Getting it into your agent
One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.
npx agentmods add commands/dsswift/ion/aligngit clone --depth 1 https://github.com/dsswift/ionWhat it costs to keep this loaded
Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.
| Model | Per session | Once invoked |
|---|---|---|
| Fable 5 | $0.00039 | $0.20974 |
| Opus 5 | $0.00019 | $0.10487 |
| Sonnet 5 | $0.00008 | $0.04195 |
| Haiku 4.5 | $0.00004 | $0.02097 |
Grade A, and why
align scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.
A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.
Nothing flagged
None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.
How it starts
The opening of the file, as written. The whole thing — 1,030 lines — stays where its author put it; the contents beside it link to each section on GitHub.
You are running the /align command. This command operates in two modes depending on context. Detect the active mode first, then follow the instructions for that mode.
Invocation arguments. The raw arguments passed to this invocation (referred to as ARGS throughout this document) are:
$ARGUMENTS
If the block above is empty, the command was invoked with no arguments. Everywhere this document says "ARGS", it means the exact raw string above — the engine substitutes it here once, at this single point, and nowhere else.
Hard rules. These apply in both modes.
- Review the whole target; never ask the operator to narrow scope by size. The review surface is fixed by the mode and arguments, not by how large it is: Mode A audits the attached plan in full; Mode B (Local) reviews the entire
{base}...HEADdiff in full; Mode B (PR/branch) reviews the entire target diff in full. A large diff — any number of commits, files, or scopes — is reviewed completely; it is never a reason to stop and ask the operator which slice to review. The only scope narrowing that exists is explicit operator input parsed in B-Step 1 (a<focus>instruction, or a PR / branch target — explicitin PR/in branch, or a bare PR reference like287/#287/PR 287). Absent that input, there is no scope question. Do not emit anAskUserQuestion(or any prose prompt) asking the operator to pick a subset, confirm scope, or choose between "recent commits" and "whole branch" — proceed and review everything. A genuinely enormous diff yields a large report, not a smaller review. - You will not squash, re-cut, split, reorder, or force-push commits, and you will not push or open/modify a PR. The commit-rewrite lifecycle (
/squash) and the PR lifecycle (/create-pr) belong to the operator and are invoked when the operator decides. - Amending in place is the default delivery mechanism for a finding whose defect originated in a commit on the active branch — see B-Step 6. Amending a branch-local commit is not a commit rewrite in the
/squashsense: the commit count, order, and scope seams are unchanged; only the content of the commit that introduced the defect changes, so the branch's history reads as though the defect never shipped. This applies only to commits that originated on the active branch (present in{base}..HEAD, where{base}is the resolved review base — a worktree's source branch, elsemain) and have not been pushed. Never amend a commit that exists on{base}, onmain, on a remote, or in a PR's published history. - You will not run
gh pr create,gh pr merge,gh pr review,gh pr comment,git push,git push --force, or any other remote-mutating command.git commit --amendand agit rebase -ilimited toedit-ing branch-local commits are permitted only as the B-Step 6 amend mechanism described above; a rebase that reorders, squashes, drops, or rewords beyond the fix is not. - Committing is allowed — and only in Mode B, only after the operator approves the fix plan. When Mode B implements an approved fix plan (B-Step 6), it delivers the completed work into the branch's history: amended into the originating commit when the defect came from a commit on the active branch, or as a new conventional, correctly-scoped commit when it did not (see root
AGENTS.md§ "Commits"). In PR mode fixes always land as new commits on top of the PR's head branch in a dedicated worktree (see B-Step 6) — that history is published, so it is never amended; local commits on the PR branch are not "modifying the PR", and only pushing updates the PR, which stays the operator's. It never squashes commits together, never reorders or splits them, and never pushes — the operator handles squashing and PRs. Mode A never commits (no code exists yet — there is nothing to commit). During the review/plan phase of either mode (everything before an approved Mode B plan), nogit commithappens. - You review the content of the work, never the commit-shaping or PR lifecycle. You will not author findings, amendments, recommendations, plan steps, or open items that direct the operator to squash, split/re-cut/reorder commits, choose a merge strategy, or open/sequence a pull request. Commit-shaping (one-scope-per-commit, squash seams) is owned by
/squash; PR creation is owned by/create-pr. Both are the operator's lifecycle, invoked when the operator decides — running/alignnever implies a squash or a PR is the next step. You may mention in the report's prose that a follow-up squash or PR will eventually happen, but never as a finding, plan step, amendment, recommendation, or open item. Amending a fix into its originating branch-local commit is a delivery decision made in B-Step 6, not a commit-shaping finding — it never appears as a finding either. - Your analysis output is a single markdown report in this chat response. After the report, Mode A writes only the plan file (folding in the amendments); Mode B writes the plan file and then — once the operator approves — implements the fix plan, which edits source and lands the result (amended into the originating branch-local commits, or new commits). No other lifecycle action (squash, reorder, push, PR) is permitted in either mode.
- The report is followed by the mode's plan write — Mode A applies the amendments to the audited plan, Mode B authors a fix plan in planning mode. Mode A stops after the plan write and never implements. Mode B stops after authoring the plan and waits for operator approval; only after approval does it implement the fixes and commit them (B-Step 6).
What this file has done since we first saw it
Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.
- yesterday First seen · 1,030 lines · 39 tokens per session scan A f5b020af6612
align is a command published in the GitHub repository dsswift/ion (4 stars, last pushed yesterday), licensed MIT. It adds 39 tokens to every session and 20,974 once invoked, about $0.0002 per session on Opus 5. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.
Other commands, from other repositories
memory-self-review
Mine recent agent history (claude-mem + usage stats) for recurring failures and repeated patterns, audit MEMORY.md health, and PROPOSE (never apply) CLAUDE.md/memory edits for human approval.
code-review
Provide a code review for the given pull request.
echo
This is a placeholder command definition for plugin framework tests.
dashclaw-quality
Recurring find-and-fix quality pass over the whole DashClaw app — browser smoke (frontend-verify) + code gates → triage → parallel worktree fixes → verify → ship. The goal prompt that drives the dashclaw-find-and-fix workflow.
dashclaw-retire-legacy
Officially retire the published dashclaw/legacy Node SDK the safe way — deprecate now, delete at v5 — without breaking external npm consumers. Drives the legacy-sdk-deprecation-sweep workflow.
apply-self-review
Apply the approved improvements from the latest memory-self-review PROPOSAL. Auto-applies the safe, reversible doc/memory edits (with backups); queues code/config/rule-rewrite edits for an explicit go. Never commits or pushes.