Skip to content

Record three mechanics the promotion loop had to learn twice - #468

Merged
ptr727 merged 1 commit into
developfrom
docs/runbook-cli-drift-and-closing-keyword
Jul 31, 2026
Merged

Record three mechanics the promotion loop had to learn twice#468
ptr727 merged 1 commit into
developfrom
docs/runbook-cli-drift-and-closing-keyword

Conversation

@ptr727

@ptr727 ptr727 commented Jul 31, 2026

Copy link
Copy Markdown
Owner

Three things that cost a wrong turn while driving promotion #460, none of them written down anywhere. Each is recorded as the general rule plus the concrete symptom, so the next agent recognizes it rather than rediscovering it.

A closing keyword never fires under this branching model

GitHub closes a referenced issue only on a merge into the repository's default branch, and every feature pull request here targets develop. #464 carried Closes #462 and merged, and #462 stayed open, reading as unfixed until it was closed by hand. The promotion that later reaches main carries the commit rather than the keyword, so it does not close it either.

This is fleet law rather than a Copilot mechanic, and it binds any agent regardless of provider, so per this file's own rule it belongs in GOVERNANCE.md rather than the runbook. It is a top-level branching-model rule rather than a third entry under "Executing a develop -> main promotion safely", because the pull request it bites is a feature one merging into develop, not the promotion itself. That parent says "two traps" and means it.

gh pr edit is broken by the classic-Projects sunset

It fails with GraphQL: Projects (classic) is being deprecated ... (repository.pullRequest.projectCards) and mutates nothing. That matters during a long review loop, where the promotion body goes stale as fixes land under it. The runbook now names gh api -X PATCH \"repos/<owner>/<repo>/pulls/<N>\" -F body=@<file> and says to read the body back afterwards, since a failed call is not evidence the pull request is untouched.

Filed as its own list rather than added to the existing one, which is specifically about review request paths. This is an edit the loop makes between rounds.

A poll that tests a captured result against != \"0\" reads empty as success

An empty string is exactly what a mis-written --jq filter returns, so the two cases that must be distinguished, a query finding nothing and a query running wrong, both satisfy the test. One such poll during #460 reported a review that had not landed, and the next message said so before the correction. Counting matches and testing -gt 0 makes them read alike.

The specific trap named with it: --arg is a gh api graphql flag. Passing it to a plain gh api --jq call fails with accepts 1 arg(s), received 4 while the surrounding $(...) still yields the empty string.

Placed in "Verify Review Covered Current Head", one paragraph above the existing warning about exiting on mergeStateStatus, since both are ways a poll exits early on a false signal.

Verification

Documentation only, no behavior change, so no test accompanies it. 71 self-tests and 19 repo_gate tests pass, repo_gate is clean, charset and dupword exit 0, and the warn-only backlog is unchanged at dash 962 and semicolon 388, so the new prose adds no findings of its own.

Fidelity note

Both files carry verbatim sections, so downstream copies are stale until re-vendored. That is already true of them from #460.

🤖 Generated with Claude Code

Each of these cost a wrong turn during #460 and none was written down.

A closing keyword never fires under this branching model, because GitHub only
closes on a merge into the default branch and every feature pull request targets
`develop`. #462 sat open with `Closes #462` in the merged body, looking unfixed.
That is fleet law rather than a Copilot mechanic, so it goes in the branching
model beside the other two promotion traps.

`gh pr edit` is broken by the classic-Projects sunset. It fails on
`repository.pullRequest.projectCards` and mutates nothing, which matters when a
promotion body needs refreshing as fixes land under it. The runbook now names
the REST form that works, and says to read the body back rather than trust that
a failed call changed nothing.

A poll that tests a captured `gh api --jq` result against `!= "0"` reads an
empty string as success, and an empty string is what a mis-written filter
returns. One such run reported a review that had not landed. Counting matches
and testing `-gt 0` makes a query that finds nothing and a query that ran wrong
read alike.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Copilot AI review requested due to automatic review settings July 31, 2026 01:10

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Records three promotion-loop “learned twice” mechanics as durable governance/runbook documentation so future promotions avoid the same wrong turns.

Changes:

  • Add a branching-model governance note that closing keywords don’t auto-close issues when PRs merge to develop (default branch is main).
  • Document that gh pr edit is broken due to Projects (classic) deprecation and provide an alternative gh api edit path.
  • Strengthen the “Verify Review Covered Current Head” polling guidance to avoid false positives from empty/failed query results.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File Description
GOVERNANCE.md Adds a branching-model rule about why Closes #N doesn’t fire under the develop-targeting PR model.
.github/copilot-instructions.md Updates the Copilot review runbook with gh pr edit failure mode and safer polling guidance.

Comment thread GOVERNANCE.md

Copilot AI left a comment

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.

@ptr727
ptr727 merged commit f35e8c7 into develop Jul 31, 2026
8 checks passed
@ptr727
ptr727 deleted the docs/runbook-cli-drift-and-closing-keyword branch July 31, 2026 01:16
ptr727 added a commit that referenced this pull request Jul 31, 2026
Both findings from Copilot's review of promotion #469, against the
mechanics #468 had just recorded. Both were written from inference
rather than from checking, in the commit whose subject was recording
mechanics accurately.

## The closing keyword is not "never"

Verified rather than reasoned:

```
gh repo view --json defaultBranchRef --jq '.defaultBranchRef.name'  ->  main
```

A keyword fires whenever the pull request **itself** merges into the
default branch. A promotion, or a Dependabot security update opened
against `main`, therefore closes what it references. Only a feature pull
request, which targets `develop`, cannot.

The rule is inverted to state when the keyword *does* work, so the
absolute has nowhere to hide, and it names the promotion and bot cases
explicitly. It also adds something the original missed: a promotion body
can carry a keyword deliberately, which is a real option rather than a
trap.

## `--arg` is jq's flag, not gh's

Copilot filed this as "likely to go stale or be incorrect across GitHub
CLI versions". It was incorrect on arrival, which is the more useful
finding:

```
gh api graphql --help   ->  --cache, --hostname, --input, --paginate, --silent, --verbose
jq --help               ->  --arg name value      set $name to the string value
```

`gh api graphql` takes `-f` and `-F`. The `accepts 1 arg(s), received 4`
error came from `gh` treating the extra tokens as positional arguments,
and I reasoned the flag's ownership from that error instead of reading
`gh api --help`.

The paragraph now leads with the durable general rule, that a failed
command writes to stderr and leaves stdout empty so the surrounding
`$(...)` yields the empty string the test reads as success. The error
string stays as a recognizable symptom, without the claim about which
tool owns the flag.

## Why this matters more than its size

`GOVERNANCE.md` and `.github/copilot-instructions.md` are vendored
verbatim across the fleet. A false statement in either propagates to
every repository that re-vendors, and it arrived in the commit that
exists specifically to stop the next agent reasoning from a symptom.
Both corrections came from a review round rather than from me re-reading
my own work.

## Verification

Documentation only, no behavior change, so no test accompanies it. 71
self-tests and 19 repo_gate tests pass, `repo_gate` is clean, `charset`
and `dupword` exit 0, and the warn-only backlog is unchanged at dash
962, comment-wrap 454, semicolon 388, comment-case 56.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
ptr727 added a commit that referenced this pull request Jul 31, 2026
From the low-confidence block of Copilot's second review of promotion
#469. It flagged one duplicate. Checking the other two additions from
#468 found a second, and the second is worse than a duplicate.

## What #468 got wrong

That pull request said it was recording three mechanics "none of them
written down before". I did not grep the files before writing to them.

**`gh pr edit`** was already covered at
`.github/copilot-instructions.md` "PR Edits and Merge-State Gotchas",
which gives **both** the GraphQL and the REST form and says to verify
the edit took. The existing entry is better than the one I added beside
it, so the addition goes and the original stays untouched.

**Issue-closing keywords** were already covered in `GOVERNANCE.md` under
the release model, and that entry is not merely earlier but **correct
where mine was not**:

> Issue-closing keywords (`Closes #N`, `Fixes #N`) go in the `develop ->
main` promotion PR, not the feature -> develop PR.

That is a working mechanism. My branching-model bullet said to close the
issue by hand instead. Two bullets, one topic, **different procedures**,
which is a contradiction in fleet law rather than a repetition, and
exactly the drift Copilot warned the duplicate would cause.

It also means my reply on #470, that the promotion mechanism was
untested and so the rule should stay silent on it, was answering a
question this repository had already answered. #462 could have been
closed by putting the keyword on promotion #460.

## What survives

The release-model rule gains the one thing mine had that it lacked: the
fallback for a develop pull request that already merged with a keyword
on it. That is the case #462 actually hit, and without it a reader who
has already made the mistake finds no instruction.

The polling guard from #468 **stays**. Guarding an empty bot node id
before a mutation is documented in two places already, but the **exit
test of a poll loop** is not, and the numeric comparison is what stops
an empty result reading as a landed review. Kept as the one genuinely
new thing in that commit.

## Why this happened, and what would prevent it

The closing-keyword rule lives under "Release Model" while the question
I was answering was a branching one, so reading the Branching Model
section did not surface it. That is a findability problem in a 400-line
governance file rather than an excuse: a `grep` for the topic would have
found it in either section, and adding to a rules file without grepping
it first is the actual failure.

Not proposing a reorganization here. Moving committed rule text between
sections changes two verbatim sections at once and every downstream copy
with them, which is not something to ride along with a fix.

## Verification

```
gh pr edit mentions      ->  1 (copilot-instructions.md:209)
closing-keyword bullets  ->  1 (GOVERNANCE.md:83)
```

71 self-tests and 19 repo_gate tests pass, `repo_gate` is clean,
`charset` and `dupword` exit 0.

🤖 Generated with [Claude Code](https://claude.com/claude-code)

---------

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>
ptr727 added a commit that referenced this pull request Jul 31, 2026
Promotes `f35e8c7` (#468), `a5e12a2` (#470), and `7477c42` (#472).
Conflict-free, three commits ahead.

The net change against `main` is **three lines in two files**. Reviewing
this promotion is what reduced it to that, so the history and the result
are described separately below.

## What actually lands

**A polling guard, in the Copilot review runbook.** A poll that captures
a `gh api --jq` result and exits on `[ "$found" != "0" ]` treats an
**empty** string as a landed review, and an empty string is what a
mis-written filter returns. Counting matches and testing `-gt 0` makes a
query that finds nothing and a query that ran wrong read alike. One such
poll during #460 reported a review that had not landed.

**A fallback on the existing issue-closing rule, in `GOVERNANCE.md`.**
The rule already said to put `Closes #N` on the promotion pull request
rather than the feature one. It did not say what to do once a develop
pull request has already merged carrying the keyword. It now does: move
the keyword to the promotion body, and close by hand only when the
promotion has merged without it. That is the case #462 hit.

## What was reverted, and why that matters more

#468 opened claiming three mechanics "none of them written down before".
Two of them **were**, in the files it edited.

`gh pr edit` being broken by the classic-Projects sunset was already
documented under "PR Edits and Merge-State Gotchas", with both the
GraphQL and the REST form. The addition was a plain duplicate and is
gone.

Issue-closing keywords were already documented under the release model,
and that entry is **correct where the addition was wrong**. It gives a
working mechanism, the keyword on the promotion pull request. The added
branching-model bullet said to close the issue by hand instead. Same
topic, two sections, different procedures, which is a contradiction in
fleet law rather than a repetition, and is the drift a duplicate is
supposed to risk only later. That bullet is gone and the surviving rule
absorbed the one thing it lacked.

Two further corrections landed on the way. The closing-keyword bullet
first said a keyword "never fires under this model", which is false
because a pull request merging **into** `main` closes what it
references. The polling note called `--arg` a `gh api graphql` flag,
which is false because `--arg` belongs to `jq` and `gh api graphql`
takes `-f` and `-F`. Both files are vendored verbatim across the fleet,
so either statement would have propagated on the next re-vendor.

## Why promote now

What survives is small and true. The polling guard is the only new
mechanic in the set, and it is the one that produced a wrong report
during #460 rather than a hypothetical. The `GOVERNANCE.md` sentence
completes a rule that was already right by covering the state a reader
reaches only after getting it wrong.

`main` is what a newly scaffolded or realigning repository carries, and
what the fleet audit reads as ground truth, so a rule that is complete
on `develop` and partial on `main` is a rule that argues with itself
across the fleet.

## Fidelity note

`GOVERNANCE.md` and `.github/copilot-instructions.md` both carry
verbatim sections, so downstream copies stay **stale** until
re-vendored. They were already stale from #460 and this does not change
that state, only its size, which is now three lines rather than the
eleven #468 first proposed.

Documentation only, no behavior change. 71 self-tests and 19 repo_gate
tests pass, `repo_gate` is clean, `charset` and `dupword` exit 0, and
the warn-only backlog is unchanged at dash 962, comment-wrap 454,
semicolon 388, comment-case 56.

🤖 Generated with [Claude Code](https://claude.com/claude-code)
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants