Run the DESIGN phase. The design-author decides the approach — recording\nevery self-resolved choice as an auditable assumption — and the\nadversarial design review gates advancement. No mid-run prompt fires.
\n$ARGUMENTS is the artifact directory: docs/plans/<id>/. If empty, the\ndiscovery block below resolves it.
The design-author reads:
$ARGUMENTS/task.md — what we are building (intent)$ARGUMENTS/questions.md — the questions that drove research$ARGUMENTS/research.md — what exists (facts)Resolve the artifact directory by running this self-contained block (one bash\ncall — agent threads reset cwd between calls):
\n# Three-tier artifact-directory discovery (archetype A).\n# ID_RE + PHASE_FILES canonical from hooks/session-start-recover.mjs.\n# PHASE_FILES recency mirrors findActiveTopic() in session-start-recover.mjs.\n# NOTE: this block is duplicated across 8 skills by design (see docs/architecture.md); future: shared discover-topic.sh.\nID_RE='^([A-Za-z][A-Za-z0-9_]*-[0-9]+|[0-9]{4}-[0-9]{2}-[0-9]{2})-[a-z0-9][a-z0-9-]*$'\nPHASE_FILES=\"task questions research design structure plan\"\nPRED=\"research.md\" # predecessor artifact this skill consumes\n# Tier 1 — explicit: $ARGUMENTS names an existing dir → use verbatim.\nif [ -n \"$ARGUMENTS\" ] && [ -d \"$ARGUMENTS\" ]; then\n echo \"$ARGUMENTS\"; exit 0\nfi\n# Tier 2 — discover: newest ID_RE dir under docs/plans/ that holds PRED.\nbest=\"\"; best_mtime=-1\n# Assumes cwd is the repo/worktree root (where docs/plans/ lives).\nfor dir in docs/plans/*/; do\n name=\"$(basename \"$dir\")\"\n printf '%s' \"$name\" | grep -qE \"$ID_RE\" || continue # ID_RE filter\n [ -f \"$dir$PRED\" ] || continue # predecessor filter\n m=-1\n for p in $PHASE_FILES; do\n f=\"$dir$p.md\"\n [ -f \"$f\" ] || continue # skip racing/absent\n s=\"$(stat -f %m \"$f\" 2>/dev/null || stat -c %Y \"$f\" 2>/dev/null)\" || continue\n [ \"${s:-0}\" -gt \"$m\" ] && m=\"$s\" # max-mtime over PHASE_FILES\n done\n [ \"$m\" -gt \"$best_mtime\" ] && { best_mtime=\"$m\"; best=\"$dir\"; }\ndone\n[ -n \"$best\" ] && { echo \"$best\"; exit 0; }\n# Tier 3 — none found: print nothing → fall to AskUserQuestion (prose below).\n$ARGUMENTS for the rest of this\nskill (tier 1 explicit arg, or tier 2 discovery). When the path came from\ntier 2 (no explicit arg), announce the resolved directory to the user before\nproceeding, so an auto-picked topic is never silent.research.md),\ndo not hard-error. Fire AskUserQuestion with a Setup header and labeled\noptions:\n/team-research docs/plans/<id>/ to produce the\nmissing research.md.docs/plans/<id>/ directory\ndirectly (run ls docs/plans/ to find your topic directory).\n\nFollow
\nskills/progress-tracking/SKILL.md: when this procedure has two or more steps, seed one todo item per step before starting and mark each complete as you go.
Use the directory resolved in ## Input.
Dispatch design-author, which:
## Decisions made marked as an assumption (see the agent file)$ARGUMENTS/design.md with frontmatter revision: 0If $ARGUMENTS/design.md already exists, skip this dispatch and\nresume at step 3 — never re-draft an existing design.
Design review gate. If the latest\n$ARGUMENTS/design-review-<n>.md already carries a passing verdict\n(APPROVE or COMMENT), skip straight to step 4 — never re-review a\npassed design. Otherwise, before each review dispatch, run the\nexternal cross-model pass by following ## Design-review pass in\nskills/cross-model-review/SKILL.md — reference that procedure,\nnever duplicate it here. Its one gate: the\nTEAM_DISABLE_CROSS_MODEL kill-switch. Run the runner's detect\nverb, then run per ready CLI — each through its own named courier\nsub-agent per that skill's vendor-courier block, with its inline\nfallback — naming any unavailable CLI to the\nuser per that skill's ## When a vendor CLI is unavailable; a\nmissing runner is\nskip: cross-model runner not found per CLI. Fence each CLI's raw\noutput as a DATA block at capture time (fence longer than any\nbacktick run in the output, per that section), append one\n## External review input section — opening with the\nuntrusted-content line that section specifies — holding the fenced\nblocks to the review brief, and append the round's transcript to\n$ARGUMENTS/cross-model-raw.md in the result-line format that\nsection pins (created on first use; a zero-call round appends\nnothing). Any skip continues with the\nreviewer alone — the pass never blocks the gate. Then dispatch the\nadversarial design review (the\n## Review brief in skills/eng-design-doc-review/SKILL.md, run by\na fresh-context read-only Explore subagent each round) and write\nthe findings + verdict to $ARGUMENTS/design-review-<n>.md, where\n<n> is the highest existing <n> + 1 (1 when none exists) — never\noverwrite an earlier verdict record. Derive the verdict:\nfrontmatter from the last verdict token in the report body — the\nreviewer's verdict is the terminal line of its report. When the\nreport contains a ### Cross-model disposition section, append that\nsection as one block to $ARGUMENTS/cross-model-notes.md,\nblockquote-wrapped — prefix every line with > at append time, per\nthe design-review gate in skills/team/SKILL.md — opening with the\norchestrator-authored label\nline — the literal > **Design round <n>** — prepended inside the\nwrap; same frontmatter-on-first-append rules as the other gates\n(schema in skills/artifact-frontmatter/SKILL.md). Then act on the\nverdict:
design-author with the\nreviewer's findings verbatim. The agent re-drafts and increments\nrevision: <n+1>, then a fresh review round runs. Cap at\nrevision: 5; at cap, halt terminally and report the unresolved\nfindings. Recovery: a human revises $ARGUMENTS/design.md by\nhand and re-invokes /team-design bare — the run resumes at this\ngate. The revision counter persists in design.md frontmatter;\nhand-lower it to restore the revision budget.Stop once $ARGUMENTS/design.md exists and the latest\n$ARGUMENTS/design-review-<n>.md verdict is APPROVE or COMMENT.
Report design path and tell the user:\n\"Next: run /team-structure docs/plans/<id>/\"