Create Plan

Reading workflow files (TOC protocol)

When you Read any file under .claude/workflow/ or .claude/skills/, follow the protocol in conventions.md §1.8:

1. Read the TOC region: from <!--Document index start--> to <!--Document index end--> (read to the closing delimiter, not a fixed line count). If the file has no TOC region (a file whose only ## heading is this bootstrap block carries none, per §1.8(d)), read the file in full.
2. Match TOC rows where Roles contains any of your roles (or your role is any, or the row's Roles is any) AND Phases contains any of your phases (or your phase is any, or the row's Phases is any).
3. Use Read(offset, limit) to read only matched sections; if no row matches your role/phase, the file holds nothing for you — do not read further.

Your role: planner. Your phase: determined by the auto-resume State in workflow.md § Startup Protocol.

Inline refs you find inside workflow files carry the same name:roles:phases suffix; apply file-level filtering before opening: a ref matches when any of your roles is in its roles and any of your phases is in its phases, your own any on either axis matches every ref on that axis, and a ref whose own roles or phases is any matches you. Backtick-wrapped refs carry no suffix; open or skip them at your discretion.

Read and follow the workflow for Phase 0 (Research) and Phase 1 (Planning).

House style for chat-scale prose. User-facing prose produced from this file (status updates, escalation prompts, replanning summaries, review-mode loop turns, handoff notes, whichever apply) follows the AI-tell subset of house-style.md: ## Banned sentence patterns, ## Banned analysis patterns, ## Orientation, and ## Plain language. Structural rules (§ BLUF lead, § Structural rules for the ≤200-word section cap, § Document-shape rules (design / ADR-specific)) do not apply to chat-scale prose. See conventions.md:planner:0,1 §1.5 for the workflow-level anchor and tier mapping.

Stamp discipline. Every _workflow/** artifact this SKILL creates carries a line-1 <!-- workflow-sha: <40-char SHA> --> stamp written at creation. Direct-mutation kinds applied later by edit-design (content-edit, section-add, section-remove, section-rename, section-move, structural-rewrite, mechanics-edit, design-sync) leave the stamp untouched and preserve its line-1 position; only artifact creation, migration replay, and no-drift normalization write the stamp. The format definition, parser idioms, and the paired SHA-computation idiom this SKILL copies into its planning-transition step are anchored in conventions.md:planner:1 §1.6. Read that section for the single source of truth.

Step 1 — Read workflow documents.

Read these in order before doing anything else (do NOT ask the user anything yet):

1. .claude/workflow/conventions.md — shared formats, glossary (including the complexity axes and design gate terms), plan file structure, the §1.2 Per-axis artifact set, scope indicators, review iteration protocol
2. .claude/workflow/research.md — Phase 0 instructions: interactive research, code exploration, internet research, the research log (the durable Phase-0/1 decision ledger Phase 0 writes), transition rules

Do NOT read .claude/workflow/planning.md or .claude/workflow/design-document-rules.md yet — they are only needed when the user asks to create the plan (Step 4). Load them on demand at that point.

Resolve <dir-name>. All subsequent steps reference docs/adr/<dir-name>/_workflow/; resolve the placeholder once before running any command that uses it. If "$ARGUMENTS" is non-empty, use it. Otherwise, default to $(git branch --show-current).

Step 1.5 — Workflow drift check (mandatory, before any other on-disk work).

Ordering: this step depends on the <dir-name> resolver above being complete and Step 1b's mkdir not yet having run — see the trailing paragraph below for the gate's Skip-#1 rationale.

Invoke the drift gate defined in workflow-drift-check.md:planner:1. The gate is shared with /execute-tracks; its intro names both callers and its body is caller-symmetric, so this step is a thin orchestration handoff that defers to the gate rather than restating its detection. Run the gate's § Detection against the resolved <dir-name> from the previous block. Detection now runs the two-phase drift walk inside .claude/scripts/workflow-startup-precheck.sh under --mode full and reads the resulting drift JSON object; the script resolves the plan dir from the active branch, so no inline PLAN_DIR= bash line runs here. Follow its § Skip conditions, § No-drift normalization, and § Resolutions flow verbatim.

The three-resolution prompt fires only when drift surfaces and no skip condition matched. The user picks one:

• Migrate now — print Run /migrate-workflow from this worktree, then re-invoke /create-plan afterward. (the single instruction line per workflow-drift-check.md § Migrate now, with the /create-plan re-invocation hint appended), then end the session. Exit immediately; no on-disk work has run yet (Step 1b's mkdir, Step 2's aim prompt, and Step 5's commit and push are all downstream of Step 1.5).
• Defer — continue this session. Record the deferred-drift count via the TaskCreate todo described in workflow-drift-check.md § Defer; Step 5's deferred-drift recital reads that todo and prints the same line shape workflow.md § What to do before ending a session uses for /execute-tracks. If TaskCreate is unavailable in this session, hold the <count> and <short-stamp-base-SHA> (or the unstamped variant flag) in in-context memory instead, matching the gate file's § Defer paragraph.
• Suppress — continue this session with no recital at session end.

No-drift (with or without the gate's normalization commit), Defer, and Suppress all proceed to Step 1a without further user prompt. Ordering: Step 1.5 runs after the <dir-name> resolver (so the resolved name is available when the script's --mode full walk resolves the plan dir from the active branch) and before Step 1b's mkdir (so the script's internal Skip-#1 directory check reads the pre-creation _workflow/ state on fresh /create-plan invocations). The skip check is now internal to the script, not an inline gate-bash [ -d … ].

Interaction with Step 1a's handoff scan. Step 1.5 fires before Step 1a. On a /create-plan resume where handoff-*.md exists in docs/adr/<dir-name>/_workflow/, the drift gate fires before the handoff loader notices. No failure mode loses the handoff: on Migrate now the handoff file persists on disk (it is already committed) and the next /create-plan invocation's Step 1a picks it up after the drift gate clears; on Defer or Suppress, Step 1a's handoff resume runs after Step 1.5 in the same session. Per-session TaskCreate todos do not survive /clear, so a paused Session A's Defer state is not carried into Session B — Session B's Step 1.5 re-evaluates drift independently.

Step 1a — Handoff check (mandatory, before any other on-disk work). Run:

ls -t docs/adr/<dir-name>/_workflow/handoff-*.md 2>/dev/null

If any files exist, load mid-phase-handoff.md:planner:1 and follow its §Resume protocol BEFORE Step 1b. A previous /create-plan session paused mid-research or mid-planning and left a handoff to be re-presented. Do NOT ask for the aim, start fresh research, or write plan files until the handoff is resolved.

Step 1b — Create the workflow directory.

As the first durable action of /create-plan, ensure the workflow directory exists so research handoff files have a home if context fills up before Step 4:

mkdir -p docs/adr/<dir-name>/_workflow/plan

This is idempotent — safe to re-run on resume. The directory carries the research log, plan, design, track files, review files, and handoff files; the Phase 4 cleanup commit removes it before merge (see .claude/workflow/conventions.md §1.2).

Step 1c — Resume check (before the aim prompt).

After Step 1.5 (drift) and Step 1a (handoff) have cleared, check the design-first artifacts plus the phase ledger and the single-track glob on disk:

ls docs/adr/<dir-name>/_workflow/design.md \
   docs/adr/<dir-name>/_workflow/implementation-plan.md \
   docs/adr/<dir-name>/_workflow/phase-ledger.md \
   docs/adr/<dir-name>/_workflow/plan/track-1.md 2>/dev/null

The routing signal is the three complexity axes the ledger now carries (D10): the design gate (does a design.md exist), the plan-presence / track-count signal (how many track files the planner authored — more than one means an implementation-plan.md exists), and the Phase-1-complete marker (did Phase 1 finish cleanly). A plan-less change has no plan (D2), so implementation-plan.md presence cannot disambiguate it; the phase ledger is the signal — when the ledger records design_gate=no and the plan-presence signal is single (tracks=1), an interrupted plan-less session resumes off the ledger, with the plan/track-1.md glob as the secondary signal that the one track file was written. When the plan is present (tracks > 1), implementation-plan.md presence stays the routing signal, and the design gate is read from the ledger design_gate field, never from a plan line (the plan no longer carries a tier line; the classification moved to the ledger per D4) and never from a fresh read of the research log, which would be a third decision-content read site and break S2.

The new shape the three axes make expressible is a design with one track and no plan (design_gate=yes, tracks=1): on disk it is a design.md, no plan, and one track file — a file set identical to a mid-authoring crash that wrote design.md before deriving any plan. File presence alone cannot tell the two apart, so the Phase-1-complete marker is the disambiguator: set (phase1_complete=yes) means the design+single-track steady state (do not re-author); unset means a mid-authoring crash (re-enter Step 4a). The marker check runs first to separate "Phase 1 is done" from "Phase 1 is not done"; the existing committed-and-clean design.md check still applies within the crash arm to tell a frozen design apart from an unfrozen one.

Read the three fields from the ledger once before routing — the same last-value-wins fields the script's --append-ledger seeds at Phase 1 (conventions.md §1.1 Phase ledger):

LEDGER="docs/adr/<dir-name>/_workflow/phase-ledger.md"
# design gate (yes/no), track count (integer), and the Phase-1-complete
# marker (yes); each last-value-wins, empty if no ledger or no such line.
LEDGER_DESIGN_GATE="$(sed -n 's/.* design_gate=\([a-z]*\).*/\1/p' "$LEDGER" 2>/dev/null | tail -n 1)"
LEDGER_TRACKS="$(sed -n 's/.* tracks=\([0-9]*\).*/\1/p' "$LEDGER" 2>/dev/null | tail -n 1)"
LEDGER_PHASE1_COMPLETE="$(sed -n 's/.* phase1_complete=\([a-z]*\).*/\1/p' "$LEDGER" 2>/dev/null | tail -n 1)"

Route on what exists. Evaluate the branches in order; the first whose condition holds wins. The design.md-present branch fans out on the Phase-1-complete marker (steady state vs mid-authoring crash); the no-design branches differ from it by design.md absence. The plan-less resume branch and the fresh-start branch both describe the no-plan/no-design state, so the order matters: the more-specific plan-less resume is reached first and the fresh-start branch is the catch-all that fires only when no earlier branch matched.

• design.md exists, implementation-plan.md does not — a design.md is written (design_gate=yes) and no plan exists. Two on-disk-identical states share this signature — the design+single-track steady state and a mid-authoring crash — so check the Phase-1-complete marker first:

  • Marker set (LEDGER_PHASE1_COMPLETE = yes) — Phase 1 finished cleanly. This is the design+single-track steady state (design_gate=yes, tracks=1): the design and its one track file are the durable Phase-1 artifacts and there is no plan by derivation (a cross-track summary is vacuous for one track). This is a normal resume, not a Step-4 entry: the drift / handoff / state routing above already handled it; do not re-author the design or re-derive a plan. Proceed to Step 2 only if the user explicitly asks to start a new aim against the same dir (rare). The marker alone is sufficient by construction here: a clean Phase-1 seed co-writes phase1_complete=yes and design_gate=yes on the same line, so a set marker implies the gate and no cross-check of LEDGER_DESIGN_GATE is needed on this arm. The LEDGER_DESIGN_GATE / LEDGER_TRACKS locals parsed at the top of the step are read only by the lower no-design branches — they are intentionally unused here, not a missing check.

  • Marker unset (LEDGER_PHASE1_COMPLETE blank) — Phase 1 did not finish: the prior /create-plan invocation authored (and possibly committed) design.md but ended — crash, context-full /clear, or the user stopping the session — before the plan derivation completed and the marker was seeded. This is the crash-recovery arm. File presence alone is not proof the design is frozen: edit-design writes design.md to disk in its apply step, before the cold-read review runs and before the design commit lands. A session interrupted after the write but before the review passed leaves an unreviewed, uncommitted design.md on disk. So confirm the design is committed and clean — the on-disk proxy for "frozen and reviewed", since the Add initial design commit lands only after its review passes:

    # committed: at least one commit touches design.md
    git log -1 --format=%h -- docs/adr/<dir-name>/_workflow/design.md
    # clean: no uncommitted changes to design.md (empty output = clean)
    git status --porcelain docs/adr/<dir-name>/_workflow/design.md

    • Committed (non-empty git log) AND clean (empty git status) — the design is frozen and reviewed, and the prior invocation crashed between the design commit and the plan derivation. Auto-resume into Step 4b (plan derivation): skip Step 2's aim prompt and Step 3's Phase 0 research loop entirely — the aim and research are already captured in the frozen design.md and the conversation that produced it. Read planning.md (deferred from Step 1) and derive the plan from the frozen design. This crash-recovery resume reaches the same Step 4b the collapsed happy path flows into directly, so the plan derives identically whether or not the prior session was interrupted.
    • Uncommitted (empty git log) OR dirty (non-empty git status) — a session was interrupted mid-design-authoring (before the Add initial design freeze-and-commit). Resume Step 4a, not Step 4b: re-enter the edit-design review loop so the adversarial gate and cold-read pass run and the design is committed before any plan derives from it. This arm is retained by the collapse: even though the happy path no longer crosses a session boundary, a crash mid-authoring still leaves an unfrozen design.md, and re-entering Step 4a is how that state recovers. Re-entering the loop on an already-good design is idempotent and harmless, so this branch is safe even on a false alarm (e.g., a stray editor write left the file dirty).
• implementation-plan.md exists, design.md does not — a no-design multi-track change whose plan is already derived (design_gate=no, tracks > 1; a one-track change has no plan, so it never reaches this branch). The missing design.md is by derivation (design_gate=no), not a sign of an interrupted Step 4a. This is a normal resume, not a Step-4 entry: the drift / handoff / state routing above already handled it; do not re-run Step 4 and do not route to design authoring. If the ledger instead records design_gate=yes (or is absent / unreadable), a multi-track plan with no design.md is malformed (the design should have been authored and committed first) or the branch predates the ledger scheme; treat it as the Both files exist normal resume below and surface the inconsistency to the user rather than silently re-deriving. Proceed to Step 2 only if the user explicitly asks to start a new aim against the same dir (rare).
• Plan-less resume — ledger records design_gate=no and tracks=1, plan/track-1.md present, no implementation-plan.md, no design.md — a no-design single-track session whose one track file is already written (D10). The ledger, not a plan, is the resume signal here; the plan/track-1.md glob is the secondary signal that the track file landed. This is a normal resume, not a Step-4 entry: the drift / handoff / state routing above already handled it; do not re-run Step 4 and do not author a plan (a one-track change has none). Proceed to Step 2 only if the user explicitly asks to start a new aim against the same dir (rare).
• Ledger absent, plan/track-1.md present, no implementation-plan.md, no design.md — a no-design single-track Phase-1 session interrupted between the track-file write and the ledger seed (the seed runs after the track file is written; Step "Seed the phase ledger"). The durable Phase-1 artifact for a one-track no-design change is the track file, and it landed, so this is not a fresh start: re-authoring plan/track-1.md would clobber work already on disk. Resume by seeding the ledger (--design-gate no --tracks 1, plus the matched categories and any §1.7 staging mode) and continuing from the recorded state, not by re-running research, classification, or the Step-4b track-file Write. (A multi-track change cannot reach this branch: its durable Phase-1 artifact is implementation-plan.md, whose presence is matched by an earlier branch above.)
• Neither implementation-plan.md nor design.md exists, and no plan-less resume signal is present — fresh start. The "no plan-less resume signal" condition is the OR of three testable arms, evaluated against the LEDGER_DESIGN_GATE / LEDGER_TRACKS values parsed at the top of this step: the ledger is absent; OR the ledger is present but its design_gate field is empty or unreadable (LEDGER_DESIGN_GATE blank); OR the ledger records design_gate=no and tracks=1 but plan/track-1.md has not been written yet (the branch immediately above already claimed the case where the track file exists). Proceed to Step 2 (aim), then Step 3 (research), then Step 4 (the design-gate classifier + adversarial gate, then Step 4a/4b). This also covers the narrow /clear window where Step 4's gate cleared but no artifact was written yet: with no plan, no design.md, no seeded ledger fields, and no track file on disk there is no resume signal, so the resume correctly reads as a fresh start and Step 4's classifier re-runs, re-deriving the classification from the now-populated log through its existing sanctioned authoring read — no extra read site, S2 intact.
• Both files exist — the plan is already derived (a design-and-plan change with both committed); this is a normal resume, not a Step-4 entry. The drift / handoff / state routing above already handled it; do not re-run Step 4. Proceed to Step 2 only if the user explicitly asks to start a new aim against the same dir (rare); the common case is the session has nothing new to plan.

This check has a defined resume path for every artifact combination, so the "never a dead end" invariant holds for every arm. A design.md with no plan is never a dead end — with the marker set it is the design+single-track steady state (do not re-author); with the marker unset it is a crash, and the committed-and-clean check then routes a frozen design to Step 4b and an unfrozen one back to Step 4a. A derived plan resumes normally without re-entering design authoring, and a plan-less single-track session resumes off the ledger and its single track file (or, when the seed had not yet run, by seeding the ledger and continuing) rather than reading as a fresh start. The check runs after the drift and handoff gates so a pending migration or handoff resolves first (those can change what is on disk), and before the aim prompt so a Step-4b crash-recovery resume does not re-ask for an aim already captured in the design or the research log.

Step 2 — Ask the user for the aim, then seed the research log.

Skip this step when Step 1c auto-resumed into Step 4b (the aim is already captured in the frozen design.md and the research log). Otherwise, after you have finished reading the workflow documents, ask the user to describe the aim and goal for this session. Do NOT proceed until the user provides the aim. Wait for the user's response before starting any research or planning work.

Once the user provides the aim, write the research log's ## Initial request (the verbatim aim) as the first durable Phase-0 action. The _workflow/plan/ directory already exists from Step 1b (mkdir -p .../plan), so write the log directly: create docs/adr/<dir-name>/_workflow/research-log.md (a Write, not a shell command) with the six sections research.md §The research log defines: ## Initial request (the verbatim aim, written once); the empty ## Decision Log, ## Surprises & Discoveries, ## Open Questions continuous logs; ## Baseline and re-validation filled only on a workflow-modifying branch; and the empty ## Adversarial gate record the Step 4 gate appends its verdict headings to. The log is created unstamped: it is on the §1.6(f) never-stamped list (D19), so no line-1 workflow-sha comment is written and the §1.6(b) paired-idiom does not run for it. Idempotent on resume: if the log already exists (a prior Phase-0 session created it), leave its ## Initial request intact and append to the continuous logs only. The log is the agent's internal memory: seed it without narrating the seeding to the user (research.md §Rules, the Keep the research log agent-internal rule).

The plan will be saved to: docs/adr/<dir-name>/_workflow/implementation-plan.md (the _workflow/ subdir holds every ephemeral working file — research log, plan, design, track files, reviews — and is removed in the Phase 4 cleanup commit before merge; see conventions.md §1.2 and workflow.md § Final Artifacts). The codebase is at the current working directory.

Step 3 — Research phase (Phase 0).

Once the user provides the aim, enter research mode. In this mode:

• Answer user questions about the codebase, architecture, and design
• Explore code (read files, search for patterns, trace call chains)
• Perform internet research when asked (web search, fetch documentation)
• Present findings and intermediate conclusions
• Help the user evaluate trade-offs and alternatives
• Append decisions, surprises, and open questions to the research log as they settle — each entry an ISO timestamp and a [ctx=<level>] tag, each ## Decision Log entry carrying the **Why:** and **Alternatives rejected:** fields the Step-4 adversarial gate challenges (research.md §The research log for the append cadence). Do this silently: the log is agent-internal, so surface its content to the user as plain conversational prose, never as log quotes, section names, or D-numbers (research.md §Rules, the Keep the research log agent-internal rule)
• Do NOT produce plan files, design documents, or track decompositions

Stay in research mode until the user explicitly asks to create the plan (e.g., "create the plan", "let's plan this", "proceed to planning").

Step 4 — Classify the design gate, gate the research log, then transition to planning (Phase 1).

Phase 1 is complexity-adaptive (planning.md §Tier classification): a one-line fix does not pay the ceremony a durability rework needs. Step 4 runs in three parts at the Phase 0 → 1 boundary, before any Phase-1 artifact is authored:

1. Design-gate classification — propose the change-level design gate (does a design.md exist) from the now-rich research log, with the centrally-matched HIGH-risk categories, and let the user confirm.
2. The adversarial gate — run the relocated adversarial review on the research log as a gate (loop on blockers, gate on should-fix, no skip), domain-primed by the confirmed matched categories.
3. Transition — branch to Step 4a (design-first, when the design gate is yes) or straight to Step 4b (thinned plan + tracks, or one track file when the change is single-track).

Step 4 part 1 — The design-gate classifier.

When the user asks to create the plan, before authoring anything, confirm the research log captures the conversation's decisions (append any settled but unlogged), then classify the change. Read the now-rich research log's ## Decision Log, ## Surprises & Discoveries, and ## Open Questions — this is a sanctioned Step-4 authoring read (S2). Classify the change on two orthogonal gates (planning.md §Tier classification):

Gate 1 is the design gate this part persists: it decides whether a design.md (and the Phase-4 design-final.md) exists. Gate 2's multi-vs-single answer is not decided up front here — the track count is not knowable until the planner has decomposed the change, so plan presence is decided at the end of Step 4b from the authored track count (D1). design = yes implies multi-track, so the only single-track shape Gate 2 adds beyond the design gate is the no-design single-track change.

Gate 1's "needs a design" test reuses the HIGH-risk category list in risk-tagging.md §Gate 1 reuse (change-level), read at the change level: Gate 1 is yes only when one of those categories is central to the change's purpose, not merely touched by one incidental edit. Record the centrally-matched categories — they prime the adversarial gate's lenses in part 2 and seed the Phase-4 durable carrier's lens set (D16).

Propose the design gate (design_gate=yes/no) and the matched categories to the user and wait for confirmation. The user confirms or overrides the design gate in either direction, and may add or drop an adversarial lens explicitly at this point (D16) — confirming the design gate confirms the matched categories, so an override may shift the lenses. A design_gate=no change runs its gate lens-free unless the user adds one. This is a human gate on the artifact-shedding decision; do not proceed to part 2 until the design gate and the lens set are confirmed. The confirmed design_gate value is seeded into the phase ledger at the Step "Seed the phase ledger" below — it is the resume router's design-gate signal and the consistency / structural reviews' design-presence read.

Step 4 part 2 — The adversarial gate on the research log.

Once the design gate is confirmed, run the relocated adversarial review on the research log as a gate before any Phase-1 artifact derives (D6, planning.md §Tier classification). The gate spawns the existing reviewer-adversarial in its research-log scope (prompts/adversarial-review.md §Research-log-scoped review (Phase 0→1)); no new reviewer is added.

Create the gate's review-file directory once (idempotent), before the first spawn — the canonical track-anchored plan/track-N/reviews/ home does not exist yet, so the Phase-0→1 gate writes to a plan-scoped directory (conventions-execution.md §2.5 §Third-scope review-file home):

mkdir -p docs/adr/<dir-name>/_workflow/reviews

Spawn the adversarial sub-agent via the Agent tool (the same recipe the sibling edit-design/SKILL.md uses for the identical reviewer; the research-log Inputs block in prompts/adversarial-review.md §Research-log Inputs substitutes the inputs):

• subagent_type: general-purpose

• description: "Adversarial research-log gate (Phase 0→1)"

• prompt: the full content of .claude/workflow/prompts/adversarial-review.md. The prompt's TOC-protocol header resolves the reviewer's phase to 0→1, which routes it to the § Research-log-scoped review (Phase 0→1) section. Substitute these inputs into that section's ### Research-log Inputs block:

  - research_log_path: docs/adr/<dir-name>/_workflow/research-log.md
  - matched_categories: the centrally-matched HIGH-risk categories from the
    confirmed design gate's Gate 1, plus any user-added lens — or (none) for a
    `design_gate=no` change with no user lens
  - output_path: docs/adr/<dir-name>/_workflow/reviews/research-log-adversarial-iter<N>.md
    (one file per gate iteration, <N> starting at 1 — the <type>-iter<N>.md
    naming conventions-execution.md §2.5 §Third-scope review-file home fixes)
  - codebase_path: the repo root

Model and effort (D14). Pin the spawn's model on the Agent tool's model field by the confirmed design gate:

• design_gate=yes → model: fable
• design_gate=no → model: opus

The Agent tool has no per-spawn effort field, and there is no adversarial-reviewer agent file under .claude/agents/ to carry effort in frontmatter — the adversarial reviewer is a prompt-file plus a general-purpose spawn. So the model half lands on the model field as above, and the xhigh-effort half rides the session default (it cannot be pinned per-spawn through this surface). D14 accepts the effort caveat: the effort half degrading to the session default does not reopen the decision.

Output handling (D17). The reviewer's output mode is file: it persists the conventions-execution.md §2.5 manifest-plus-sections review file to output_path and returns only the thin manifest. Validate the manifest's findings count against the file with the §2.5 count grep (grep -cE '^### [A-Z]+[0-9]+ ' <file>) before trusting the index, then partial-fetch ## Findings from disk — do not pull the whole file into context. This caps the gate loop's context cost and makes a mid-gate /clear resumable from the committed file. Commit the review file at reviewer-return as a Workflow-update commit (the resume precondition; conventions-execution.md §2.5 §Third-scope review-file home).

Gate semantics (no skip). This run is a gate, not an advisory pass:

• A blocker sends the decision back to research to be re-decided; the gate loops — re-spawn the reviewer (incrementing <N>) after the log decision is revised, until no blocker remains. The iteration-1 run is a fresh finding set; iteration ≥2 runs use the verdict-producer manifest variant (per-prior-finding VERIFIED / STILL OPEN / REJECTED verdicts plus any new finding), per conventions-execution.md §2.5 §Verdict-producer manifest variant.
• A should-fix gates: the log's rationale must strengthen before the gate clears.
• Iteration cap (iteration_budget, default 3). The re-spawn loop is bounded exactly as the sibling edit-design cold-read loop is (edit-design/SKILL.md § Inputs iteration_budget, § Failure modes and recovery). Cap the gate at iteration_budget re-spawns. On exhaustion with blockers or unaddressed should-fix findings still open, do not loop further: the user is the gate — surface the still-open findings and the decision history and let the user accept the risk, revise the decision, or abandon the change, mirroring edit-design § Failure modes "the user is the gate when the budget is exhausted". A gate must never spin unbounded on a contested decision.
• There is no skip — the log is not a track that can be dropped; a would-be skip is raised to a blocker so the change is re-justified in research before any artifact derives.

Record the verdict on the log. At each gate-clear or re-challenge, append one verdict heading to the research log's ## Adversarial gate record section in the canonical shape defined once in research.md §The research log (Gate-record cadence): ### Adversarial review of this log (<ISO>) — <PASS | NEEDS REVISION[: <counts>]>, followed by a one-line pointer to the iteration's _workflow/reviews/research-log-adversarial-iter<N>.md file. This on-log record is the gate's durable verdict carrier. The gate's review files are ephemeral (they die at the Phase 4 cleanup); the durable verdict carrier the Phase-4 consumers and the S3 freeze-order gate read is this ## Adversarial gate record section, not the review files (conventions-execution.md §2.5 §Third-scope review-file home).

Pre-presentation re-trigger vs post-presentation queue. While authoring a Phase-1 artifact (the full-tier Step-4a design), a load-bearing decision appended to the research log re-opens the gate on that entry immediately (D5). Once a frozen-ready artifact is presented for user review, findings instead queue and batch through one gate run — the D15 review-iteration batching, whose queue mechanics (the tagged [clarification]/[decision] queue, the three-step batch, and the multi-session handoff queue block) live in this SKILL's review-hold batching section and mid-phase-handoff.md. Step 4 here owns the first, pre-presentation gate run; the batch loop is the consumer of the same gate.

Step 4 part 3 — Transition to Phase 1.

After the gate clears, branch on the confirmed design gate (part 1's only confirmed output — the multi-vs-single track count is not decided here; it is settled at the end of Step 4b, where the plan-presence decision lives):

• design_gate=yes — design-first: Step 4a (author + review + freeze design.md) then Step 4b, within one /create-plan invocation (the freeze-and-commit between them stays the logical gate and crash checkpoint, but is no longer a session boundary), exactly as the rest of this Step describes.
• design_gate=no — no design.md. Skip Step 4a; go straight to Step 4b and author the track files directly from the research log in a single Phase-1 session, with the full inline Decision Records and no design seed to derive from. The thinned derived-mirror plan is authored at the end of Step 4b iff the planner decomposed into more than one track (the plan-presence decision, D1); resume state otherwise lives in the phase ledger (conventions.md §1.2 Per-axis artifact set).

The design_gate=yes design-first split keeps the design-authoring and plan-derivation work in order, but both run in one /create-plan invocation:

• Step 4a (design authoring, design_gate=yes only) — author design.md via edit-design, run its review, and freeze it. The design's review passing (or the user accepting open risks) is the gate that releases Step 4b; the freeze-and-commit is the crash checkpoint but no longer ends the session.
• Step 4b (plan derivation) — in the same /create-plan invocation for full (flowing on from Step 4a once the design is frozen and committed), or the same Phase-1 session for lite/minimal, derive the Architecture Notes, Decision Records, and track files (from the frozen design.md in full; from the research log in lite/minimal).

Design→plan flow within one invocation (full tier only). In full, Step 4a freezes and commits design.md (Add initial design) and then flows straight into Step 4b in the same /create-plan invocation; the two no longer span a session boundary. The freeze-and-commit stays the logical gate and the crash checkpoint — the plan derives only from a committed, frozen design — but it no longer ends the session. The context isolation the old boundary forced is supplied directly by sub-agent authoring: Step 4b's track derivation runs through the design-author spawn, a fresh cold spawn that reads the frozen committed design regardless of session, so no /clear is needed to keep design-authoring context out of plan derivation. The collapse therefore rests on a by-reference orchestration invariant (built in Track 1, confirmed by gate A6 before this collapse was applied): the author spawn returns only a thin summary, never the drafted document, so the combined session does not re-accumulate the design and plan context the boundary kept apart. The running skill always takes the collapsed path — there is no runtime branch here that the executor evaluates. The "retain the boundary instead" fallback is a design-time decision rule, not a live one: it governs whether this collapse ships at all. By-reference was confirmed statically (gate A6, held green) for this staged change; the live-harness re-confirmation is a deferred Phase-4-promotion / first-live-run gate. Were a future change to break the invariant, the boundary would be reinstated at that authoring step, not skipped at runtime.

The startup protocol's auto-resume into Step 4b is now crash-recovery-only: it fires when design.md is committed and clean, implementation-plan.md does not exist, AND the Phase-1-complete marker is unset — the state a crash leaves between the design commit and the plan derivation. The marker check comes first: a set marker with that same on-disk file set is the design+single-track steady state (do not re-author), not a crash; the committed-and-clean test then (within the unset-marker crash arm) proves the design is reviewed rather than abandoned mid-authoring. This is checked after Step 1.5 (drift) and Step 1a (handoff) have cleared and before the aim prompt (Step 2): a crash-recovery resume into Step 4b skips the aim prompt and the Phase 0 research loop, because the aim and research are already captured in the frozen design.md and the conversation that produced it. Step 1c above is the single decision-rule home for every artifact combination — it spells out the marker fan-out, the exact git log / git status check, the branch order, the never-a-dead-end fallback, and the resume-Step-4a arm for a dirty or uncommitted design; this block does not re-derive that routing.

A design_gate=no change has no design.md and so no Step 4a at all: its Step-4b plan derivation runs in the same Phase-1 session that Step 4 part 1/2 ran in, with no design freeze in between. After the collapse full also runs Step 4a and Step 4b in one invocation, so the difference is no longer single-session vs two-session — it is whether a design.md is authored and frozen first (full) or not (lite/minimal). Step 1c's tier-aware branch keeps an interrupted no-design tier (plan on disk, no design.md by design) routing to a normal resume rather than back into design authoring.

Step 4a — Author the design first (design_gate=yes only).

This sub-step runs when design_gate=yes only. When design_gate=no there is no design.md; skip directly to Step 4b. When the user asks to create the plan with design_gate=yes (and design.md does not yet exist):

First, read the design workflow document (deferred from Step 1):

• .claude/workflow/design-document-rules.md — design document rules, structure, and examples

Summarize the key research findings and decisions from the conversation, then author design.md via the edit-design skill (phase1-creation kind) — not direct Edit / Write. phase1-creation runs the dual-clean inner loop (see edit-design/SKILL.md § Workflow and design-document-rules.md § Working / sync): each round spawns the code-grounded author plus a per-round readability-auditor plus a separate per-round warm absorption-check that runs the absorption-completeness cross-check (every load-bearing research-log decision in scope appears as a seed D-record in design.md). After the inner loop converges to dual-clean, the cold comprehension-review gate runs once; it assesses whether a fresh reader can build a working mental model and runs no absorption cross-check inline (that check is the separate absorption-check spawn's job). The cold comprehension gate is gated behind the log-adversarial gate clearing (S3): the gate cannot run while a log-adversarial entry is open, so a load-bearing decision surfaced while authoring the design is appended to the log, re-challenged at the gate, and cleared before the comprehension gate assesses the draft, matching the ordering the in-edit-design adversarial pass used to give. Iterate until the comprehension gate passes (or the user accepts open risks), then write the design document to docs/adr/<dir-name>/_workflow/design.md using the structure below. The design document must incorporate findings and decisions from the research phase — it reflects the design choices discussed with the user.

Commit the frozen design.md (Step 5 carries the commit/push/draft-PR mechanics; the Add initial design commit is the logical gate and crash checkpoint), then flow straight into Step 4b in the same /create-plan invocation — do not end the session. The freeze-and-commit no longer ends the session; sub-agent authoring supplies the context isolation the old boundary forced (see the Design→plan flow block above). The auto-resume condition above is the crash-recovery path that re-enters Step 4b only when this invocation ended before deriving the plan.

Step 4b — Derive the plan and track files.

In full, Step 4b runs in the same /create-plan invocation, flowing on from Step 4a once design.md is frozen and committed (the design seed the plan derives from); the crash-recovery resume re-enters here only when a session ended after the design commit but before the plan derived (design.md committed and clean, implementation-plan.md absent). In lite/minimal, Step 4b runs in the same Phase-1 session immediately after the Step 4 gate clears — there is no design.md, so the research log is the seed the carriers absorb. Read the planning workflow document (deferred from Step 1):

• .claude/workflow/planning.md — Phase 1 instructions: goal, tier classification, plan file structure, architecture notes format, track descriptions, scope indicators, checklist decomposition rules

Then derive the plan and track files. The decision seed is tier-keyed (S2): in full, seed Decision Records from the frozen design.md seed's D-records; in lite/minimal, Step-4b authoring is itself the research log's sanctioned read point, so the track Decision Records absorb the log's load-bearing decisions directly. The thinned plan (in lite/full; absent in minimal), the track Decision Records, and track files must incorporate findings and decisions from the research phase (and, in full, the frozen design):

• Track Decision Records reflect alternatives explored during research, absorbed as full inline records in each relevant track's ## Decision Log (the track-canonical live decision, D7)
• Architecture Notes build on codebase exploration findings
• Track descriptions incorporate constraints discovered during research
• In full, the design.md seed reflects design choices discussed with the user; the track inline records are seeded from (and stay faithful to) it

Help the user develop the plan:

1. Understand the relevant parts of the codebase — explore the modules, packages, and classes relevant to the goal. Build a mental model before proposing anything.

2. Identify key decisions and constraints — technical, performance, compatibility, and process constraints that will shape the plan.

3. Produce Architecture Notes following the workflow rules:

   • Component Map (required): Mermaid diagram if 3+ components with non-trivial relationships, always paired with annotated bullet list.
   • Decision Records (required): one per non-obvious design choice, with alternatives, rationale, risks, and track references.
   • Invariants & Contracts (if applicable): must map to testable assertions.
   • Integration Points (if applicable): how new code connects to existing code.
   • Non-Goals (if applicable): explicit scope boundaries.

4. Decompose the work into tracks with full descriptions following the workflow rules:

   • Every track gets an intro paragraph in the plan checklist entry (a short paragraph of high-level context) and a matching plan/track-N.md track file whose ## Purpose / Big Picture section carries a one-line BLUF followed by the same intro paragraph. The track's detailed content spreads across three other plan-at-start homes (no length cap on any of them): ## Context and Orientation carries the codebase state at the start of the track and the concrete deliverables it produces; ## Plan of Work carries the prose sequence of edits and additions plus ordering constraints and invariants to preserve; ## Interfaces and Dependencies carries in-scope/out-of-scope file boundaries, inter-track dependencies, and library/function signatures. See conventions-execution.md §2.1 for the canonical section list and lifecycle.
   • Populate each track file's ## Decision Log with the full inline Decision Records the track owns — the track-canonical live decision carrier (D7) in every tier. Each is a complete four-bullet DR (Alternatives considered, Rationale, Risks/Caveats, Implemented-in), seeded from the frozen design.md D-records in full and authored directly from the research log in lite/minimal. The track file is the live authority in every tier; in full, the design.md seed copy is historical provenance with an optional **Full design** line into the seed's mechanism — it never substitutes for the inline record.
   • Populate each track file's combined ## Invariants & Constraints section (D9) with the track's testable technical / performance / compatibility constraints and the testable invariants the plan's Architecture Notes used to carry — they share this home because both are a property that must hold, backed by a test (each invariant becomes a test assertion in the relevant step). A process-only, non-testable constraint goes to ## Context and Orientation or ## Decision Log instead; Integration Points fold into ## Interfaces and Dependencies, and Non-Goals move to the research log and the PR ## Motivation (and design.md in full).
   • Include a track-level Mermaid component diagram inside the track file's ## Context and Orientation section when the track has 3+ internal components with non-trivial interactions. Track-level diagrams are never rendered in the plan file.
   • Track sizing rule: size each track by its in-scope file footprint, not its step count. Maximize — pack autonomous units in up to the soft footprint ceiling (related or not), opening a new track only when the next unit breaches the ceiling or breaks independent mergeability. A track ≤~12 in-scope files that folds into a neighbor is a merge candidate (flag-only); a track over ~20-25 in-scope files is a split candidate. Both bounds are soft: an out-of-bounds track passes when its track file carries a written justification. The full rule lives in planning.md §Track descriptions. The execution agent handles sequencing and episode propagation between dependent tracks.

5. For each track, include a Scope indicator:

   • Format: > **Scope:** ~N files covering X, Y, Z
   • Approximate file footprint + brief list of major work pieces. The footprint is a per-track soft heuristic, not the per-step ~12 split cap; the in-scope file set already lives in the track file's §Interfaces.
   • These are strategic signals, not tactical commitments — step decomposition happens during Phase 3 execution.
   • Do NOT include full - [ ] Step: items or (provisional) markers.
   • Focus energy on track descriptions and architecture, not premature step decomposition.

6. Order the tracks so dependencies are respected — earlier tracks don't depend on later ones. Annotate dependencies with > **Depends on:** Track N.

7. Identify key test scenarios and invariants that must be covered — this is strategic (what to test and why), not tactical (how to implement tests).

8. Anchor the carriers to the tier's decision seed. In full, anchor every Architecture Note, Decision Record, and track description to the frozen design.md authored in Step 4a: the design is the seed the plan derives from, so the track inline Decision Records mirror its D-records, the Component Map matches its class / workflow diagrams, and each DR that needs long-form support links to a design section via **Full design**. The design is not re-authored here — it is frozen (design-document-rules.md Rule 15). If plan derivation surfaces a design gap that the frozen design cannot answer, route the design intent through a fresh edit-design mutation in this Step 4b session before the freeze re-applies; do not back-fill it silently into the plan. In lite/minimal there is no design.md: anchor the carriers to the research log directly — Step-4b authoring is the log's sanctioned read point (S2), and the track inline Decision Records absorb the log's load-bearing decisions and their rejected alternatives.

9. Author the track sections through the dual-clean loop (Step-4b loop, every tier). Do not author the track-file prose inline. The track files are the durable human-facing carrier in every tier (the only one in lite / minimal, where no design.md buffers the reader from log-derived density), so they go through the same code-grounded author plus dual-clean inner loop that edit-design runs on design.md. Items 1-8 above settle the decisions, the track boundaries, the Decision Records, and the section homes; this item hands that settled shape to the design-author spawn, which writes the prose, and runs the readability auditor and the absorption check against what it wrote. This is the track-path analog of the edit-design phase1-creation loop (edit-design/SKILL.md § Workflow, § Step 6), parameterized to target=tracks. Run it after items 1-8 produce the track shape and before the Step 5 commit.

   The loop reuses three of the four roles Track 1 built (the design-author, the readability-auditor, and the absorption-check) plus the de-warmed comprehension-review gate. Each is an agent definition with a minimal tools: allow-list spawned by subagent_type basename, with its per-spawn parameters in a params file under _workflow/reviews/ (one file per spawn; the plan-scoped authoring-loop review home, conventions-execution.md §2.5 Third-scope review-file home — the same home edit-design/SKILL.md § Step 4 uses) and a byte-identical spawn-prompt body that names only that file, so the shared prompt body caches across the fan-out. The spawn contracts, the params-file keys, and the fan-out cache warm-up are the same ones edit-design/SKILL.md § Step 4 defines for the creation kinds; this item reuses them with the track-path values below rather than restating them.

   Round 1: spawn the author (subagent_type: design-author, description: "Track authoring (Step-4b round 1)"). Its params file carries target=tracks, output_path set to the plan/ directory the track files land under, research_log_path set to the research log under _workflow/, round=1, and (in full only) design_path set to the frozen design.md the track prose derives from. The author's seed source is the frozen design.md in full, the research log directly in lite / minimal. The author grounds the whole change, writes every track file's plan-at-start sections, and returns a thin summary only, never the drafted track files (the by-reference contract; if it returns the draft, the track files are still on disk, so proceed with them and note the violation, per edit-design/SKILL.md § Failure modes). On later rounds the author is re-spawned with the auditor's flagged passages and re-grounds only those, as in the inner loop below.

   Per round, run the pair, then evaluate dual-clean. Each round spawns the cold readability-auditor (range-sliced fan-out, one spawn per slice, sequenced behind the cache warm-up) and the warm absorption-check, then the round is dual-clean when neither returns a blocker or should-fix finding:

   • The readability-auditor (subagent_type: readability-auditor, description: "Readability audit (Step-4b round <N>)") audits the track prose with target=tracks and a slice range per spawn. Slice the track-path fan-out per track file: one readability-auditor spawn per plan/track-N.md (in track-number order), its params file carrying target=tracks, target_path set to that one track file, and a whole-file range (line 1 to the file's last line). This per-file fan-out omits slice_count and total_lines — those two params are passed only on the design-path creation-kind fan-out, so the agent-side whole-doc guard stays inert on the track path (per readability-auditor.md § The whole-doc guard, the guard applies only when both params are present). The single (target_path, range) slice model that edit-design/SKILL.md § Step 4 defines for the design path (one ~200-line window in the single design.md) does not partition N track files on its own; this per-file rule is the deterministic partition, so every orchestrator run produces the same set of slices and the same per-slice anchor visibility. The standing anchors fill the cross-file vocabulary a single-file slice lacks: with target=tracks they are the plan Component Map and each track's ## Purpose / Big Picture (the design-path anchors ## Overview and ## Core Concepts do not exist on the track path). It owns the prose AI-tell axis on this surface (S4) and reads no log (S1). Its too-terse findings become the next round's author flagged_passages.
   • The absorption-check (subagent_type: absorption-check, description: "Absorption check (Step-4b round <N>)") is the second check: a separate spawn, not a clause inside the comprehension gate. Its params file carries target=tracks, research_log_path set to the research log, and draft_path set to the plan/ directory (a directory, not a single file: the absorption check reads every plan/track-N.md ## Decision Log under it, per absorption-check.md § Inputs), plus design_path set to the frozen design.md in full. It two-way coverage-matches the load-bearing research-log (or, in full, design.md seed) decisions against each track's ## Decision Log records: a load-bearing decision missing from the tracks is a finding, and a track record inventing a decision the log lacks is a finding that re-opens the S3 gate (below) when it is load-bearing, exactly as a decision-shaped comprehension finding does. It carries design_path only as that seed decision source; the seed↔track fidelity criterion is owned by the comprehension gate (below), so the absorption check does not run it.

   A round that is not dual-clean re-spawns the author with the auditor's flagged passages and any absorption log-missing-from-draft decision to seed, then re-runs the pair. The loop is bounded by iteration_budget (default 3) and exits to the user on budget exhaustion (S5), the same termination contract the Step 4 part 2 gate and edit-design/SKILL.md § Step 6 carry — the cap is restated here so the loop carries its own termination rather than borrowing it across skills.

   Cross-round convergence: the same mechanism, parameterized. The track path runs the identical cross-round convergence mechanism the design path does — edit-design/SKILL.md § Step 6 "The canonical convergence mechanism (section-keyed settled-state)" is the canonical statement; this loop cross-references it rather than restating it. Apply it with the track-path parameters: the settled-state key is per track-N.md file (not per ## / # Part section as on the design path), and the standing-anchor set the anchor-folded content hash folds in is the plan Component Map plus each track's ## Purpose / Big Picture (not ## Overview + ## Core Concepts). Fold in only the track-path anchors that exist, mirroring the design-path "when present" tolerance: each track's ## Purpose / Big Picture, plus the Component Map when the plan carries one. An absent Component Map on a thin minimal plan is not an error and does not force a re-audit by itself — the hash simply folds fewer anchors. A track that returned clean and is unchanged (its anchor-folded hash matches last round) has its re-flags dropped; a Component Map edit re-opens every track's settled-state, the same way an Overview edit does on the design path. The standing anchors are byte-stable for the loop's duration — items 1-8 settle the plan Component Map and the track skeletons before this item's dual-clean loop runs — so the hash does not churn (this holds in lite / minimal too: the Component Map is not still in flux during the loop). This adds only the convergence-mechanism cross- reference. The slicing unit stays per-file as already stated above — one readability-auditor spawn per plan/track-N.md, whole-file range; do not read this as "apply the same partition," which is the design-path per-window rule that does not partition N track files.

   The gate-A7 warm-up deferral (inherited from Track 1). The fan-out cache warm-up is a cost lever, not a correctness dependency: the loop must produce correct dual-clean output with the warm-up disabled (the disabled path just pays N cold prefixes). When the harness offers no non-blocking fixed-delay mechanism, disable the warm-up; Step-4b acceptance does not require a working warm-up.

   The S3 freeze-order gate, then the comprehension gate. After the inner loop reports dual-clean, hold the S3 freeze-order gate across the loop before running the comprehension gate: the comprehension gate must not run while a log-adversarial entry is open. Read the research log's ## Adversarial gate record (the gate's durable verdict carrier; the open/resolved and latest-dated-entry rules are in research.md §The research log under Gate-record cadence), and run the comprehension gate only when the latest entry is resolved (the same ordering edit-design/SKILL.md § The S3 freeze-order gate applies on the design path). Then spawn the de-warmed comprehension-review gate once (subagent_type: comprehension-review, description: "Cold comprehension gate (Step-4b track cold-read)"). Its params file carries the ## Inputs it forwards to prompts/design-review.md (target=tracks, scope=whole-doc, plan_dir, and plan_path, plus design_path in full for the full-tier seed↔track fidelity criterion) and names no research_log_path and no output_path: the absorption cross-check moved off this role onto the absorption-check spawn above, so the comprehension reviewer runs no prose axis (S4) and no absorption cross-check, and with no output_path it returns its verdict inline. The Step-4b gate omits the output_path file-write branch that phase4-creation uses (edit-design/SKILL.md § Step 4) deliberately: its return is the bounded comprehension verdict plus a summary-shaped ## Structural findings list, not the long-form structural detail phase4-creation persists, so the inline return stays small even on a wide full surface of N track files plus the frozen design.md. A decision-shaped comprehension-gate finding re-opens the S3 gate and re-enters the inner loop (the author seeds the surfaced decision, the absorption check confirms coverage, the auditor re-checks the prose), so the comprehension gate re-runs only once the gate clears again. A comprehension-gate re-open consumes a round of the same iteration_budget the inner loop counts down — it does not start a fresh count. The dual-clean inner-loop rounds and any gate re-entries share one budget, so the whole Step-4b review is bounded by iteration_budget total rounds across both; this is the single-budget reading edit-design/SKILL.md § Step 6 settles on the design path, restated locally so the loop's termination is fully self-contained rather than borrowed across skills.

   The whole Step-4b review therefore exits when the inner loop is dual-clean and the comprehension gate passes, or the budget is spent. The written plan and tracks are then presented for the user's pre-persist confirmation, which is the presentation D15's review window opens at.