| name | adr |
| description | Anton ADR lifecycle — author new architectural decision records, list existing ones by status or affects-category, and mark old decisions superseded. Use when capturing a decision (especially after `cluster-intake-gatekeeper` returns an ADD/DEFER/REJECT verdict), when reviewing prior decisions before changing direction, when checking if a candidate component has been removed before, or when promoting a decision out of memory into a durable record. ADRs live in `context/adrs/` and are immutable — supersession is the only way to change a decision. The ADR index is built by scanning ADR files directly and injected into every Claude Code session by `.claude/hooks/inject_adr_index.py`. Keywords — ADR, architecture decision record, decision log, supersede, decision history, why did we, prior decision, recorded decision, MADR, immutable, intake handoff, cluster-intake-gatekeeper handoff, removal graveyard, reverted decision. |
| allowed-tools | Read, Write, Edit, Glob, Grep, Bash |
ADR
Anton-local skill for the architectural decision record lifecycle. Owns three workflows: new, list, supersede. ADRs live at context/adrs/NNNN-kebab-slug.md and are immutable — once accepted, the body never changes. Supersession is the only way to revise a decision.
Why this skill exists
Anton needs one durable home for architectural decisions so that:
cluster-intake-gatekeeper has somewhere to hand off its ADD/DEFER/REJECT verdicts (instead of stopping at "ADR-ready").
- Future Claude Code sessions can discover prior decisions automatically via the SessionStart hook (
.claude/hooks/inject_adr_index.py) without the operator typing them in.
cluster-intake Step 2 can query past Reverted decisions by category instead of grepping a single shared file.
- The operator (and future-Will) can recover why the cluster looks the way it does without mining
git log.
The skill is anton-local and has no runtime dependency on ~/.claude/skills/context-repo-adr — the format is intentionally compatible with the global context-repo conventions, but invoking this skill never delegates to that one.
Hard rules
- ADRs are immutable. Once an ADR's
status: is Accepted (or Reverted / Rejected / Deferred), the body never changes. The only edit allowed on an existing ADR is flipping the status: line to Superseded-by NNNN. Any other change must go through supersede.
- Numbers are never reused. The next number is always
max(existing) + 1, zero-padded to four digits, regardless of whether earlier numbers were superseded.
Status: Reverted is anton-specific. Use it for any decision that was once live and is now removed. Standard MADR has no equivalent.
References:
Workflow — new
Triggered by:
- The user says "create an ADR" / "record this decision" /
/adr new "<title>"
cluster-intake-gatekeeper hands off an ADD / DEFER / REJECT verdict (the verdict object substitutes for the interactive interview — see "Handoff from intake" below)
Steps
-
Allocate the next number. Glob('context/adrs/0[0-9][0-9][0-9]-*.md'). Sort. The next number is max + 1, zero-padded to 4 digits. Never reuse a number.
-
Generate the slug. Lowercase the title, kebab-case it, drop articles and prepositions, take the first 4–6 meaningful words. Example: "Adopt Cilium over Calico for CNI" → adopt-cilium-over-calico-for-cni → file 0018-adopt-cilium-over-calico-for-cni.md.
-
Interview the user (skip this when invoked from intake — see Handoff from intake). Ask only the questions whose answers aren't already known:
- Title (one short sentence)
- One-sentence summary (the blockquote after the H1)
status — usually Accepted for forward-looking decisions; Deferred / Rejected / Reverted if known
affects — pick the broadest applicable category from: storage, observability, databases, registries, demos, networking, security, compute, all, or a new category if none fit
intent — concrete-need / learning / unknown
supersedes — list of NNNNs being superseded by this one (usually empty)
- Context (1–3 paragraphs — what situation prompts this decision, what forces are in play)
- Decision (active voice, one or two sentences — "We will adopt X" / "We will not")
- Alternatives considered (do-nothing + 1–2 alternatives, with a one-line "why rejected" each)
- Consequences — accepted costs, lessons (if
Reverted), follow-ups
-
Write the ADR file at context/adrs/NNNN-slug.md using the shape from template. Frontmatter is mandatory; missing frontmatter fields default to the template's defaults.
-
If supersedes: is non-empty, also flip the line of each superseded ADR to . This is the allowed in-place edit on an existing ADR. Use with a precise old/new string for the status line only — never touch the body.
Handoff from intake
When cluster-intake-gatekeeper invokes this skill at the end of its evaluation, it passes the verdict object directly. Map the verdict to ADR fields without re-interviewing:
| Gatekeeper verdict | ADR status: | Notes |
|---|
| Add (concrete need) | Accepted | intent: concrete-need. Decision section names the hand-off skill (add-flux-app / flux-app-author). Consequences section includes the Renovate-PR tax line. |
| Add (learning) | Accepted | intent: learning. Consequences section must include the timebox, exit plan, and review date — these are load-bearing under the contained-learning rubric. The frontmatter should add a review-by: field with the review date. |
| Defer | Deferred | Body holds the exact measurable unblock conditions (what would have to be true to convert to Accepted). |
| Reject | Rejected | Body holds the failed gate (1–7) or known-bad pattern, and — if appropriate — whether reframing as a learning intake would change the answer. |
After writing the ADR, return to the gatekeeper with the ADR number so it can include it in its memory append.
Workflow — list
Triggered by:
/adr list or /adr list --status Reverted or /adr list --affects databases
cluster-intake Step 2 (the removal-detection lookup) — internally invokes this with --status Reverted --affects <candidate-category> to find prior reverts
Steps
Glob('context/adrs/0[0-9][0-9][0-9]-*.md').
- For each file, read the frontmatter and extract
status, date, affects, intent, and the H1 title.
- Apply filters:
--status <X> → keep only ADRs whose status: matches X (case-insensitive)
--affects <X> → keep only ADRs whose affects: contains X
- With no filter, return all
- Format as a markdown table sorted by NNNN (columns: #, Title, Status, Date, Affects, Intent).
- If used as the data source for
cluster-intake Step 2: also read the matching ADRs' "Re-adoption guidance" section so the intake can surface the conditions under which a re-attempt would be valid.
Workflow — supersede
Triggered by:
/adr supersede <NNNN> "<reason>" — the user wants to revise an existing decision
Steps
- Verify the target exists.
Read context/adrs/NNNN-*.md. If missing, abort with a clear error.
- Verify the target's current status. If it is already
Superseded-by NNNN, abort — supersession chains exist (NNNN can be superseded by MMMM which is superseded by PPPP), so the user must supersede the latest in the chain.
- Author the replacement by invoking the
new workflow with supersedes: [<NNNN>] pre-populated. Walk the interview as normal.
- Flip the old ADR's status line — and only the status line — from
Accepted (or whatever it was) to Superseded-by <new NNNN>. Use Edit with the precise old/new strings. Do not touch the body, the date, or any other field.
- Confirm to the user with both ADR numbers (old → new) and the file paths.
What this skill does NOT do
- Does not write Flux manifests — that is
add-flux-app.
- Does not run cluster intake — that is
cluster-intake (which hands off here).
- Does not edit ADR bodies after the fact. Ever. Use
supersede.
- Does not auto-detect when a decision should be superseded. Manual via
/adr supersede.
- Does not push, commit, or tag — the user owns git operations.
- Does not delete ADRs. ADRs are append-only forever, even superseded ones.
Related skills and agents
cluster-intake / cluster-intake-gatekeeper — the dominant authoring path. Hand off here on every verdict.
add-flux-app / flux-app-author — invoked after an Accepted ADR that involves new manifests.
~/.claude/skills/context-repo-adr (user-level) — format-compatible but not invoked at runtime. If the user adopts the full context-repo convention later, the existing ADRs will Just Work.
Anti-patterns
- Editing an ADR body to reflect a changed decision. That's what
supersede is for. Editing in place destroys the historical record.
- Re-using a number after a delete. ADRs are never deleted, so this shouldn't come up — but if it does, allocate the next number, never the gap.
- Authoring an ADR before the decision is actually made. ADRs document decisions that happened, not aspirations. If the user is still considering, they're not ready for an ADR.
- Verbose ADRs. One screen of body is the target. If the decision genuinely needs more context, link to a doc — don't inline it.