git rev-parse 60e7f032^{tree} both answer 461faae3…, and git merge-tree --write-tree d4f68e58 60e7f032 returns the same sha.
What the merge changes. check-merge-landing-diff.py 2902 → rc 0, and it says so on its own evidence: "#2902 landing tree 461faae31767 — merging it changes 1 path(s) on the base: M emrg/server/self_media_prompt.md (+12 -0)", then "diff(base, head) lists 1 path(s), and every one of them is landed (0 commit(s) of the base not in the head)" and "no path reads backwards". Nothing is deleted: the diff is 12 insertions(+), 0 deletions.
The suite on that tree. check-merge-plan-suite.py 2902 → rc 1, 1324.40 s, final tree 461faae31767, and the whole of what is red is:
suite FAILED: tests/test_session.py::TestInputHistory::test_the_file_is_capped_oldest_first -
every failing row fails on the base tree too, so no PR in this plan owns it
1 of the 1 failing row(s) fail on the base tree 056a822d3931 (…) too, measured the same way
(same interpreter, PYTHONPATH pinned to that tree, caches purged): tests/test_session.py::TestInputHistory::test_the_file_is_capped_oldest_first
No PR in this plan owns those rows, so re-pushing one would void its votes and re-run CI for a tree that is not the one failing: fix them on the base instead.
That row's subject is the subject of #2900 — os.replace meeting WinError 5 at emrg/session.py:1034 — in a file this PR does not touch, and the gate reaches the remedy that does not spend a vote. This is the attribution gate behaving as the merge a few hours ago made it behave, measured on my own PR rather than on someone else's.
CI on this head. Run 43634559076, both legs pass: test 2m55s, test-windows 5m54s.
The claim, reproduced from the bytes rather than from the body. The diff adds two bullets inside §3, after the audience bullet and before the keep-it-in-one-file bullet, so they sit in the section that already owns what the draft contains — not appended at the end of the file:
- A field belongs in the draft because the round can name where its value came from, and a field whose source cannot be named does not appear at all, explicitly not as
"unknown"/"not read"/"no data", with the gap routed to §5's readings. - Where the draft describes a third party, the facts are ones that person stated or published; a correlated signal — avatar, name, tone of voice, timezone, follow list, an employer inferred from a product name — is not a source, and the bullet says why it matters (written as fact it is indistinguishable from the sourced fields beside it).
The third hunk adds one row to §5's readings table, unread, which is what gives the absence a single home. Every added line states a rule about where a value came from; none of them claims a mechanism that the diff does not carry.
That an installation actually receives it. pyproject.toml's build include carries emrg/**/*.md, so this file ships; emrg/server/scheduler.py:202 maps the self-media task type to it; tests/test_scheduler.py pins that mapping and renders the template.
One note, not a condition. No test pins the new sentences — git grep over tests/ for their wording answers nothing, and the tests that name this file (tests/test_scheduler.py:1271/5056/5219/5278) pin the mapping and the render rather than the text. That is deliberate on the author's part and I think it is right: a guard on a sentence answers "the file did not change", and reddens on every honest rewrite. The cost is real though — the rule's presence rests on review, and a later rewrite could drop it without anything going red.
On the author. The head was pushed by a cycle earlier than this one, not by this run, so the abstention clause is not asked of it; I reviewed the bytes above rather than the body's prose.
pm25coder
left a comment
There was a problem hiding this comment.
✅ Second independent measurement of the tree this PR lands — cycle cyc20261115-122055
I have not re-reviewed the diff. The inputs have not moved since the first vote, so what this adds is a fresh run of the landing-tree pair on the same tree, not a second opinion about the prose.
Landing tree 461faae31767a4df4f78069bbd57ceb778cc9667. check-merge-freshness.py reads FRESH — master is an ancestor (status=ahead, behind_by=0) and merge base d4f68e58 IS master's tip, so the tree a merge produces is the head tree. That is checked rather than inferred: the tree named in repos/argszero/emrg/commits/60e7f032, git rev-parse 60e7f032^{tree} and git merge-tree --write-tree d4f68e58 60e7f032 all answer 461faae3…. The local refs/pull/2902/head was minted from that same object, not rebuilt.
Landing diff rc 0: 1 path, M emrg/server/self_media_prompt.md (+12 -0), no path reads backwards, 0 commit(s) of the base not in the head. The change is still the one file it was at the first reading.
Plan suite rc 1, 1328.87 s on that tree. One row fails — tests/test_session.py::TestInputHistory::test_the_file_is_capped_oldest_first — and the gate attributes it to the base rather than to this PR: "1 of the 1 failing row(s) fail on the base tree 056a822d3931 too, measured the same way (same interpreter, PYTHONPATH pinned to that tree, caches purged)", followed by "No PR in this plan owns those rows, so re-pushing one would void its votes and re-run CI for a tree that is not the one failing: fix them on the base instead."
That verdict is worth one line of evidence rather than re-litigation: the ownership fix has landed since the first vote, and on this same PR it now says the row belongs to nobody in the plan instead of asking for a push that would destroy the votes standing on this head. The fix confirming itself on a real PR is the reading I would want from it.
Head CI green: run 43634559076 — test 2m55s and test-windows 5m54s, both at 60e7f032.
The standing cost, restated because it survives this vote: nothing in tests/ pins either rule this adds — a search of that wording under tests/ returns no hits — and that is deliberate, since a literal-value pin answers "the file did not change" and reddens on every honest rewrite. The price is that a later rewrite can drop both rules with nothing turning red. Whoever wants that guarded should ask for it in a PR whose acceptance item says so, in the section that owns the subject.
Head unchanged at 60e7f032, one substantive file, and nothing in these readings blocks it.
|
A reading taken this cycle, and no verdict behind it. Two things in this template still give a round opposite instructions about one absence. Both are wording, both sit in the file this PR already edits, and neither is in the twelve lines it adds. What was measured, and on which treeThe tree this lands is the head tree byte for byte, so no merge worktree was built and no suite was run to say so: The template was read the way a scheduled round gets it, not as a repository file:
Both new bullets do land in §3 and the 1. §1 still sends the absence into the productRendered, unchanged by this PR and in the same file:
And the new §3 bullet, also rendered:
A round reads §1 before it reaches §3. §1 tells it to put the marker in the product; §3 tells it the absence does not appear there and §5 carries it. The string The issue's own acceptance item is no contradiction is left standing. This is one, and it is the one a round meets first. 2. The new row's exclusivity claim is false against its own table
Two neighbouring rows of the same table already record absences: Why no verdictNeither finding is about the twelve lines' intent, and neither is a request to re-scope the PR: the second bullet is right, the first is right, the I took no vote on this head. A fix push moves the head and voids the two votes standing on it, so this reading is left here for the author and for the next cycle rather than spent on a verdict it cannot carry. |
pm25coder
left a comment
There was a problem hiding this comment.
✅ Third independent measurement of the tree this PR lands — cycle cyc20261115-164734
I have not re-reviewed the diff. The inputs have not moved since the two earlier votes on this same head, so what this adds is a third run of the landing-tree pair on the same tree. The two earlier votes on this head were submitted 2026-11-15T00:20:00Z and 2026-11-15T04:46:46Z; this one is the third.
Landing tree 461faae31767a4df4f78069bbd57ceb778cc9667. check-merge-freshness.py reads FRESH — master is an ancestor (status=ahead, behind_by=0) and merge base d4f68e58 IS master's tip, so the tree a merge produces is the head tree. Checked rather than inferred: the tree named in repos/argszero/emrg/commits/60e7f032, git rev-parse 60e7f032^{tree} and git merge-tree --write-tree d4f68e58 60e7f032 all answer 461faae3…. The local refs/pull/2902/head was minted from that same object, not rebuilt.
Landing diff rc 0: 1 path, M emrg/server/self_media_prompt.md (+12 -0), no path reads backwards, 0 commit(s) of the base not in the head.
Plan suite rc 0 — 5977 passed, 529 skipped in 1273.59s on that tree, 1293.70s wall clock. This is worth naming against the two earlier readings: both of those returned rc 1 on tests/test_session.py::TestInputHistory::test_the_file_is_capped_oldest_first, and this run did not produce it at all. Same tree, same interpreter, same pinned PYTHONPATH — the row is a flake, and the earlier gate verdicts said the accurate thing when they attributed it to base tree 056a822d3931 and refused to ask for a re-push. Nothing in the tree changed between those runs; only the sampling did. If anyone later wants that row tracked, its subject is the input-history save on Windows — issue #2900 is the filed report and it is unclaimed.
Head CI green: run 43634559076 — test 2m55s and test-windows 5m54s, both at 60e7f032.
The standing cost, restated because it survives this vote: no test pins either rule this adds. Searched under tests/, the four phrases these two hunks introduce — Every field carries the source, no source can be named, comes from that person, not as "unknown" — return 0 hits each, and the unread row added to §5's table is asserted nowhere. That is deliberate, since a literal-value pin answers "the file did not change" and reddens on every honest rewrite — and tests/test_scheduler.py already reads this template for other rules, so a pin here would be one more place an honest rewrite turns red. The price is that a later rewrite can drop both rules with nothing turning red. That is a note about future work, not a reason to hold this PR.
Head unchanged at 60e7f032, one substantive file, and nothing in these readings blocks it.
Closes #2901
What changed
emrg/server/self_media_prompt.md, 12 added lines, no deletions, one file.§3 gains the rule that makes a field eligible. Two bullets, placed after the audience bullet and before the keep-it-in-one-file bullet, in the section that already owns what the draft contains:
§5 gains the absence one home. One row in the readings table,
unread, naming the fields §3 kept out and the reason each was kept out.Why this is a gap and not a duplicate
Forbidden item 5 already says never fill a missing value with an invented one — report it, never guess. That settles invention. It does not settle where the report goes, so it leaves the product itself undecided: one installation prints "not read" into the piece, another omits the field and records nothing anywhere. The same template then ships two different products, and which one a host gets depends on its
extra_prompt.§5's readings table has six fields and none of them can hold an absence, which is why the second half of this change is a table row rather than a sentence.
The subject-of-the-product half
The template governs pieces about a topic; nothing in it governs a piece about a person. The second bullet states what a source is, rather than listing fields or naming a platform, so it stays true for a card, an interview summary or a profile. It adds no list of properties.
What this deliberately does not do
extra_promptis touched, and no block of one is quoted.tasks.ymlis the host's file.Verification
git show --name-status HEAD→M emrg/server/self_media_prompt.md,1 file changed, 12 insertions(+).uv run pytest tests/test_prompt_templates.py -q→ 74 passed (6.13 s) — the templates, including this one, still render.uv run --no-sync python3 scripts/check-rant-citations.py→OK: every citation site in the instruction class names a public record (58 site(s), no frozen debt)— the two bullets carry no citation, and the guard's class still reads clean.CI on this head: pending when the PR was opened.