When a user (often via Slack, GitHub, or email) sends a pre-built patch + a
structured instructions file (UPSTREAM-PROMPT.md style) and asks you to open
the PR — clone the target repo, branch from origin/main, git am the patch,
run the user's listed test suites verbatim, push, and open the PR with the
title/body the user supplied. Distinct from `always-pr-never-local-edit`
(which dispatches fix-X work to AO), and from `github-pr-workflow` (general
PR lifecycle); this is the bounded "land this patch" workflow where the
intent, title, body, and verification gates are pre-specified.
v1.2.0 adds Pitfalls P8 (force-push after origin branch deletion), P9
(two-tier "delete X gate" distinction — workflow gate vs pipeline node
type), and P10 (pre-existing test-failure attribution via stash/reset/
restore). Verified case #3 (2026-07-21, dark-factory PR #407 skeptic-
gate deletion + receipt-gate-reviewer follow-up commit).
Standardmäßig ist der Prompt ausgewählt, der zuerst die Quelle prüft. Sie können zu einem direkten Befehl wechseln oder eine lokale Kopie herunterladen.
Quelldateien prüfen
Lesen Sie SKILL.md und alle von SkillsMP angezeigten Begleitdateien, bevor Sie sich für eine Installation entscheiden.
Mit Codex oder Claude installieren Kopieren Sie diesen Prompt, fügen Sie ihn in Codex, Claude oder einen anderen Assistant ein und lassen Sie die Skill-Seite prüfen und installieren.
Ein direkter Befehl überspringt den Prüf-Prompt. Prüfen Sie die Quelle, bevor Sie ihn ausführen.
When a user (often via Slack, GitHub, or email) sends a pre-built patch + a
structured instructions file (UPSTREAM-PROMPT.md style) and asks you to open
the PR — clone the target repo, branch from origin/main, git am the patch,
run the user's listed test suites verbatim, push, and open the PR with the
title/body the user supplied. Distinct from `always-pr-never-local-edit`
(which dispatches fix-X work to AO), and from `github-pr-workflow` (general
PR lifecycle); this is the bounded "land this patch" workflow where the
intent, title, body, and verification gates are pre-specified.
v1.2.0 adds Pitfalls P8 (force-push after origin branch deletion), P9
(two-tier "delete X gate" distinction — workflow gate vs pipeline node
type), and P10 (pre-existing test-failure attribution via stash/reset/
restore). Verified case #3 (2026-07-21, dark-factory PR #407 skeptic-
gate deletion + receipt-gate-reviewer follow-up commit).
when_to_use
- User message has an attached `.patch`/`.diff` file AND a text description
of where it should land ("apply this to disk_magician", "open a PR on
<repo> with this", etc.)
- Slack/GitHub/email message body is short ("Handle", "apply this", "open
the PR") because the *files* are the ask, not the body text
- User-provided instructions file lists exact test suites and exact PR title
+ body template (often called UPSTREAM-PROMPT.md, RUNBOOK.md, HANDOFF.md)
triggers
["Handle","apply this patch","open the PR with this","land this upstream","git am this","use the instructions in the attached file"]
Apply a user-supplied patch to its target repository, run the listed verification
suites, push, and open the PR with the title + body the user provided. Distinct
from always-pr-never-local-edit (route-through-AO for general fix work) and
from github-pr-workflow (general PR lifecycle).
When this skill fits
This is the right skill when ALL of the following hold:
The user has sent a pre-built patch file (typically .patch or .diff
attached to a Slack/GitHub/email message). The diff is not something you
need to write.
The user has sent structured instructions — either inline in the
message body or in a sibling instructions file (often named
UPSTREAM-PROMPT.md, RUNBOOK.md, HANDOFF.md, or <name>-PROMPT.md).
These instructions explicitly call out:
The target repository (e.g. jleechanorg/disk_magician)
The branch name to create (e.g. findings-wiki-contract)
The list of test suites to run (e.g. tests/test_safety_lib.sh)
The PR title + body template
Whether to merge or leave for review
The work is bounded — pre-specified, not iterate-to-green. If the patch
doesn't apply cleanly, fall back to git apply --3way per the instructions,
then conflict resolution per the recipe below.
If ANY of these is missing (no patch, no instructions, OR unclear target),
this is NOT the right skill — fall back to either:
always-pr-never-local-edit if the user said "fix X" and you have to
figure out the diff yourself
github-pr-workflow if it's a general PR task without a supplied patch
Trigger-1 anti-pattern: don't narrate before reading attachments
When a Slack message has a body word like "Handle" (1-2 words, generic) AND
file attachments, the attachments are the ask, not the body. The wrong
move is to immediately write narration messages like:
"I see file attachments..."
"Let me work on this..."
"On it — applying the patch..."
before checking what the attachments actually are. The right move is to
download the attachments first, read them, then compose a single coherent
response. Verified failure mode (2026-07-15, jleechanorg/disk_magician
PR #17): the bot wrote 4 narration messages before realizing the body text
was empty/short and the actual ask lived in the patch + UPSTREAM-PROMPT.md
file pair.
The rule: if mcp__slack__conversations_replies shows a parent message
with files: [...] populated and a body of ≤5 words, the first tool call
in the response turn MUST be reading the files — not writing a narration
message.
Recipe — the 7-step bounded workflow
Step 1 — Pre-flight checks (parallel, fast)
# A) gh auth status
gh auth status 2>&1 | head -3
# B) Does the target repo exist + what is its default branch HEAD?
gh api repos/<OWNER>/<REPO> --jq '{default_branch, pushed_at, archived}'
gh api repos/<OWNER>/<REPO>/branches/main --jq '{sha, date}'# C) Does a PR for this topic already exist? (REST, not GraphQL — GraphQL hits# rate limits fast and you can avoid them)
gh api "repos/<OWNER>/<REPO>/pulls?state=all&head=<OWNER>:<branch>&per_page=10" \
--jq '.[] | {number, title, state, head_sha, url}'# D) List open PRs + branches matching the topic to avoid stepping on existing work
gh api "repos/<OWNER>/<REPO>/pulls?state=open&per_page=20" \
--jq '.[] | {number, title, head_ref, author: .user.login}'
gh api "repos/<OWNER>/<REPO>/branches?per_page=100" \
--jq '.[] | select(.name | test("<topic-regex>")) | .name'
Decision matrix (after pre-flight):
Finding
Action
Auth OK, repo exists, HEAD matches the patch's base SHA exactly (git log --oneline origin/main -1 returns the SHA named in the patch message)
Proceed — git am will apply cleanly
Auth OK, repo exists, HEAD differs from the patch base SHA
Try git am; on conflict, abort and use git apply --3way <patch>, resolve, then commit. Expected conflict points per the instructions file (e.g. pyproject.toml version bump, CLAUDE.md append)
Auth OK, repo exists, but PR for this branch/topic already exists
STOP. Confirm with the user — do NOT push onto a non-owned PR head (see pr-cleanup-replay skill, Phase -1). Either push to the existing branch and comment on the PR, OR close the existing PR and open a fresh one.
A prior-session PR for the SAME topic is OPEN but the new patch is a "refreshed/supersedes" version (new instructions file says "REFRESHED … supersedes the YYYY-MM-DD patch", OR the existing PR's base SHA is N commits behind current origin/main)
This is the supersede pattern — see Pitfall P7. Open a NEW branch from current origin/main (e.g. <branch>-v2), apply the refreshed patch, open a NEW PR, then close the prior PR with a "superseded by #N" comment linking to the new PR. Do NOT rebase the prior PR; do NOT push onto its branch head. The new PR is the source of truth.
A related branch exists with the same topic but not yet PR'd
Inspect it: gh api .../compare/<patch-base-sha>...heads/<related-branch> — if ahead_by > 0 and the topic matches, consider whether to push onto that branch instead of opening a parallel one
Step 2 — Clone to a scratch dir (NOT the user's home directory)
SCRATCH=/tmp/<ta[REDACTED_OPENAI_KEY]>-<unix-ts>
mkdir -p "$SCRATCH"cd"$SCRATCH"
git clone --quiet https://github.com/<OWNER>/<REPO>.git repo
cd repo
# Verify local HEAD matches origin/main (sanity)
git rev-parse HEAD # should equal the SHA from Step 1.B
git fetch origin main --quiet
git rev-parse origin/main # same as above
Why scratch /tmp/: the user's home directory has a directory-state tracker
(CLAUDE.md "Working directory lock — stay in the session's primary cwd"),
and writing to an existing home dir creates the "file modified since read"
divergence error. A scratch dir is fully isolated.
Step 3 — Branch + apply the patch
# Branch from origin/main (clean, no carried-over history)
git checkout -B <branch-name> origin/main
git rev-parse HEAD # confirm same as origin/main# Apply the patch (provided the base SHA matches)
git am <path-to-patch>
echo"git am exit: $?"# If git am fails (HEAD has moved past the patch base):# 1. git am --abort# 2. git apply --3way <patch># 3. Resolve conflicts per the instructions file (often only pyproject.toml# version + CLAUDE.md appends are conflict points)# 4. git add <resolved-files># 5. git commit -m "<original-commit-message>"
Step 4 — Run the user-listed test suites verbatim
The instructions file lists specific test paths. Run EACH one and capture the
exit code. Do NOT substitute "I think these tests are equivalent" — the user
gave you an exact list because those are the gates they want on the PR.
echo"=== <test-suite-1> ==="
bash <path-to-test-1> 2>&1 | tail -15
echo"exit=${PIPESTATUS[0]}"echo"=== <test-suite-2> ==="
bash <path-to-test-2> 2>&1 | tail -15
echo"exit=${PIPESTATUS[0]}"# ...repeat for each suite listed in the instructions file
Required evidence per suite:
Exit code (printed)
Last 10-15 lines of output (so the user can see PASS/FAIL counts)
For pure-PASS suites: the line that says "ALL TESTS PASSED" or equivalent
If a test FAILS: STOP. Do NOT push. Do NOT open the PR. Report the
failing test verbatim in the Slack reply and ask the user how to proceed.
Step 5 — Verify guardrails from the instructions file
Most UPSTREAM-PROMPT.md files list guardrails like:
"No machine paths (/Users//...) belong in any committed file"
"Must not add any file under findings_wiki/ other than README + TEMPLATE"
"Run findings_lint.sh --upstream for purity"
Verify each guardrail with a grep/ls/the linter the instructions file
calls out. The pitfalls section below has worked examples.
Open the PR with title + body from the instructions file (verbatim, but
un-indented if it was indented by the user):
~/.hermes/scripts/gh-safe-publish pr create \
--repo <OWNER>/<REPO> \
--base main \
--head <branch-name> \
--title "<exact title from instructions file>" \
--body-file /tmp/<ta[REDACTED_OPENAI_KEY]>/PR_BODY.md
# Capture
gh pr list --head <branch-name> --json number,title,url,state
Do NOT merge unless the instructions explicitly say "merge it yourself".
Most UPSTREAM-PROMPT.md files say "leave the PR for the repo owner to review"
out of politeness — but the agent's default is also "no merge without explicit
MERGE APPROVED" per SOUL.md merge discipline.
Step 7 — Post the Slack reply
If the user sent this from Slack, post ONE reply with:
PR URL (linked markdown, see .cursor/rules/pr-hyperlink.mdc)
Branch name + final SHA
Exit code + tail of each test suite
A summary of the guardrail checks
The patch location if archived (gist URL or local path)
If mcp__slack__conversations_add_message returns not_in_channel (sub-class
5f, cross-workspace bot-token hard-block), follow SOUL.md
slack-cross-workspace-fallback-xoxp — fall back to SLACK_USER_TOKEN via
Path B curl. Note in the message body if the message will appear under the
user's identity instead of the bot identity.
Schedule a one-shot 20m check-back cron per
## COMMIT: one-time-status-cron-after-every-task so the user gets a
follow-up only if CodeRabbit posts a verdict or the user comments.
Step 8 — Cleanup
rm -rf /tmp/<ta[REDACTED_OPENAI_KEY]>/repo /tmp/<ta[REDACTED_OPENAI_KEY]>/PR_BODY.md
# Optional: preserve the patch itself at /tmp/<ta[REDACTED_OPENAI_KEY]>/<patch-file># for re-application if the user wants to redo the PR
If Step 1.C finds an existing PR whose branch or title matches the new topic,
DO NOT push onto that branch's head. The user's instructions file may not
mention an existing PR because they were written assuming a clean slate. Stop
and ask. The verifying-SHA check (gh api repos/.../branches/<branch>) tells
you whose PR it is — if it's not yours, the right move is to open a NEW PR
on a NEW branch from origin/main, leaving the existing PR alone.
P2 — The instructions file lists "git am" but git am fails
The instructions file usually has a section like "If git am fails... use
git apply --3way". But the actual conflict resolution is project-specific.
Common conflict points across jleechanorg/* repos:
pyproject.toml version bump — instructions usually say "take the
higher version and bump patch". The pattern is to read the current version,
read the patch version, take max(current, patch) and increment patch by 1.
CLAUDE.md append — instructions usually say "append the two new
sections at the end". The pattern is cat >> CLAUDE.md << 'EOF' then
commit with the same message.
Sub-tree duplication — some jleechanorg/* repos mirror scripts/ to
src/<package>/scripts/. If the patch touches one but not the other, the
conflict resolution is to also apply the patch to the mirroring path
(often a separate cp or rsync).
If the conflict is none of the above, STOP and ask. Do not invent a resolution.
P3 — Tests fail on placeholder paths the user told you to ignore
The instructions file may say "No machine paths (/Users//...) belong
in any committed file" — but the patch itself may contain
/Users/nobody/protected-tree as a TEST FIXTURE in tests/test_*.sh. These
are deliberate placeholders (/Users/nobody/ = "no specific user"). Do NOT
report them as a guardrail violation. Verify the placeholder is in a test
file (tests/ or scripts/test*) AND uses /Users/nobody/ (not the user's
real home), then move on.
P4 — Both Slack tokens lack files:write for binary evidence
Per evidence-attach-to-slack v1.8.0, the OAuth scope gap on both
HERMES_SLACK_BOT_TOKEN and SLACK_USER_TOKEN means you cannot upload the
patch to the Slack thread directly. v1.9.0 covers the text-file case: push
the patch to a public gist via gh gist create <file> (text is text, no
clone-and-replace dance). Capture the SHA via git clone https://gist.github.com/<id>.git
git rev-parse HEAD and embed https://gist.githubusercontent.com/<user>/<id>/raw/<sha>/<file>
in the Slack reply.
P5 — The user wants to be the assignee, not you
By default, the PR author is whoever ran gh pr create (which is whoever
authenticated gh). That's fine for jleechanorg/* repos where
jleechan2015 is the active gh user. But if a different account did the
auth, the PR is owned by that account. Verify gh pr view <N> --json author
matches the user the instructions expect. If it doesn't, add --assignee
to the gh pr create call.
P6 — Slack thread routing for the summary reply
After opening the PR, post the summary to the SAME thread the user started
in (per SOUL.md slack-reply-inherit-thread-ts). If the user's most recent
message was a HERMES-bot self-message, scan further back — do NOT reply to
the bot self-message. The not_in_channel error on conversations_add_message
indicates you need the xoxp fallback (P5 above cross-references this).
P8 — Force-pushing a follow-up commit when the origin branch was deleted by a prior session
When the prior session shipped a PR by force-pushing, then the agent
crashed/ended before merge — and a NEW session is asked to follow up on the
same PR by adding commits — the workflow is:
Verify origin/PR state via gh pr view <N> --json headRefOid — head
should still resolve even if the feature branch is gone locally.
When pushing, use plain --force, NOT --force-with-lease. The lease
target (the deleted remote ref) doesn't exist on origin, so
--force-with-lease will return "stale info" and fail. Plain
--force rewrites the same path (refs/heads/<branch>) and is correct
here because no other agent or branch is sharing that ref.
Verify with gh pr view <N> --json headRefOid returns your new SHA
before claiming the push landed.
Verified 2026-07-21 on jleechanorg/dark-factory PR #407: prior session
shipped commit f461f93da, then origin branch receipt-gate-reviewer was
deleted (no on-disk lease target). Follow-up commit b04df6f44 pushed
with --force succeeded; gh pr view 407 --json headRefOid then
returned b04df6f44960db5dca15b0539e9360d8956736bd confirming PR head
moved correctly.
P9 — Two-tier "delete X gate" distinction (workflow gate vs pipeline node type)
When the user says "delete the X gate" in a dark-factory / agent-orchestrator
project, distinguish TWO unrelated layers:
Workflow-level gate: .github/workflows/X-gate*.yml files. These
define GitHub Actions checks that block PRs (e.g. skeptic-gate-caller.yml
triggers on pull_request_target and posts a skeptic status check).
Deleting these removes the CI gate.
Pipeline-level gate type: a handler registered in
runner/handlers.py TYPE_REGISTRY (e.g. gate_skeptic →
_gate_skeptic(node, ctx) in runner/handler_universal_prompts.py)
used by DOT pipelines as a type="gate_X" node. Many .dot pipelines
many tests + many fixtures depend on this registration.
Deleting the workflow gate does NOT affect the pipeline node type. Most
"delete X" requests target the workflow gate (the CI blocker); deleting
both is rare and requires updating fixtures + tests.
P10 — Pre-existing test-failure attribution via stash/reset/restore
When you delete code that some tests depend on, the deletion will break
those tests. To attribute which failures are caused by your deletion vs.
pre-existing on origin/main (per SOUL.md same-test-name-rule):
# 1. Capture current changes
git stash
# 2. Reset to origin/main
git reset --hard origin/main
# 3. Run the failing tests on clean main
.venv/bin/python -m pytest tests/test_X.py -v --no-header 2>&1 | tail -20
# 4. Restore the branch + changes
git checkout -B <branch> refs/remotes/origin/pr/<N> # or origin/<branch>
git stash pop
# 5. Compare results — if same SHA, same exit code, same error line →# PRE-EXISTING (not your fault). Attribute correctly in the PR summary.
Verified 2026-07-21 on jleechanorg/dark-factory PR #407: 4 test
failures observed after skeptic-gate deletion; stash/reset confirmed all
4 had identical SHA + identical output on origin/main (3× git-lfs PATH
drift, 1× macOS /bin/false not-found). Zero new failures caused by
deletion.
P11 — Fresh destructive user comments OVERRIDE prior closeouts
When the user previously claimed a PR was closed out (e.g. comment 5111574771
on your-project.com PR #8661 said "all CodeRabbit blockers fixed"), then
later posts a file comment like "Delete this" or "Do not add things specific
to a campaign", the destructive comment overrides the closeout — even
if the closeout was posted by the same agent. Workflow:
Before executing any closeout work, re-fetch
gh api "repos/<OWNER>/<REPO>/issues/<N>/comments?per_page=100" and
gh api "repos/<OWNER>/<REPO>/pulls/<N>/comments?per_page=100" and
sort by created_at. If any comment is destructive-verbing
(delete, remove, revert, do not add, drop, simplify to)
AND newer than the most recent closeout claim, the closeout is stale.
Reclassify the work as destructive-removal: delete the named
files/constants, then update the tests that referenced them, then push.
Do NOT do another round of "fix the CodeRabbit findings" — the user
is telling you the whole PR is the wrong shape.
Ask one question, not five. If the destructive comment is ambiguous
(e.g. "delete this" without specifying scope), ask one clarify with
the scope choices. Do not invent a destructive scope and do not redo
the full closeout.
Reference the newer comment in the PR comment so the reviewer can
see why the work shifted from "fix CodeRabbit" to "delete the artifact".
Verified 2026-08-01 on $GITHUB_REPOSITORY PR #8661: a
2026-07-29 closeout claimed all 5 CodeRabbit P1/P2 items were fixed.
Two 2026-08-01 user file comments ("Do not add things specific to a
campaign" on constants.py, "Delete this" on the overlay) overrode the
closeout. Net effect: 4 Spellblade-specific artifacts deleted, 910 lines
removed, +44/-0 net diff left on the PR (just generic shared-contract
text). The earlier fix work was not wasted — it informed the right answer
("delete the artifacts, keep the generic sections") — but the agent must
NOT have blindly executed the closeout instructions.
P12 — Explain rebase-before-push necessity in the approval ask, not after
If a destructive-Removal rebase is required (e.g. the local branch needs
to be rebased onto current origin/main before push, so the cleanup
commit is a fast-forward onto the remote tip), make the rebase reason
part of the approval ask, not a follow-up explanation. The user
should not have to ask "why is force-push needed?" after the push lands.
Approval-ask template (use before any force-push that requires prior user
approval):
Need force-push to feat/<branch> because the local branch was rebased
onto current origin/main (N newer commits behind your remote tip
<sha>), so the cleanup commit can't be a fast-forward onto the old
remote tip. Alternative: open a new branch + new PR (safer, reviewable
independently). Approve force-push, or pick the new-branch route?
Why this matters: the user typed "force push is fine although i donts
ee why thats needed" — the user accepted the push, but the framing landed
as mysterious. The next session should reverse this: name the reason
in the same message as the approval, not after.
Verified 2026-08-01 on $GITHUB_REPOSITORY PR #8661: the
agent should have surfaced "the local branch is 6 commits behind
origin/main because I rebased onto current main first" in the
approval ask, not in the post-push reply.
P13 — Multi-reviewer fan-out loop on destructive-removal cleanup (iteration gate)
After a P11 destructive cleanup, dense PRs (multiple CodeRabbit items,
multiple reviewers, dangling references, generic-vs-campaign rules)
need a multi-reviewer second-opinion pass before the work is shippable.
The pattern is:
List the open review items before writing any new code: gh api "repos/<OWNER>/<REPO>/pulls/<N>/comments?per_page=100" (issue-level) + the
latestOpinion GraphQL review-thread query (per-thread resolution
state). The 2-axis matrix is: severity x staleness. Moot items (target
file deleted) you can cross off in 30 seconds; live items need a fix plan.
Run /advice (3 reviewers) + /web-advice (multi-model) in parallel.
The reviewer roles are:
Reviewer A (source-accuracy): byte-equality of any copied files
(e.g. shared contracts cherry-picked from a sibling branch), include
resolution, test suite pass count, MBTI/alignment/campaign-shaped
string scan across the diff.
Reviewer B (external-docs / spec compliance): repo conventions
(AGENTS.md, the prompts/CLAUDE.md), commit-attribution rule, README
alignment, hidden internal-only data leakage.
Reviewer C (adversarial): cross-agent coverage, runtime failure
modes for directives, orphan references, conflict with other recent
changes.
/web-advice is the cross-model independent review — it does NOT
share the prior context. Use codex exec (Codex CLI) as the
recorded second-opinion if Aside-side Web LLM tabs (ChatGPT/Grok)
are not available on the active account (Aside CLI on a Google-only
account returns "model not available" for openai/gpt-5, gpt-5.x,
anthropic/claude-sonnet-4.5, google/gemini-2.5-pro). Codex CLI is
always available. Pull the diff into evidence/<pr>.diff and feed
it as the prompt.
Convergence signal: when 2 of 3 /advice reviewers + Codex agree
on a finding, the finding is real — fix it. Don't fix reviewer-unique
nits unless they fall in the user's destructive scope.
Re-run after fix. Don't compose a one-shot pass — the iteration
is the value. Watch the live transcript (~/.hermes/cache/delegation/ live/<deleg-id>/task-N.log) for the final status=completed block
before trusting the verdict.
Final-state acceptance requires:
All reviewer APPROVED-as-is verdicts collected (or only the
deferred-to-PR-A items remaining).
Codex APPROVE.
4-lane (ponytail/ZFC/root-cause-first) all PASS for the diff.
Test suite green on the rebased branch.
Diff is N files / +X / -Y with no MBTI/alignment player-facing
exposure (verified by grep against the 16 MBTI type codes +
big-five + alignment labels).
Then force-push + PR comment + 20m babysit cron.
P13a — Cross-prompt coverage gap (the DialogAgent pattern)
When a {{PROMPT_INCLUDE:}} directive lives in a parent prompt (e.g.
narrative_system_instruction.md) and the rule is described as
"mandatory for every agent on every turn", but a lighter agent
(DialogAgent) only loads a stripped variant (narrative_lite),
the claim is false. Two real signals that this is happening:
grep -A 5 "REQUIRED_PROMPT_ORDER" $PROJECT_ROOT/agents.py DialogAgent
returns a list that includes narrative_lite but not narrative.
The prompt description says "every agent" but the lighter parent
doesn't have the same {{PROMPT_INCLUDE:}} directive.
The fix has two options:
Option
When to use
Cost
Include from the lighter parent too (recommended)
The rule is genuinely global (e.g. npc autonomy, witness rules)
+12 lines per lighter parent
Soften the wording in the heavy parent
The rule is heavy-parent-specific (e.g. long-form narrative only)
One-line edit, but you lose the "mandatory everywhere" claim
The include-from-lighter-parent option is almost always right because
the rule is genuinely global — DialogAgent narrates, narrating needs
the contract. Verified 2026-08-01 on PR #8661, where the Codex
review flagged the gap and the fix was to add 4 {{PROMPT_INCLUDE:}}
directives to narrative_lite_system_instruction.md covering the
same 4 narrative contracts.
P13b — Pulling shared contracts from a sibling PR (the PR-A cherry-pick pattern)
When the destructively-removed code referenced contracts that live
on a sister PR (e.g. PR-A feat/shared-contracts-mbti-internal-drive),
don't redeclare the contracts inline — pull them from the sister
branch via git show <branch>:<path> and write them as new files.
This is the right move when:
The sister PR has already been reviewed and the contracts are
considered load-bearing.
The contracts are pulling shared payload (rules, state schemas) that
the destructive-removal PR needs to reference.
The sister PR is the canonical source of truth for the contracts.
The pattern is:
# Verify byte-equality after pullfor f in <list-of-contracts>; do
diff -q \
<(git show origin/<sister-branch>:$PROJECT_ROOT/prompts/shared/$f) \
$PROJECT_ROOT/prompts/shared/$fdone# Should print 0 differences# Pull only the contracts you need (not the whole branch)for f in <list-of-contracts>; do
git show origin/<sister-branch>:$PROJECT_ROOT/prompts/shared/$f \
> $PROJECT_ROOT/prompts/shared/$fdone
Do NOT pull the sister PR's MBTI-internal-data file (e.g.
ai_generated_mystery_and_internal_drive_plot_arc.md) — that's the
internal-only personality model that lives on PR-A only. The user
explicitly stated "Did not touch PR #8539 (MBTI internal-only
canonical contract) or the PR-A worktree" in the closeout comment.
Verify with ls <shared>/ and grep for "mystery" / "internal_drive" /
"MBTI" before commit.
Then wire the contracts via {{PROMPT_INCLUDE:}} in the parents
that need them. The {{PROMPT_INCLUDE:}} directive is the canonical
mechanism documented in $PROJECT_ROOT/prompts/AGENTS.md and $PROJECT_ROOT/prompts/ CLAUDE.md §"Shared rules across agents". Existing examples:
Use markdown-link references (See <file>.md for the rule.) ONLY for
human reading; the {{PROMPT_INCLUDE:}} directive is what the runtime
resolves. The {{PROMPT_INCLUDE:}} form is parsed by
$PROJECT_ROOT/agent_prompts._resolve_prompt_includes (fail-fast on
missing/unsafe targets — see agent_prompts.py:817-881).
Verified 2026-08-01 on PR #8661: 5 shared contracts pulled from PR-A
feat/shared-contracts-mbti-internal-drive (byte-identical, verified
via diff -q), wired into narrative + combat + narrative_lite, all
3 lighter agents (DialogAgent, LiteDialogAgent via narrative_lite,
StoryModeAgent via mandatory narrative injection) now see the rules.
P13c — /web-advice fallback when ChatGPT Web / Grok Web are unavailable
The /web-advice skill assumes multi-model access via aside-mcp or
browser tabs to ChatGPT Web + Grok Web + Gemini Web. On a Google-only
Aside account, the model list returns "not available" for all three
official Web LLM providers:
$ aside exec -m openai/gpt-5 "test"
Error: Requested model openai/gpt-5 is not available for this account.
$ aside exec -m anthropic/claude-sonnet-4-5 "test"
Error: Requested model anthropic/claude-sonnet-4-5 is not available for this account.
$ aside exec -m google/gemini-2.5-pro "test"
Error: Requested model google/gemini-2.5-pro is not available for this account.
The fallback when this happens:
Use codex exec (CLI) as the recorded multi-model second-opinion.
It runs on gpt-5.3-codex-spark by default and has its own quota.
Build the prompt + diff, save to evidence/<pr>-prompt.md and
evidence/<pr>.diff, pass via codex exec --cd <worktree> -m <model>.
Post the synthesis as a PR comment with a transparency note
explaining which models were queried. The Green Gate accept
criterion is "independent adversarial review recorded on the PR",
not "specifically ChatGPT Web + Grok Web".
In the Slack reply, state the constraint explicitly: "ChatGPT Web
Grok Web tabs were not queryable from this account (Aside-side
model list does not include them on the active Google-only
account). Codex is the multi-model second-opinion of record."
The Codex-as-second-opinion pattern is durable across the broader PR
review workflow — it's not specific to P13. Documented here so that
when a future session hits the same "Aside model not available" wall,
the fallback is one tool call away.
Verified 2026-08-01 on PR #8661: Codex review found the same DialogAgent
coverage gap that /advice Reviewer C found (independent convergence =
strong signal). Pre-fix verdict: CHANGES REQUESTED. Post-fix verdict:
APPROVE. The Codex approval was the deciding factor for shipping.
P13d — The "iterate until /advice and /web-advice approve" user directive
When the user types "keep iterating until PR comments are handled and
/advice and /web-advice approve" (or equivalent), this is a
completion criterion, not a vibe. The work is not done until:
Both /advice (3 reviewers) /web-advice (Codex) return APPROVED.
The diff passes the 4-lane (/document-standards) check.
Tests pass.
The PR comment is posted with the full per-comment fix map.
Force-push is approved + executed.
The 20-min babysit cron is scheduled.
If any of those gate-criteria fails, fix the underlying issue and
re-run the relevant reviewer. Don't ask the user "should I keep
iterating?" — the directive is the user explicitly saying "iterate
until X is true." The right move is to keep working until X is true
or until a hard block surfaces (missing credentials, ambiguous
preference, etc.).
Verified 2026-08-01 on PR #8661: the user-typed phrase "Keep iterating
until PR comments are handled and /advice and /web-advice approve"
drove 3 rounds of reviewer fan-out (initial → fix DialogAgent gap →
re-run) before the convergence signal was strong enough to ship. The
post-fix /advice Reviewer C re-run and the /web-advice Codex re-run
both returned APPROVED, which is the ship criteria.
P13 checklist — when the user types "iterate until reviews approve"
When the user types "iterate until /advice + /web-advice approve" (or
the PR has multiple review items from a destructive cleanup):
List open review items via gh api .../issues/<N>/comments +
reviews(last:5) + the latestOpinion GraphQL thread query. Mark
each item as live vs. moot.
For live items, identify the minimal fix surface. Generic
rules → shared/ contracts via {{PROMPT_INCLUDE:}}. Campaign-
specific → delete. Duplicated localization → extract to one shared.
Verify the existing tests pass on the rebased branch
(cd <worktree> && $HOME/projects/your-project.com/venv/bin/python -m unittest <module>).
Run /advice (3 reviewers) + /web-advice (Codex) in parallel.
Use delegate_task for the /advice fan-out. Use codex exec for
the /web-advice fallback when ChatGPT/Grok Web tabs are unavailable.
Convergence signal: 2/3 /advice reviewers + Codex all agree on
a finding → fix it. Re-run only the disagreeing reviewers.
Apply the fix. Verify tests still pass. Force-push only after
the convergence signal is strong.
Post the PR comment with the full per-comment fix map + review
verdicts table. Use gh pr comment <N> --body "...".
Schedule a 20-min babysit cron targeting the same Slack channel
via hermes cron create "20m" --name "<name> status (20m)" --deliver 'slack:<channel>' --repeat 1. The cron will report CI status +
CodeRabbit review flip (or self-cancel if the PR is merged).
P7 — Refreshed patch supersedes a prior-session PR for the same topic
P7 — Refreshed patch supersedes a prior-session PR for the same topic
This is a recurring shape when the user re-runs the same workflow with a
refreshed patch (e.g. new patch supersedes the previous day's patch).
The verification case (2026-07-16, disk_magician): user sent the same patch
shape (sister files: <topic>.patch + <topic>-PROMPT.md) two days in a
row; the new instructions file literally said "REFRESHED 2026-07-16,
supersedes the 2026-07-15 patch."
Detection signals (Step 1 pre-flight):
A same-topic PR is OPEN from a prior session (gh api .../pulls?state=all&head=<OWNER>:<branch> returns a match)
gh api repos/<OWNER>/<REPO>/compare/<new-base-sha>...<prior-PR-head-sha> returns behind_by > 0 or status: diverged
The instructions file says "REFRESHED", "supersedes", "v2", or similar
What to do (no user confirmation needed — this is the natural end-state of "supersedes"):
Use a fresh branch name from current origin/main — e.g. <branch>-v2,
<branch>-refreshed-<date>. NEVER push onto the prior PR's branch head
(P1 applies: that branch is owned by the prior session's commit, and
git push --force onto it would rewrite history that other PRs may
reference).
Apply the refreshed patch on the fresh branch (the new patch's base SHA
should match current origin/main).
Open the NEW PR; in the body, reference the prior PR explicitly:
> **Note:** this PR **supersedes** [PR #N](...) — the <date> patch against base \`. PR #N is now commits behind main;
this refreshed patch (against base ``) is the new source
of truth.`
Close the prior PR with a comment that links to the new PR and quotes
the user's "supersedes" directive. Do NOT delete the prior branch —
leave it in place for diff inspection. gh pr close <N> --repo ... --comment "..."
In the Slack summary reply, name both PRs and explain the close as a
supersede (not a rejection). One-time 20m status cron targets the new
PR's URL.
Why this is non-destructive and safe to do without asking: closing a PR
with a supersede comment is a reversible action (reopen via gh pr reopen
in <2s). The user explicitly stated the new patch supersedes the old — the
"no confirmation gate" applies because the directive is unambiguous.
Failure modes to avoid:
❌ Pushing onto the prior PR's branch head (git push --force-with-lease origin HEAD:refs/heads/<prior-branch>) — rewrites shared history, breaks
reviewers' local checkouts, fails SOUL.md never-push-onto-someone-elses-pr-head
❌ Asking the user "Want me to close the prior PR?" mid-stream — the
supersede directive is the user's explicit instruction; this is the
finish-the-job end-state
❌ Deleting the prior branch — leaves no diff trail if the new PR is reverted
❌ Forgetting to mention the prior PR in the new PR body — CodeRabbit
reviewers will see two PRs for the same feature and ask which is canonical
Verified case
2026-07-15, jleechanorg/disk_magician PR #17 — Slack message "Handle" +
two attached files (0001-feat-safety-machine-loca.di[REDACTED_OPENAI_KEY]
41,016 B, and UPSTREAM-PROMPT.md 4,213 B / 71 lines). The instructions file
listed:
Target repo: jleechanorg/disk_magician
Branch: findings-wiki-contract
Base SHA: efc51ba (and git ls-remote confirmed HEAD was exactly efc51ba at apply time)
4 test suites + 1 lint to run
PR title + body template
"Do NOT merge — leave the PR for the repo owner to review"
End-state: PR #17 opened, all 4 test suites green, lint clean, base SHA
exactly matched. No existing PR for the topic. No push onto a non-owned
branch. Slack summary posted via xoxp fallback (bot token had not_in_channel).
Patch archived at a public gist via gh gist create (text-file shortcut,
v1.9.0).
2026-07-16, jleechanorg/disk_magician PR #21 (supersedes PR #17) — same
patch shape, refreshed version. Slack message body "Make sure these are
handled or do the extra work as needed. use /ms and search slack history
might already be in progress" + 2 file attachments:
0001-feat-safety-machine-loca.di[REDACTED_OPENAI_KEY] (64,987 B
— refreshed, supersedes the 2026-07-15 patch)
UPSTREAM-PROMPT.md (188,464 B / 71 lines — explicit "REFRESHED 2026-07-16,
supersedes the 2026-07-15 patch" in opening paragraph)
Deltas from the 2026-07-15 patch:
Base SHA: b187d8e8 (was efc51ba); current origin/main HEAD matched
the new base exactly → git am applied cleanly with no conflicts.
Scope grew: 35 files / +1092/-33 (was 17 / +856/-4). New gates wired into
ALL cleanup scripts (was 2: cleanup_agent_artifacts + cleanup_worktrees).
New Section N: ~/Snapchat/Dev/.cache/bazel governance with pgrep+lsof
liveness gates. New CLAUDE.md sections.
4 test suites all green: test_safety_lib.sh PASS=31, test_cleanup_safety.sh
93/93, test_package_sync.sh PASS=16 ALL TESTS PASSED, findings_lint.sh --upstream clean.
Prior PR #17 detected as diverged, behind_by=14 against current main.
Followed P7 supersede pattern: created fresh branch findings-wiki-contract-v2
from b187d8e8, pushed 3d63d703, opened PR #21 (MERGEABLE), closed
PR #17 with supersede comment. Branch findings-wiki-contract left in
place for diff inspection (NOT deleted).
2026-07-21, jleechanorg/dark-factory PR #407 — follow-up commit on a
prior-session PR. Slack message body "lets delete skeptic gate and then
directly code this PR" + 3 file attachments (GUIDE.md,
INSTRUCTIONS.md, receipt-gate-reviewer.patch). Prior session had
shipped commit f461f93da and then the origin branch
receipt-gate-reviewer was deleted (no on-disk lease target for
--force-with-lease).
P8 applied (force-push after branch deletion): reconstructed from
refs/pull/407/head, added 3rd commit b04df6f44 deleting
.github/workflows/skeptic-gate*.yml + .github/actions/skeptic-setup/
6 adversarial tests + evidence-gate.yml "Skeptic PASS" line, pushed
with plain --force. gh pr view 407 --json headRefOid confirmed
PR head moved from f461f93 to b04df6f44.
P9 applied (two-tier skeptic-gate distinction): user said "delete
skeptic gate" — workflow-level gate (CI blocker) deleted; pipeline-level
_gate_skeptic DOT node type kept (used by level5 pipelines + tests).
P10 applied (pre-existing failure attribution): observed 4 test
failures after deletion; stash/reset/restore to origin/main confirmed
all 4 byte-identical pre-existing (3× git-lfs PATH drift,
1× macOS /bin/false not-found). PR summary cited zero new failures.
End-state: PR #407 +331/-863 across 8 files, all remaining CI checks
green (test, daemon-tests, Evidence Gate, CodeRabbit SUCCESS, Bugbot
NEUTRAL). Skeptic no longer in statusCheckRollup (gate deleted).
Bead $USER-pm8f closed as resolved-via-deletion.
No patch attached. If the user said "fix this bug" with no diff, you
need always-pr-never-local-edit + AO dispatch — that's a fix-X task, not
an apply-supplied-patch task.
No instructions file. If the user just sent a patch with no
structured instructions, the workflow is much more flexible — you have to
figure out the target repo, branch name, test gates yourself. That's
closer to github-pr-workflow with an apply step added.
User said "apply this locally, don't push yet." This skill assumes the
user wants the PR opened. For "stage this and show me" tasks, do just
Steps 1-5 then report.
User gave you destructive ops (force-push, branch delete on protected
branch, etc.). STOP. Per SOUL.md merge discipline + commit/push gate,
destructive operations require explicit user approval in the current thread
(MERGE APPROVED-style signal).
Support files
scripts/apply_patch_and_open_pr.sh — end-to-end shell recipe for the
pre-flight + clone + branch + git am + test + push + pr-create sequence.
Re-runnable; takes <OWNER>/<REPO><branch><patch><instructions>
as positional args.
references/pre-existing-failure-attribution.md — the 5-step
stash/reset/restore protocol for machine-checking that a test failure
is pre-existing on origin/main (companion to P10; used to satisfy
SOUL.md same-test-name-rule + qa-test-failure-dismissal-anti-pattern).
references/upstream-prompt-format.md — the canonical structure of an
UPSTREAM-PROMPT.md / RUNBOOK.md file (what sections to expect, what
to extract from each).
references/refreshed-patch-supersede-pattern.md — the recurring
"refreshed patch supersedes prior-session PR" shape (Pitfall P7); how to
detect it, what to do, and why this is durable across repos.
