diff --git a/.chezmoitemplates/agent-cheapest-pr-cold-reader.md.tmpl b/.chezmoitemplates/agent-cheapest-pr-cold-reader.md.tmpl
new file mode 100644
index 00000000..648644ab
--- /dev/null
+++ b/.chezmoitemplates/agent-cheapest-pr-cold-reader.md.tmpl
@@ -0,0 +1,47 @@
+You simulate a **cold reader** of a pull request: someone seeing this change for the
+first time, with none of the author's context. Your job is to report what such a reader
+can and cannot understand from the change alone.
+
+## What you can see
+
+- The **net diff** (`base...HEAD`), in one file whose absolute path is given in your prompt.
+
+## What you deliberately cannot see — do not assume or ask for it
+
+- The drafted PR description.
+- Commit messages.
+- Any session or chat history, or alternatives the author considered.
+
+Your only tool is `Read`, and you use it on that one diff file and nothing else. Read the
+file in pages (`offset` / `limit`) from the first line to the last. Do not open any other
+file, even one the diff names, imports, or links to. If something isn't derivable from the
+diff, that *is* your finding — mark it `unclear`, not a reason to go looking for more.
+
+## Produce
+
+1. **Coverage** — whether you read the whole file. If you could not, say at which line you
+ stopped and why.
+2. **Significance ranking** — the changes you'd call out as *key*, most significant first.
+ This is a fresh reader's view of what matters and becomes the candidate list for the
+ PR's Key Changes. Leave out trivial or mechanical edits.
+3. **Per key change**, report:
+ - **What** changed.
+ - **Why** you believe it was done — your inferred intent.
+ - **Confidence**: high / medium / low.
+ - **Clarity**: `clear` (intent is derivable from the change itself) or `unclear`
+ (you had to guess, or couldn't tell).
+4. **Could not determine** — anything whose purpose you genuinely can't infer from the diff.
+
+## Rules
+
+- Infer intent from the code, not from imagination. Don't invent a backstory.
+- Separate confident reads from guesses. A confident *wrong* read is worse than an honest
+ "unclear" — if you're guessing, say so.
+- Judge the change as it stands. Do not speculate about implementation order, rejected
+ approaches, or how the author got here.
+- You are not given a description and must not critique one. Produce an *independent*
+ impression, not a review of someone else's text.
+
+Return structured text the caller can parse: the coverage note, then the significance
+ranking, then one block per key change with the fields above, then the "could not determine"
+list.
diff --git a/.chezmoitemplates/agent-mid-tier-pr-cold-reader.md.tmpl b/.chezmoitemplates/agent-mid-tier-pr-cold-reader.md.tmpl
new file mode 100644
index 00000000..648644ab
--- /dev/null
+++ b/.chezmoitemplates/agent-mid-tier-pr-cold-reader.md.tmpl
@@ -0,0 +1,47 @@
+You simulate a **cold reader** of a pull request: someone seeing this change for the
+first time, with none of the author's context. Your job is to report what such a reader
+can and cannot understand from the change alone.
+
+## What you can see
+
+- The **net diff** (`base...HEAD`), in one file whose absolute path is given in your prompt.
+
+## What you deliberately cannot see — do not assume or ask for it
+
+- The drafted PR description.
+- Commit messages.
+- Any session or chat history, or alternatives the author considered.
+
+Your only tool is `Read`, and you use it on that one diff file and nothing else. Read the
+file in pages (`offset` / `limit`) from the first line to the last. Do not open any other
+file, even one the diff names, imports, or links to. If something isn't derivable from the
+diff, that *is* your finding — mark it `unclear`, not a reason to go looking for more.
+
+## Produce
+
+1. **Coverage** — whether you read the whole file. If you could not, say at which line you
+ stopped and why.
+2. **Significance ranking** — the changes you'd call out as *key*, most significant first.
+ This is a fresh reader's view of what matters and becomes the candidate list for the
+ PR's Key Changes. Leave out trivial or mechanical edits.
+3. **Per key change**, report:
+ - **What** changed.
+ - **Why** you believe it was done — your inferred intent.
+ - **Confidence**: high / medium / low.
+ - **Clarity**: `clear` (intent is derivable from the change itself) or `unclear`
+ (you had to guess, or couldn't tell).
+4. **Could not determine** — anything whose purpose you genuinely can't infer from the diff.
+
+## Rules
+
+- Infer intent from the code, not from imagination. Don't invent a backstory.
+- Separate confident reads from guesses. A confident *wrong* read is worse than an honest
+ "unclear" — if you're guessing, say so.
+- Judge the change as it stands. Do not speculate about implementation order, rejected
+ approaches, or how the author got here.
+- You are not given a description and must not critique one. Produce an *independent*
+ impression, not a review of someone else's text.
+
+Return structured text the caller can parse: the coverage note, then the significance
+ranking, then one block per key change with the fields above, then the "could not determine"
+list.
diff --git a/dot_agents/skills/creating-github-pull-requests/SKILL.md b/dot_agents/skills/creating-github-pull-requests/SKILL.md
deleted file mode 100644
index 3b9a5d0b..00000000
--- a/dot_agents/skills/creating-github-pull-requests/SKILL.md
+++ /dev/null
@@ -1,145 +0,0 @@
----
-name: creating-github-pull-requests
-description: "Create GitHub Pull Requests with consistent, high-quality titles and descriptions. Use when opening a PR on GitHub — covers title formatting, summary, decision log, and before/after comparison."
----
-
-# Creating GitHub Pull Requests
-
-Create Pull Requests on GitHub with a consistent structure that is easy to review and understand.
-
-## Gathering Context
-
-Before writing anything, collect the raw material:
-
-1. **Branch name** — extract any issue number (e.g., `42` from `feature/42-local-accounts`).
-2. **Commits** — `git log ..HEAD --oneline` to understand the full scope.
-3. **Diff against base** — `git diff ...HEAD` (or `--stat` first for an overview). For a stacked PR, `` is the parent PR branch, not the default branch. Read enough of the diff to understand what changed and why.
-4. **Related context** — conversation history, plan files, or linked issues that explain motivation.
-
-Do not rely solely on commit messages — they often lack the bigger picture. Read the actual changes.
-
-All PR content must describe the net change from the base branch to `HEAD`. Do not describe intermediate implementation steps, temporary refactors made during the session, or changes elsewhere in the worktree that are not part of the branch diff.
-
-## Title
-
-Format: `: `
-
-- **Type** — one of: `feat`, `fix`, `refactor`, `docs`, `test`, `chore`, `ci`, `perf`.
-- **Description** — a short sentence (lowercase, no trailing period) that captures the change from the reviewer's perspective. Aim for under 72 characters total.
-
-Examples:
-- `feat: add support for local user accounts`
-- `fix: prevent duplicate webhook deliveries on retry`
-- `refactor: extract payment validation into dedicated service`
-
-## Description Structure
-
-Start from [assets/pr-description-template.md](assets/pr-description-template.md). The PR body has three sections, plus an optional issue-closing line at the end. All sections use plain, direct language — no filler, no marketing.
-
-In the **generated PR description**, use the fixed section headings `📝 Summary` and `🔄 Before & After`. Keep `Key Decisions` plain; its subheadings carry decision-specific emoji cues.
-
-### Summary (required)
-
-3–5 lines maximum. Answer: *what changed and why does it matter?*
-
-Write for a reviewer who has not seen the ticket. State the problem or goal, then the approach taken. Do not list files or repeat commit messages.
-
-Describe only the final reviewer-visible outcome relative to the base branch.
-
-### Key Decisions (required)
-
-Each significant component or area that was added, changed, or removed gets its own `###` sub-heading under this section. Prefix the sub-heading with an emoji that signals that decision, such as `🗄️` for storage or `🧹` for cleanup. Skip the emoji rather than using a vague or misleading one. A short paragraph describes what changed and, when the rationale is not obvious from the issue alone, why.
-
-If a change isn't worth more than a single bullet point, it's not worth calling out here — the diff speaks for itself.
-
-Do not explain the order you made changes during the session unless that order is visible in the final branch state and matters to the reviewer.
-
-Include rationale when:
-- The PR makes internal refactors or restructuring choices that are self-contained (not driven by the ticket)
-- Non-obvious trade-offs or alternative approaches were considered
-- The change deviates from prior patterns or conventions
-- A reviewer would question the approach without explanation
-
-```
-## Key Decisions
-
-### ✅ Validation Service
-Added `validation.Service` to encapsulate all input validation rules, replacing the inline checks that lived in HTTP handlers. The inline approach had zero test coverage — extracting it makes the rules unit-testable without HTTP scaffolding.
-
-### 🔁 Webhook Idempotency
-Introduced an idempotency key on webhook delivery to prevent duplicates on retry. Chose optimistic locking over pessimistic since contention is rare and the retry cost is low.
-```
-
-### Before & After (required)
-
-A table comparing the previous state to the new state. This helps reviewers understand the net effect without reading every line of the diff.
-
-Each row covers one meaningful area of change. Keep cells concise — short phrases, not sentences.
-
-Every row must compare the base branch state to the final branch state. Never use intermediate states from the implementation process as the `Before` or `After` values.
-
-```
-## 🔄 Before & After
-
-| Area | Before | After |
-|------|--------|-------|
-| User creation | Only OAuth providers supported | Local username/password accounts available |
-| Validation | Inline in HTTP handler | Extracted to `validation.Service` |
-| Retry behavior | Webhooks re-sent on any failure | Idempotency key prevents duplicate delivery |
-```
-
-### Issue Link (conditional)
-
-If the branch name contains a GitHub issue number, add a closing keyword as the last visible line of the PR body. Reference-style link definitions may follow it. This auto-closes the issue when the PR merges.
-
-Always use `Closes`: `Closes #42`
-
-Omit this line entirely if there is no associated issue.
-
-## Links
-
-Keep full `https://...` URLs out of prose. Use reference-style links for external resources, then collect each definition once at the very bottom of the body, after the optional `Closes #` line. This keeps the review narrative concise while preserving clickable references.
-
-```
-See the [migration guide][migration-guide] before reviewing this change.
-
-Closes #42
-
-[migration-guide]: https://example.com/migration-guide
-```
-
-Use GitHub-native `#` references for other issues and PRs in the same repository rather than full URLs or reference-style definitions. Reference another PR only when it adds review context, especially for stack dependencies.
-
-## Stacked PRs
-
-A stacked PR branch is based on another open PR branch rather than the default branch. To describe only this PR's net change, compare it with its parent branch: `git diff ...HEAD`. Do not claim parent PR changes as this PR's own.
-
-Still target the repository's default branch unless instructed to target the parent branch. The parent-branch diff is for accurately describing the change, not necessarily for choosing the PR target.
-
-Declare stack dependencies in prose because GitHub has no stable, general PR-to-PR dependency mechanism. At the start of a dependent PR's Summary, name the parent with a native reference and state the merge order, for example: `Stacked on #123. Review and merge it first.` While the parent remains open, add a line naming its dependent follow-up PRs.
-
-## Creating the PR
-
-Use the GitHub CLI:
-
-```bash
-gh pr create --title "" --body "$(cat <<'EOF'
-
-EOF
-)"
-```
-
-- Target the repository's default branch unless instructed otherwise.
-- Do not add reviewers, labels, or milestones unless asked.
-- Create a draft PR only when explicitly asked.
-- If the branch has not been pushed yet, push it first with `git push -u origin HEAD`.
-
-## Anti-Patterns
-
-- **Restating commits** — the commit log is one click away; the description should add context the commits lack.
-- **Describing session history instead of branch diff** — PR text must reflect `base...HEAD`, not the sequence of edits made while implementing the change.
-- **Claiming a stacked PR's parent changes as your own** — diff against the parent branch, not the default branch, so the description covers only this PR's net change.
-- **Listing every file changed** — the diff view exists for this; focus on *what* and *why*, not *where*.
-- **Vague summaries** — "various improvements" or "refactor code" tells the reviewer nothing.
-- **Over-long descriptions** — if the summary exceeds 5 lines, it is doing too much. Push detail into Key Decisions or Before & After.
-- **Spraying PR references** — use `#` only where the related PR helps a reviewer understand the change or its dependencies.
diff --git a/dot_agents/skills/creating-pull-requests/SKILL.md b/dot_agents/skills/creating-pull-requests/SKILL.md
new file mode 100644
index 00000000..3d2f1ae3
--- /dev/null
+++ b/dot_agents/skills/creating-pull-requests/SKILL.md
@@ -0,0 +1,286 @@
+---
+name: creating-pull-requests
+description: "Create Pull Requests with consistent, high-quality titles and descriptions. Use when opening a PR or merge request — covers a cold read of the diff, title formatting, summary, decision log, before/after comparison, and stacked PRs. Supports GitHub and GitLab."
+---
+
+# Creating Pull Requests
+
+Create Pull Requests with a consistent structure that is easy to review and understand.
+
+The workflow, title format, description sections, and linking conventions in this file
+apply to every hosting platform. The parts that differ per platform — the concrete
+auto-link tokens, emoji syntax, description length limits, stacked-PR targeting, and the
+create command — live in a per-platform reference file.
+
+## Select the platform
+
+Read the reference file for the platform you are targeting **before assembling and
+creating the PR**. It carries the platform-specific syntax the rest of this skill defers
+to:
+
+- **GitHub** → [references/github.md](references/github.md)
+- **GitLab** → [references/gitlab.md](references/gitlab.md)
+
+If the platform is not obvious, infer it from the remote (for example, a `github.com`
+remote means GitHub and a `gitlab.com` remote means GitLab) or ask.
+
+## Gathering Context
+
+Before writing anything, collect the raw material:
+
+1. **Branch name** - extract the issue number if present (for example, `42` from `feature/42-local-accounts`).
+2. **Commits** - `git log ..HEAD --oneline` to understand the full scope.
+3. **Diff against base** - `git diff --no-ext-diff ...HEAD` (or add `--stat` first for an overview). `--no-ext-diff` keeps the output git's own raw diff regardless of any external diff tool configured on the machine. Read enough of the diff to understand what changed and why. For a stacked PR, `` is the parent PR's branch, not the trunk (see [Stacked PRs](#stacked-prs)).
+4. **Related context** - conversation history, plan files, or linked issues that explain motivation.
+
+Do not rely solely on commit messages. Read the actual changes.
+
+All PR content must describe the net change from the base branch to `HEAD`. Do not describe intermediate implementation steps, temporary refactors made during the session, or changes elsewhere in the worktree that are not part of the branch diff.
+
+## Cold-Read the Diff
+
+Before writing the description, get an unpoisoned read of the change. You carry the whole
+session in your head, so you cannot reliably tell what a reviewer will grasp from the diff
+alone and what they won't — everything looks obvious to someone who lived it. A dedicated
+cold reader can tell, because it sees only the change.
+
+1. **Stage the diff.** Create a fresh scratch folder with
+ `mktemp -d "${TMPDIR:-/tmp}/-.XXXXXX"` and write
+ `git diff --no-ext-diff ...HEAD` to `pr-cold-read.diff` inside it. `` is the
+ last segment of the `origin` remote URL without any trailing `.git` — not the checkout
+ folder's name, which in a worktree is the worktree's name. `` is the current branch
+ with every `/` replaced by `-`. The random suffix keeps concurrent sessions apart, even on
+ the same branch. `--no-ext-diff` bypasses any external diff tool configured on the machine
+ so the reader always gets git's own raw diff text, identical across machines. This file is
+ the one source of truth for both the readers and your own later judgment — do not regenerate it.
+ Delete the folder once the PR description is written.
+2. **Run the cheapest reader.** Spawn the `cheapest-pr-cold-reader` agent and give it **only
+ the absolute path** of the staged file, with the instruction to read it. Its only tool is
+ `Read`, and its instructions keep it on that one file, so the path is all it needs — do not
+ paste the diff into the prompt. It is also blind by design to author context — **NEVER**
+ give it the draft description, commit messages, any session context, or your own summary
+ of or hints about the change; those are the author's narration, and once it sees them it
+ can no longer tell you whether the *change itself* conveys the intent. It returns a
+ coverage note (whether it read the whole file), a significance ranking plus, per key
+ change, its inferred intent, a confidence level, and a `clear`/`unclear` verdict.
+3. **Escalate the residual.** Two things send work to the `mid-tier-pr-cold-reader` agent,
+ run on the same staged diff and equally blind:
+ - **Unclear items** — any change the cheapest reader marked `unclear`.
+ - **Size** — if the staged file is large enough to risk overflowing the cheapest reader's
+ context, route it to the mid tier *from the start*. Check the file's size (`wc -l`) up
+ front rather than spending a cheapest-tier read that cannot finish. Short of that, trust
+ the cheapest reader's coverage note: if it could not read the whole file, treat that
+ like `unclear`.
+
+ A change the mid tier now reads as `clear` was a *capability* gap, not a genuine one —
+ treat it as clear. A change it still marks `unclear` is a **real gap** — the trigger to
+ add rationale (below). If even the mid tier reports it could not read the whole file, the
+ diff is too large to cold-read holistically: say so and split the PR rather than describe
+ it from a partial read. Do not tell either reader about the other; each must form an independent
+ impression.
+
+Then let that read drive the description:
+
+- **Key Decisions come from the reader's significance ranking**, not your memory of what was
+ hard. A fresh reader's sense of what matters is the reviewer's sense; yours is skewed toward
+ whatever you wrestled with. Because the reader never saw the session, its ranking cannot
+ carry interim-decision noise — that noise exists only in your head.
+- **Add rationale where the read is `unclear`.** A real gap is the strongest signal a reviewer
+ needs the why, and you — having lived the change — can supply it. A `clear` change is still
+ surfaced if significant, but stated lean; it already carries its own why, so do not pad it.
+ **NEVER FABRICATE MOTIVATION.** If you genuinely don't know why a change was made either,
+ surface it plainly or flag it — do NOT invent a rationale to fill the gap. A made-up "why"
+ is worse than none: it misleads the reviewer with false confidence and is exactly the kind
+ of leak this whole step exists to prevent.
+- **Elevate only what is in the diff.** You may promote a change the reader underweighted when
+ its impact isn't visible in the diff (a one-line flag flip with outsized effect). You may
+ **not** introduce anything absent from the diff — an interim decision that lives only in your
+ memory is exactly the leak this step exists to prevent.
+- **Correct a confident-but-wrong read** against what you actually know.
+
+## Title
+
+Format: `: `
+
+- **Type** - one of: `feat`, `fix`, `refactor`, `docs`, `test`, `chore`, `ci`, `perf`.
+- **Description** - a short phrase (lowercase, no trailing period) that captures the change from the reviewer's perspective. Aim for under 72 characters total.
+
+Examples:
+- `feat: add support for local user accounts`
+- `fix: prevent duplicate webhook deliveries on retry`
+- `refactor: extract payment validation into dedicated service`
+
+## Description Structure
+
+Start from [assets/pr-description-template.md](assets/pr-description-template.md) and fill in the three sections. The template carries the section order, the section headings (`📝 Summary` and `🔄 Before & After` keep a fixed emoji; `## Key Decisions` stays plain and lets each decision carry its own — see below), the table shape, and a placeholder describing *what* belongs in each section — including the whole Summary, which needs no more guidance than its placeholder gives.
+
+Some platforms cap the description length or write emoji differently (see the platform
+reference). When there is a limit, keep the body concise enough to stay under it.
+
+What the placeholders can't carry is *judgment* — when something is worth calling out and how to reason about it. That is what the rest of this section adds.
+
+### Key Decisions
+
+The placeholder already asks for one `###` sub-heading per significant area, and notes that a change worth no more than a bullet doesn't belong here.
+
+Give each decision a *live* emoji cue. The section heading itself stays plain — `## Key Decisions`, no fixed emoji — because a stock emoji says nothing about *this* PR and can actively mislead: a 🔑 reads as *security keys*, which is noise on most PRs and destroys the signal on a PR that genuinely is about keys. The signal belongs one level down. Prefix each decision's `###` sub-heading with one or more emojis chosen for *that* decision — a globe 🌍 for timezone or locale work, a database 🗄️ for storage, a broom 🧹 for a cleanup — so a reviewer skimming the list gets the gist of each entry at a glance. One emoji is usually enough; add a second only when it genuinely sharpens the meaning. Skip the emoji rather than reach for a vague one — a generic ⚙️ on every heading is as meaningless as a stock 🔑.
+
+Do not explain the order you made changes during the session unless that order is visible in the final branch state and matters to the reviewer.
+
+The [cold read](#cold-read-the-diff) already located where rationale is needed — the changes it marked `unclear`. The cases below are the usual shapes those take:
+- The PR makes internal refactors or restructuring choices that are self-contained (not driven by the issue)
+- Non-obvious trade-offs or alternative approaches were considered
+- The change deviates from prior patterns or conventions
+- A reviewer would question the approach without explanation
+
+Example:
+
+```
+## Key Decisions
+
+### 🌍 Set-Timezone API
+Added an endpoint so a user can set their timezone explicitly, instead of the scheduler guessing it from the browser locale.
+
+### 🕰️🗄️ Timezone Data from a Compiled Database
+Compiled the IANA timezone database into the binary rather than reading the host's zoneinfo files, so lookups stay deterministic across deployment environments.
+```
+
+### Before & After
+
+The placeholder covers the table shape and cell style. Its purpose is to show the net effect — one row per meaningful area — so a reviewer grasps the change without reading every line of the diff. Every row compares the base branch state to the final branch state; never use intermediate states from the implementation process. Example:
+
+```
+## 🔄 Before & After
+
+| Area | Before | After |
+|------|--------|-------|
+| User creation | Only OAuth providers supported | Local username/password accounts available |
+| Validation | Inline in HTTP handler | Extracted to `validation.Service` |
+| Retry behavior | Webhooks re-sent on any failure | Idempotency key prevents duplicate delivery |
+```
+
+### Issue Link (conditional)
+
+If the branch name contains an issue number, add a closing line as the last *visible* line of the PR body: `Closes #42`. Both supported platforms recognize this keyword and close the issue when the PR merges. Reference-style link definitions may follow it (see [Referencing Issues & PRs](#referencing-issues--prs)).
+
+Always use `Closes`. Omit this line entirely if there is no associated issue.
+
+## Referencing Issues & PRs
+
+These conventions apply on every platform. The concrete syntax — which token auto-links
+an issue versus a PR, and how it renders — is in the platform reference file.
+
+The principle behind all of it: **keep full `https://…` URLs out of the prose.**
+Inline URLs bloat the source and hurt readability. So collect external URLs into a single
+definitions block at the very bottom of the body — each spelled out once — and keep the
+prose itself to short labels. Reference-style links are the mechanism for that.
+
+```
+See the [migration guide][migration-guide] before reviewing this change.
+
+Closes #42
+
+[migration-guide]: https://example.com/migration-guide
+```
+
+### Native issue and PR tokens
+
+Each PR host auto-links its own issues and pull requests from a short token (see the
+platform reference for the exact tokens). Use the native token for same-project issues and
+PRs rather than a full URL or a reference-style definition.
+
+- **Reference other PRs when they add context** — most often in [stacked PRs](#stacked-prs),
+ where the reviewer needs to know which PR this one builds on.
+- **Do not spray tokens.** Auto-linked references are convenient but easy to overuse; mention
+ a related issue or PR only where it helps a reviewer understand the change or its
+ dependencies.
+
+### The link-definitions block
+
+Collect **every** reference-style definition into a single block at the very bottom of the
+body, below the `Closes` line. Link definitions render as nothing, so the last *visible* line
+stays `Closes #42` while each URL is spelled out exactly once. One definition per label.
+
+## Stacked PRs
+
+A **stacked PR** is one whose branch is built directly on another PR's branch instead of on
+the trunk. That creates a genuine dependency: the stacked PR builds on code that is still
+under review, and it cannot merge until its parent does.
+
+Stacking is deliberate. A feature is split so an early, self-contained part can ship first —
+for reviewer load, shippability, or risk — while the parts that depend on it follow, as one
+or more further PRs stacked on top.
+
+**Determine the parent.** The parent is whatever branch the user names. If they don't say,
+reconstruct it from git locally. Git records nothing about what a branch was forked from, so
+this is a best-effort inference from commit topology — reliable for a real stack, but confirm
+with the user when it's ambiguous. Run the bundled helper (path relative to this skill) from
+inside the target repository so it reads that repo's branches:
+
+```bash
+bash scripts/nearest-parent-branch.sh
+```
+
+It ranks local branches by how few commits `HEAD` is ahead of each one's divergence point
+(merge-base) and prints the closest few as ` `, nearest first. The
+top line is the nearest parent. For the first PR in a stack this comes back as the trunk
+(nothing sits between it and `main`) — the "not stacked" case — and that still holds before
+you've committed anything, when `HEAD` is still on the trunk tip. If the top two lines share
+the same count the parent is genuinely ambiguous; confirm with the user rather than pick one.
+
+You need two things from the parent: its **branch name** (to diff against, and on some
+platforms to set the target at creation) and, to declare the dependency, its **PR number** —
+look that up by the parent's source branch through the platform CLI (see the platform
+reference).
+
+**Target the nearest parent, not the trunk.** The way to make a stacked PR *independently
+reviewable* — its diff showing only what it adds, not everything its parents also change — is
+to aim it at the branch directly below it in the stack. Given a stack A → B → C where each
+branch builds on the one before:
+
+- A is first in line and targets the trunk (`main`, or whatever the repo's default is).
+- B targets A's branch.
+- C targets B's branch.
+
+Aimed this way, each PR's platform diff is exactly `...HEAD` — its own net
+change. Aiming a stacked PR at the trunk instead would fold every ancestor's still-unmerged
+changes into its diff, which is what makes it unreviewable in isolation. The platform
+reference gives the flag that sets the target.
+
+**Describe only your own net change.** Read the parent range
+(`git diff --no-ext-diff ...HEAD`) and let Summary, Key Decisions, and Before
+& After cover only what *this* PR adds on top of its parent — never the parent's own work.
+
+**Declare the dependency** so a reviewer knows the stack exists and in what order it merges:
+
+- In the **dependent PR**, state the dependency at the **start of the Summary**, naming the
+ parent with the platform's native PR token — for example, "Stacked on #123 — review and
+ merge that first." Be explicit about merge order.
+- In the **parent PR**, name the dependents too **while it is still open**, so its reviewer
+ knows follow-ups build on it. Once the parent merges this note is moot.
+
+Some platforms also have a native dependency mechanism; see the platform reference.
+
+**Retargeting after a merge (out of scope).** When a PR targets its parent, whatever was
+targeting a merged PR must be retargeted at *its* target. In A → B → C, once A merges into
+the trunk, B must be retargeted from A to the trunk, while C still targets B. This skill
+covers only *creating* stacked PRs with the correct target; performing that rotation as PRs
+merge is left to you.
+
+## Creating the PR
+
+1. Assemble the body from `assets/pr-description-template.md` with the sections filled in and
+ the link-definitions block at the bottom.
+2. Follow the **Creating the PR** section of the platform reference for the exact create
+ command and any platform flags.
+
+## Anti-Patterns
+
+- **Restating commits** - the commit log is one click away; the description should add context the commits lack.
+- **Describing session history instead of branch diff** - PR text must reflect `base...HEAD`, not the sequence of edits made while implementing the change.
+- **Claiming a stacked PR's parent changes as your own** - diff against the parent branch, not the trunk, so the description covers only this PR's net change (see [Stacked PRs](#stacked-prs)).
+- **Listing every file changed** - the diff view exists for this; focus on *what* and *why*, not *where*.
+- **Vague summaries** - "various improvements" or "refactor code" tells the reviewer nothing.
+- **Ignoring a platform's length limit** - when the platform caps description length, trim detail before creating the PR; summarize and collapse lower-value decisions instead of overflowing.
+- **Over-long descriptions** - if the summary exceeds 5 lines, it is doing too much. Push detail into Key Decisions or Before & After.
+- **Spraying auto-links** - overusing native issue or PR references makes a PR unreadable for humans; follow the referencing conventions above.
diff --git a/dot_agents/skills/creating-github-pull-requests/assets/pr-description-template.md b/dot_agents/skills/creating-pull-requests/assets/pr-description-template.md
similarity index 66%
rename from dot_agents/skills/creating-github-pull-requests/assets/pr-description-template.md
rename to dot_agents/skills/creating-pull-requests/assets/pr-description-template.md
index 1e0b5034..e1417143 100644
--- a/dot_agents/skills/creating-github-pull-requests/assets/pr-description-template.md
+++ b/dot_agents/skills/creating-pull-requests/assets/pr-description-template.md
@@ -4,11 +4,11 @@
## Key Decisions
-
+
-###
+###
-
+
## 🔄 Before & After
@@ -18,5 +18,5 @@
| ------ | ---------------- | ------------ |
| | | |
-
+
diff --git a/dot_agents/skills/creating-pull-requests/references/github.md b/dot_agents/skills/creating-pull-requests/references/github.md
new file mode 100644
index 00000000..daceae73
--- /dev/null
+++ b/dot_agents/skills/creating-pull-requests/references/github.md
@@ -0,0 +1,81 @@
+# GitHub
+
+Platform-specific details for creating a PR on GitHub. The workflow, title format,
+description sections, and linking *conventions* live in `SKILL.md`; this file supplies the
+concrete syntax that only applies to GitHub.
+
+## Description length limit
+
+GitHub caps PR bodies at **65,536 characters**. A description that follows `SKILL.md`
+stays far below that, so keep the body concise for readability (the anti-patterns in
+`SKILL.md` still apply), not to fit a budget.
+
+## Auto-link tokens
+
+GitHub uses one token for both issues and pull requests in the same repository:
+
+- `#` → **issue or pull request**. GitHub renders it compactly as `#123` and shows the
+ title on hover, so it can be used on every mention without read fatigue. No bottom-block
+ definition is needed for it.
+
+To reference an issue or PR in a **different** repository, use the fully-qualified form
+`owner/repo#123`.
+
+```
+This is the second PR in the stack on top of #123, which extracted the
+validation service. It moves the webhook handler onto that service.
+
+Because the handler now delegates to the service introduced in #123, the
+inline retry logic could be deleted entirely.
+```
+
+## Closing keyword
+
+`Closes #42` closes the issue when the PR merges into the repository's default branch.
+GitHub does not close the issue when the PR merges into another branch, such as a stack
+parent.
+
+## Stacked PRs
+
+Applying the Stacked PRs section in `SKILL.md`:
+
+**Target the parent branch at creation.** Pass the parent's branch to `gh pr create` with
+`--base ` (`-B`) so the PR's diff shows only its own net change. The first PR
+in the stack omits this and takes the default branch as usual.
+
+When the parent merges and its head branch is deleted, GitHub retargets the dependent PRs to
+the parent's base automatically. If the parent's branch is kept, the dependent PR must be
+retargeted by hand (out of scope, see `SKILL.md`).
+
+**Closing keywords wait for the default branch.** GitHub ignores `Closes #42` while the PR
+targets a branch other than the default branch. Still add the line as `SKILL.md` describes,
+but do not promise in the description that merging this PR closes the issue.
+
+**Find the parent's PR number** to reference it with `#`:
+
+```bash
+gh pr list --head --state open --json number --jq '.[0].number'
+```
+
+**Declare the dependency in prose.** GitHub has no general PR-to-PR dependency mechanism, so
+follow the dependency rules in `SKILL.md` with the `#` token above, for example:
+"Stacked on #123 — review and merge that first."
+
+## Creating the PR
+
+Use the GitHub CLI:
+
+```bash
+gh pr create --title "" --body "$(cat <<'EOF'
+
+EOF
+)"
+```
+
+- Target the repository's default branch unless instructed otherwise or the PR is stacked;
+ pass `--base ` to override — a stacked PR targets its parent branch (see
+ [Stacked PRs](#stacked-prs)).
+- Do not pass `--fill` — it replaces the title and body with commit info.
+- Do not add reviewers, assignees, labels, or milestones unless asked.
+- Create a draft PR (`--draft`) only when explicitly asked.
+- If the branch has not been pushed yet, push it first with `git push -u origin HEAD`.
diff --git a/dot_agents/skills/creating-pull-requests/references/gitlab.md b/dot_agents/skills/creating-pull-requests/references/gitlab.md
new file mode 100644
index 00000000..a9f386ea
--- /dev/null
+++ b/dot_agents/skills/creating-pull-requests/references/gitlab.md
@@ -0,0 +1,109 @@
+# GitLab
+
+Platform-specific details for creating a PR on GitLab (`gitlab.com`). The workflow, title
+format, description sections, and linking *conventions* live in `SKILL.md`; this file
+supplies the concrete syntax that only applies to GitLab.
+
+## Terminology
+
+GitLab calls a pull request a **merge request** (MR), and the CLI is `glab mr`. Wherever
+`SKILL.md` says "PR" — including "Stacked PRs" — read it as "merge request". Every
+convention in `SKILL.md` applies unchanged; only the name differs.
+
+## Emoji
+
+GitLab renders **emoji shortcodes** — `:memo:`, `:arrows_counterclockwise:`,
+`:earth_africa:` — as native inline emoji, which display more consistently than a pasted
+Unicode character. So throughout the MR body, write **every** emoji as its shortcode rather
+than a literal Unicode glyph. This applies to both:
+
+- **The template's fixed section headings.** Use `## :memo: Summary` and
+ `## :arrows_counterclockwise: Before & After` in place of the literal `📝` / `🔄` that
+ `assets/pr-description-template.md` shows.
+- **The per-decision cues in Key Decisions.** Choose the emoji exactly as `SKILL.md`
+ describes, then write it as its shortcode — a globe cue becomes `:earth_africa:`, a
+ cleanup `:broom:`, a database `:card_file_box:`, and so on.
+
+## Description length limit
+
+GitLab caps MR descriptions at roughly **1,048,576 characters (1 MiB)** — effectively no
+limit for our purposes. You never need to trim for length. Keep the body concise for
+readability (the anti-patterns in `SKILL.md` still apply), not to fit a budget.
+
+## Auto-link tokens
+
+GitLab auto-links two reference tokens inside an MR body. Both render compactly — the
+visible text stays `#42` or `!15` and the title shows only on hover — so use the raw token on
+every mention. No bottom-block definition is needed for either.
+
+- `#` → **issue**.
+- `!` → **merge request**.
+
+```
+This is the second MR in the stack on top of !15, which extracted the
+validation service. It moves the webhook handler onto that service.
+
+Because the handler now delegates to the service introduced in !15, the
+inline retry logic could be deleted entirely.
+```
+
+To reference an issue or MR in a **different** project, use the fully-qualified form
+`group/project#42` or `group/project!15`.
+
+## Closing keyword
+
+`Closes #42` closes the issue when the MR merges into the project's default branch.
+
+## Stacked PRs
+
+GitLab has **no diff-scoping stacked-MR support** — an MR aimed at the trunk shows every
+ancestor's still-unmerged changes. Applying the Stacked PRs section in `SKILL.md`:
+
+**Target the parent branch at creation.** Pass the parent's branch with `-b `
+(`--target-branch`) so the MR's diff shows only its own net change. The first MR in the stack
+omits this and takes the default branch as usual.
+
+**Find the parent's MR number** to reference it with `!`. GitLab identifies an MR by its
+project-scoped `iid`; list the parent's source branch and read it:
+
+```bash
+glab mr list --source-branch -F json --jq '.[0].iid'
+```
+
+GitLab *also* has a native **merge request dependencies** feature (Premium and Ultimate
+tiers): an MR can be marked as depending on another, and GitLab then blocks merging until the
+dependency merges. This is a merge-order block, **not** diff scoping — it complements
+parent-targeting, it does not replace it. When the project's tier supports it, also set the
+dependent MR's "Merge request dependencies" field to the parent; there is no `glab` flag, so
+set it in the MR's UI after creation.
+
+Whether or not the native field is available, still **declare the dependency in prose** (see
+the Stacked PRs section in `SKILL.md`), linking the parent with the `!` token above, for
+example: "Stacked on !15 — review and merge that first."
+
+## Creating the PR
+
+Use the GitLab CLI:
+
+```bash
+BODY="$(cat <<'EOF'
+
+EOF
+)"
+
+glab mr create \
+ --title "" \
+ --description "$BODY" \
+ --yes
+```
+
+- `glab` auto-detects the project and the target branch from the git remote. It targets the
+ project's default branch unless instructed otherwise or the MR is stacked; pass
+ `-b ` to override — a stacked MR targets its parent branch (see [Stacked PRs](#stacked-prs)).
+- `--yes` skips the interactive confirmation. Do not pass `--fill` — it overwrites the title
+ and description with commit info.
+- Do not add assignees, reviewers, labels, milestones, or auto-merge unless asked.
+- If project detection fails, pass the project explicitly with `-R /`.
+- If the branch has not been pushed yet, push it first with `git push -u origin HEAD` (or
+ add `--push` to the create command).
+- Only mark the MR as a draft (`--draft`) if asked.
diff --git a/dot_agents/skills/creating-pull-requests/scripts/nearest-parent-branch.sh b/dot_agents/skills/creating-pull-requests/scripts/nearest-parent-branch.sh
new file mode 100755
index 00000000..d7cdc7bb
--- /dev/null
+++ b/dot_agents/skills/creating-pull-requests/scripts/nearest-parent-branch.sh
@@ -0,0 +1,27 @@
+#!/usr/bin/env bash
+#
+# Print the nearest parent branch of HEAD — the local branch the current branch
+# most likely built on — for inferring a stacked PR's parent when it isn't given
+# explicitly. Run it from inside the target repository.
+#
+# Git records nothing about what a branch was forked from, so this reconstructs
+# the relationship from commit topology: it ranks every local branch by how few
+# commits HEAD is ahead of that branch's divergence point (merge-base) and prints
+# the closest few, nearest first, as tab-separated "\t".
+# The top line is the best guess; equal top counts mean it is genuinely ambiguous.
+
+current=$(git branch --show-current)
+head_rev=$(git rev-parse HEAD)
+
+git for-each-ref --format='%(refname:short)' refs/heads/ | while read -r b; do
+ # Skip the current branch itself.
+ [ "$b" = "$current" ] && continue
+ # Skip strict descendants of HEAD (children further up the stack). A branch
+ # sitting exactly at HEAD is kept, so the trunk still wins before the current
+ # branch has any commits of its own.
+ if git merge-base --is-ancestor HEAD "$b" && [ "$(git rev-parse "$b")" != "$head_rev" ]; then
+ continue
+ fi
+ base=$(git merge-base HEAD "$b")
+ printf '%s\t%s\n' "$(git rev-list --count "$base..HEAD")" "$b"
+done | sort -n | head -3
diff --git a/dot_claude/agents/cheapest-pr-cold-reader.md.tmpl b/dot_claude/agents/cheapest-pr-cold-reader.md.tmpl
new file mode 100644
index 00000000..1b3aba21
--- /dev/null
+++ b/dot_claude/agents/cheapest-pr-cold-reader.md.tmpl
@@ -0,0 +1,9 @@
+---
+name: cheapest-pr-cold-reader
+description: "Use this PR cold-reader for work assigned to the cheapest model tier. It is intended for an independent read of a PR diff, with no access to the author's context."
+model: haiku
+color: cyan
+tools: Read
+---
+
+{{ template "agent-cheapest-pr-cold-reader.md.tmpl" . -}}
diff --git a/dot_claude/agents/mid-tier-pr-cold-reader.md.tmpl b/dot_claude/agents/mid-tier-pr-cold-reader.md.tmpl
new file mode 100644
index 00000000..240e4e77
--- /dev/null
+++ b/dot_claude/agents/mid-tier-pr-cold-reader.md.tmpl
@@ -0,0 +1,10 @@
+---
+name: mid-tier-pr-cold-reader
+description: "Use this PR cold-reader for work assigned to the mid-tier model tier. It is intended for an independent read of a PR diff, with no access to the author's context."
+model: sonnet
+effort: high
+color: blue
+tools: Read
+---
+
+{{ template "agent-mid-tier-pr-cold-reader.md.tmpl" . -}}
diff --git a/private_dot_config/opencode/agents/cheapest-pr-cold-reader.md.tmpl b/private_dot_config/opencode/agents/cheapest-pr-cold-reader.md.tmpl
new file mode 100644
index 00000000..9f9f215c
--- /dev/null
+++ b/private_dot_config/opencode/agents/cheapest-pr-cold-reader.md.tmpl
@@ -0,0 +1,17 @@
+---
+name: cheapest-pr-cold-reader
+description: Use this PR cold-reader for work assigned to the cheapest model tier. It is intended for an independent read of a PR diff, with no access to the author's context.
+mode: subagent
+model: openai/gpt-5.6-luna
+permission:
+ edit: deny
+ bash: deny
+ read: allow
+ webfetch: deny
+ task:
+ "*": deny
+ skill:
+ "*": deny
+---
+
+{{ template "agent-cheapest-pr-cold-reader.md.tmpl" . -}}
diff --git a/private_dot_config/opencode/agents/mid-tier-pr-cold-reader.md.tmpl b/private_dot_config/opencode/agents/mid-tier-pr-cold-reader.md.tmpl
new file mode 100644
index 00000000..655d28e2
--- /dev/null
+++ b/private_dot_config/opencode/agents/mid-tier-pr-cold-reader.md.tmpl
@@ -0,0 +1,18 @@
+---
+name: mid-tier-pr-cold-reader
+description: Use this PR cold-reader for work assigned to the mid-tier model tier. It is intended for an independent read of a PR diff, with no access to the author's context.
+mode: subagent
+model: openai/gpt-5.6-terra
+reasoningEffort: high
+permission:
+ edit: deny
+ bash: deny
+ read: allow
+ webfetch: deny
+ task:
+ "*": deny
+ skill:
+ "*": deny
+---
+
+{{ template "agent-mid-tier-pr-cold-reader.md.tmpl" . -}}