/flow-next:make-pr — PR-as-cognitive-aid

A reviewable PR body is itself an artefact: it lets a human decide where to focus before skimming the diff. flow-next already collects every input that body needs — the spec with R-IDs, per-task done summaries and evidence commits, decisions / bug / architecture-patterns memory entries, glossary changes, strategy alignment, deferred review findings, the git diff itself. This skill stitches those into a structured body, optionally adds mermaid diagrams for module-boundary changes, and pushes via gh pr create.

The host agent (Claude Code / Codex / Droid) reads the structured payload from flowctl spec export-cognitive-aid and synthesizes the body directly. Every claim in the body must trace to a structured field in the export payload — never fabricate file paths, SHAs, R-ID attributions, or "why" reasoning. Unknown attribution is honest ("uncovered" / "unclear") rather than invented. The host is competent at "what looks important here?" given the rich input; no second-model review pass is needed (the structured payload does the heavy lifting).

flowctl provides only thin plumbing: flowctl spec export-cognitive-aid <spec-id> --base <ref> --json aggregates the inputs into a single JSON payload (Task 1 of this spec). The skill renders the body, then pushes and creates the PR directly — no confirm prompt (invoking make-pr is the intent; the body is deterministic; the default is a reversible draft). --dry-run prints the body without creating; --ready/--draft set draft state.

Read workflow.md for Phases 0–3 (pre-flight → gather → render body → mermaid) + the §4.0 --dry-run short-circuit — each phase ends with its inline ### Done when checklist. You MUST also read pr-cognitive-aid.md before composing any body — it IS Phase 1.5 (compose → flowctl pr-cognitive-aid validate/write → deterministic render), runs on EVERY invocation including --dry-run, and its rendered walkthrough supersedes the legacy Verification section, and the legacy R-ID coverage section when coverage is fully evidenced (with any unevidenced or undeclared criterion that table renders beside the walkthrough - pr-cognitive-aid.md §4 owns the rule); a body composed without executing it is a contract violation, not a style choice. Phase 1.5b additionally loads html-lens.md only when HTML artifacts are enabled and the run is not --dry-run. The post-render create + finalize machinery (§4.1 title → §4.6 gh pr create/--update → Phase 5 receipt/footer) lives in create-and-finalize.md, read ONLY on a real create (after §4.0 does not short-circuit) — a --dry-run preview never loads it. Read mermaid-rules.md before emitting any mermaid codefence — it defines reserved words, escape patterns, shape selection, the hard caps + allocation rule, the prose-summary rule, the pre-emission validation checklist, the Phase-3 hallucination guardrails, and the diff-fenced structural sketch alternate emission (§8); the --no-mermaid / no-trigger / skip-rule paths never load it. Phase 2's §2.11b Live QA section is likewise gated: references/live-qa-section.md is read only when the spec's qa_verdict receipt is present (the uncommon case).

Preamble

CRITICAL: flowctl is BUNDLED — NOT installed globally. which flowctl will fail (expected). Define once; subsequent blocks (here and in workflow.md) use $FLOWCTL:

FLOWCTL="${CODEX_HOME:-$HOME/.codex}/scripts/flowctl"
[ -x "$FLOWCTL" ] || FLOWCTL="<plugin-root>/scripts/flowctl"   # <plugin-root> = the directory two levels above this skill's SKILL.md file (the harness gave you that file's absolute path when the skill loaded); substitute it literally
[ -x "$FLOWCTL" ] || FLOWCTL=".flow/bin/flowctl"

Inline skill (no context: fork) — plain-text numbered prompt must stay reachable for the Phase 0 info prompts (resolve a missing base ref / undetected spec id — never a confirm gate). Subagents can't call plain-text numbered prompts (Claude Code issues #12890, #34592). There is no Phase 4 confirm prompt — make-pr creates the PR directly.

Mode Detection

Parse $ARGUMENTS as a flag list. Recognized flags: --draft, --ready, --no-mermaid, --memory, --dry-run, --base <ref> (consumes the next token), and the literal token mode:autonomous. Strip recognized tokens; the remainder (if any) is the optional spec id.

RAW_ARGS="$ARGUMENTS"
DRAFT_FORCE="auto"      # auto | draft | ready
NO_MERMAID=0
WRITE_MEMORY=0
DRY_RUN=0
BASE_REF=""
SPEC_ID=""
AUTONOMOUS=0

# Tokenize and walk the argument list. The loop handles both `--base=<ref>`
# and space-separated `--base <ref>` via a PREV token holder. Deliberately NO
# bash positional parameters here — the host's argument interpolation rewrites
# positional tokens inside skill code blocks (pilot dogfood finding, 1.13.0).
PREV=""
for ARG in $RAW_ARGS; do
  case "$PREV" in
    --base) BASE_REF="$ARG"; PREV=""; continue ;;
  esac
  case "$ARG" in
    --draft)      DRAFT_FORCE="draft" ;;
    --ready)      DRAFT_FORCE="ready" ;;
    --no-mermaid) NO_MERMAID=1 ;;
    --memory)     WRITE_MEMORY=1 ;;
    --dry-run)    DRY_RUN=1 ;;
    --base)       PREV="$ARG" ;;
    --base=*)     BASE_REF="${ARG#--base=}" ;;
    mode:autonomous) AUTONOMOUS=1 ;;
    -*) echo "Unknown flag: $ARG" >&2; exit 2 ;;
    *)  SPEC_ID="$ARG" ;;
  esac
done
[[ -n "$PREV" ]] && { echo "Flag $PREV given without a value" >&2; exit 2; }

# Secondary signal: process-level autonomous driver (env survives only
# within one process tree; the token is the primary, prose-safe carrier).
if [[ "${FLOW_AUTONOMOUS:-}" == "1" ]]; then
  AUTONOMOUS=1
fi

Ralph mode (FLOW_RALPH=1 or REVIEW_RECEIPT_PATH set) is detected separately in workflow.md §0.0 — the skill is not Ralph-blocked. Under Ralph the skill hard-errors instead of asking the Phase 0 info prompts, forces --draft, and emits the PR URL to stdout. (The PR is created directly in both modes — the only difference is forced-draft + no Phase 0 prompts under Ralph.) Autonomous mode is a SEPARATE flag: AUTONOMOUS=1 derives only from the mode:autonomous token or FLOW_AUTONOMOUS=1 and never sets RALPH. Under RALPH || AUTONOMOUS the Phase 0 info prompts hard-error and --draft is forced (--ready ignored with a note); the PR_URL= stdout contract and all receipt/harness semantics remain Ralph-only.

Interaction Principles

Ask the user via plain text. Render the options below as a numbered list 1. … N., followed by a final option N+1. Other — type your own answer. Print the question, then the numbered list, then stop and wait for the user's next message before continuing. Parse the reply as: a bare number 1–N+1 → that option; the literal text of an option label → that option; free text after Other → custom answer.

• Ask one question at a time via plain-text numbered prompt. Never silently skip the question.
• Lead with the recommended option and a one-sentence rationale.
• No confirm gate. make-pr opens the PR without asking. Phase 0 asks only to resolve info it cannot derive (no --base and no detection match; no spec detected) — never "do you want to create it?". Not-all-tasks-done warns and proceeds (the open items make it a draft). Skip questions when context resolves cleanly.
• Ralph and autonomous modes skip all questions. Detect both once at Phase 0 and route deterministically; a genuinely unanswerable gap hard-errors with a clear message (NEEDS_HUMAN-style) instead of hanging on a prompt.

Hallucination guardrails

The body is synthesized from the export payload. Every claim must trace to a structured field. The skill explicitly forbids:

• Inventing file paths. Only paths returned by git diff --name-status (via the diff.files array) appear in Critical Changes / Review plan. No "I think there's also a config file" content.
• Inventing risk or verification claims. Every "must review because…" clause in the Review plan traces to a diff_summary risk signal (churn / public export / security path / cross-module edge / user-facing surface); every "the pipeline verified…" line in the How-to-review block traces to tasks[].evidence / R-ID coverage / reviews.*. No narrated risk and no claimed verification without a payload anchor — absent verification is stated honestly ("no cross-model review recorded on this PR").
• Fabricating commit SHAs. SHAs come from tasks[].evidence[].commits and git log --oneline base..HEAD only.
• Guessing R-ID coverage. Coverage is computed from task satisfies frontmatter. Declared and evidenced are distinct: an R-ID no task claims (undeclared_r_ids) gets a ⚠️ flag; one claimed by a task that is not done yet renders as ⏳ claimed, not yet evidenced. Never a confident attribution either way.
• Inventing "why" reasoning. Decision context comes from memory.decisions[] entries' bodies. If no decision entry exists for a change, the body says so explicitly rather than narrating a plausible-sounding rationale.
• Quoting raw diff content. The body talks ABOUT the diff (paths, churn, modules). Never includes code snippets — privacy + secret-leakage risk; GitHub renders the actual diff below the body.
• Synthesizing review findings. Findings come from reviews.deferred[] and reviews.suppressed_count. The body never editorializes severity or fabricates findings.
• Generating fictitious memory IDs. When the body references memory entries (decisions / bugs / patterns), the IDs come from the export payload — never interpolated.
• Synthesizing strategy alignment. Strategy section content comes verbatim from strategy.tracks[] and the spec's ## Strategy Alignment block. The body never invents alignment claims.
• Inventing glossary terms. Glossary section content comes from glossary.changes[]. New terms / renamed terms are surfaced only if the export reports them.
• Hallucinating mermaid relationships. Diagram nodes + edges come from real cross-module imports detected via git diff analysis (Phase 3 details in the mermaid-rules.md ref file). The skill never adds "I think module X also imports Y" edges.

When data is missing, the body says so honestly (e.g. *No decision-track memory entries for this spec. Surface decisions in PR review comments if needed.*) rather than confabulating content. Honest "unclear" beats plausible "wrong".

Forbidden

• Ralph-blocking the skill. This skill is the autonomous-loop terminus per spec R24. Detect Ralph but proceed (with --draft forced). A FLOW_RALPH/REVIEW_RECEIPT_PATH exit-2 guard at the top of the skill has broken this.
• Re-adding a confirm gate. make-pr creates the PR without prompting. A run that asks "do you want to create it?" before push has broken this — the escape hatch is --dry-run, not a question.
• Pushing or creating PRs in --dry-run mode. Phase 4 short-circuits before any git push or gh pr create. The body lands on stdout only.
• Squashing the existing-PR check. A bare gh pr view --json url 2>/dev/null returns rc=0 for CLOSED and MERGED PRs as readily as OPEN. Filter .state == "OPEN" via jq (validated empirically during fn-42 spike). Closed/merged PRs on a reused branch must NOT trigger refusal.
• Manual git push workflows when gh is missing. When gh isn't installed or authenticated, surface the install / gh auth login instructions and exit. Don't try to fall back to half-baked PR creation.
• Writing memory entries without --memory. Default off. The user opts in for structurally-significant specs — every-PR memory inflation is the failure mode this gate prevents.
• Quoting raw diff content in the body. See hallucination guardrails — the body describes the diff, never copies code.
• Calling gh pr merge. Out of scope. The skill creates and exits; merge is a human decision.
• git add -A (or any broad stage) for a legacy PR-artifact commit. A supported current v1 lens stays local-only so it cannot stale its own head-bound input. The fallback committed path stages exactly .flow/artifacts/<spec-id>/pr.html with the fixed message chore(flow): pr artifact <spec-id> — unrelated working-tree changes are not make-pr's concern.
• Opening a Lavish session or running lavish-axi poll from make-pr. The PR artifact is a read-only review instrument — no annotate loop, interactive or autonomous. Review conversation belongs to the code host.
• Emitting an artifact blob link that can 404. Gitignored .flow/artifacts/ → local-open guidance only; committed mode links only after the narrow artifact commit landed on the branch being pushed.

Workflow

Execute the phases in workflow.md in order:

0. Pre-flight — gh installed + authenticated; resolve spec id (arg or branch-match); base-branch detection cascade; branch validity (HEAD ahead of base); all tasks done (warn + proceed as draft if not — no prompt; Ralph exits 2); existing-PR refusal filtered on .state == "OPEN". Detects Ralph environment for downstream phases.
1. Gather inputs — single call to flowctl spec export-cognitive-aid <spec-id> --base <ref> --json; parse the structured payload (spec / tasks / memory / glossary / strategy / diff / reviews). 1.5. Structured PR cognitive aid (ALWAYS — read pr-cognitive-aid.md and execute it) — the existing host composes one grounded v1 walkthrough from the export-time head. flowctl validates/persists/selects/renders it before optional HTML and final body creation. Not optional and not HTML-gated — only 1.5b is opt-in. No extra model call. Supported current artifacts render compact/full deterministically; stale/invalid artifacts select the existing fallback without mixing fields. The post-creation makePr tracker facade remains unchanged. 1.5b. HTML render lens (opt-in) — only when artifacts.html.enabled is true AND not --dry-run: generate .flow/artifacts/<spec-id>/pr.html (read-only review instrument per the shared disclosure reference plugins/flow-next/references/html-artifacts.md §5 — diff-derived, R-ID-verified with flagged mismatch rows). A supported current v1 object is embedded as the exact HTML-safe semantic carrier and stays local-only so HEAD remains equal to its headSha; only the visibly labeled legacy fallback may use the narrow committed-artifact path. Record the render-lens line for the body summary block. Never opens a Lavish session or polls (interactive AND autonomous). Failure is non-fatal — one stderr note, PR proceeds. With the mode off/unset there is zero artifact-related behavior or output beyond the single config read.
2. Render body — TL;DR, R-ID coverage table, Critical changes, How to review this PR (trust-calibration coaching block), Review plan (risk-ranked Must review / Spot-check / Safe to skim with a ≤~30% focus budget), Decisions made, Memory left behind, Glossary/strategy notes, Open items, footer breadcrumb. Sections without content are omitted (never empty placeholder headings).
3. Mermaid generation — gated by 5 trigger conditions (cross-module imports, public interface changes, new/removed top-level dirs, high fan-out spec). Hard caps: 3 diagrams, 12 nodes, 25 edges, 12K characters. Each codefence preceded by a 3-5 sentence plain-language prose summary (load-bearing for forges that don't render mermaid). Validates each codefence against the mermaid-rules.md §6 checklist (reserved words, escape patterns, no emoji / MathJax, no inheritance cycles) before emitting. Skipped under --no-mermaid or when no triggers fire / a skip rule applies (pure-additive single-module diff <50 LOC, flat-layout repo).
4. Push + create PR — git push -u origin HEAD, then gh pr create --title --body. Draft when OPEN_ITEMS_COUNT > 0 OR Ralph OR --draft; ready when --ready. --dry-run short-circuits before push.
5. Output + footer — emit PR URL on success; print breadcrumb (Generated by /flow-next:make-pr from <spec-id> against <base>) plus the invisible machine marker <!-- flow-next:make-pr spec=<spec-id> base=<base-ref> --> land's authorship probe keys on (issue #274); optionally write knowledge/architecture-patterns/ memory entry under --memory.