Rewrite Architect — Compatibility Bar, Then Design

Phase 2. First decide what must stay compatible (ADR-000), then design the Elixir system that meets the goals inside that bar. The bar comes first because "idiomatic" and "compatible" pull in different directions, and only the user can trade them.

/rewrite:analyze → /rewrite:architect  <-- YOU ARE HERE
                         ↓
                   /rewrite:roadmap → /phx:plan → /phx:work → /rewrite:verify

Iron Laws

1. ADR-000 BEFORE ANY DESIGN — never spawn the architect without an accepted or ASSUMED bar
2. THE USER OWNS TRADE-OFFS — cutover mode, forced logouts, schema changes and UI changes are asked, never inferred
3. CONSULT BY QUESTION — spawn a phx specialist only for a named open question, with an evidence packet
4. LEDGER MUST CHECK CLEAN — run ledger.py check and fix or report every error before presenting
5. ASSUMED IS NOT APPROVED — --auto decisions stay ASSUMED until the user accepts them
6. SPAWN FROM THIS SESSION — agents never spawn phx orchestrators (nesting depth)

Workflow

1. Preconditions

Require .claude/rewrite/analysis/{features,mechanisms,runtime,contracts}.md; if any is missing, stop and suggest /rewrite:analyze. Read the paths from status.md, stop if the cwd is not Target path, and check that the source is unchanged:

SOURCE=$(sed -n 's/^Source path: //p' .claude/rewrite/status.md)
TARGET=$(sed -n 's/^Target path: //p' .claude/rewrite/status.md); [ "$(pwd -P)" = "$TARGET" ]
python3 "${CLAUDE_SKILL_DIR}/../analyze/scripts/inventory.py" "$SOURCE" \
  --compare .claude/rewrite/inventory/inventory.json

Exit 1 from --compare means stale analysis: tell the user and offer to re-run /rewrite:analyze first. Resume: when ADR-000 exists but phase is architecting or architecture.md, decisions.md or divergences.md is missing, keep ADR-000 and consults/ and continue at step 4 (step 5 when all four exist and only ledger.py check failed).

2. Compatibility bar → ADR-000

Read ${CLAUDE_SKILL_DIR}/references/compatibility-bar.md for the question set, the defaults and the ADR-000 template. Ask with AskUserQuestion in two rounds, at most 4 questions each:

1. Cutover mode · Data · Identity · Clients/UI
2. Contracts (the Decisions Needed for ADR-000 list from contracts.md; default by class: keep contract, relax incidental, decide each defect) · Performance goal (numeric, against a measured legacy baseline) · Steering conflicts from charter.md

With --auto, apply the per-class defaults from the reference, set Status ASSUMED, and list every assumption at the top of the ADR. A security defect is never kept by default.

Write .claude/rewrite/architecture/ADR-000-compatibility-bar.md. When the mode is staged cutover or coexistence, ADR-000 also names the transition concerns the roadmap must cover per slice (write ownership, schema authority, identity, jobs, data movement, cutover, rollback).

3. Consults (only for open questions)

Read the Questions sections of the four analyses. Map each design-relevant question to at most one specialist, in parallel and at most 4 at a time:

Open question about	Spawn
Legacy schema mapping, transactions, constraints	phx:ecto-schema-designer
Runtime topology, processes, supervision	phx:otp-advisor
Durable work, legacy queue transition	phx:oban-specialist
UI redesign in LiveView (only if the bar allows it)	phx:liveview-architect
Legacy auth, crypto, session handoff	phx:security-analyzer
A library not in kb/	phx:hex-library-researcher

Each prompt states the question, the evidence paths to read (analysis sections, kb/ files, SOURCE anchors) and the output file .claude/rewrite/architecture/consults/{agent}.md. It says: "You are a consult: answer the question, do not plan or implement. Write the output file with your current answer before further research, and update it as you go." No open questions means no consults.

Stay in this turn until every consult reports (see analyze step 6). Then check each file exists; for a missing one, resume the agent once with SendMessage (or re-spawn it once), and record a still-missing consult as a gap in status.md.

4. Synthesize

Set status.md phase architecting first, so a cut-off run is detectable. Spawn rewrite:elixir-way-architect (opus) in the FOREGROUND (run_in_background: false): it is one long agent with nothing to run in parallel, and headless claude -p stops background agents 10 minutes after the last turn. Pass CHARTER, BAR, ANALYSIS, CONSULTS, KB=.claude/rewrite/kb/, PACKS (each pack in inventory.json stacks), TARGET, OUT=.claude/rewrite/architecture/, and RESUME (the architecture files already present, or none). It writes architecture.md and ledger.md first, then the ADRs, then divergences.md and decisions.md. If it returns BLOCKED, fix the precondition it names.

5. Check the ledger

python3 "${CLAUDE_SKILL_DIR}/scripts/ledger.py" check --rewrite-dir .claude/rewrite
python3 "${CLAUDE_SKILL_DIR}/scripts/ledger.py" stats --rewrite-dir .claude/rewrite

check counts an analysis ID only from a row's ID cell (one row per alias), so fix an unaccounted ID with a row, e.g. a DELETE row whose purpose reads "absent in legacy", never with prose. Rows need every header cell and Departure none; an AR-NN departure needs the user's approved record in verify/accepted-risks.md, which neither you nor the architect writes. On errors (unaccounted or duplicate IDs, short rows, invalid verdicts, missing ADR files), send the error list back to the architect once via SendMessage (or re-spawn it with the list); if errors remain, show them to the user instead of hiding them and keep phase architecting. Only after check passes and decisions.md and divergences.md exist, set status.md phase architected (architected (ASSUMED) under --auto).

6. Present for approval

Show, from the architect's return and architecture.md: the context map, the process table, verdict counts (UNDECIDED and what they block), the three decisions that most need review, and pending divergences. Ask:

• Accept — set ADR Status accepted and divergences approved
• Revise — name an ADR or area; re-run step 4 with the feedback
• Stop — keep everything as proposed

--auto: leave Status ASSUMED, do not mark divergences approved.

--revise ADR-NNN: re-open one decision, re-run steps 3–5 for it only, and mark ledger rows that cite it with evidence stale.

Next step: /rewrite:roadmap.

Arguments

• --auto — conservative per-class defaults for every bar question, marked ASSUMED
• --revise ADR-NNN — re-open one decision

References

• ${CLAUDE_SKILL_DIR}/references/compatibility-bar.md — questions, defaults, ADR-000 and transition template
• ${CLAUDE_SKILL_DIR}/scripts/ledger.py — check and stats