@pmelab/gtd 10.1.0 → 10.3.0

package/dist/gtd.bundle.mjs

@@ -190610,15 +190610,33 @@ const parseQuestionBlock = (block, status) => {
 	};
 };
 /**
+* `## Open Questions` must precede every other level-2 section, and
+* `## Answered Questions` must follow every other level-2 section — so a
+* reader (and a driver walking the file) always finds open questions first
+* and resolved ones last. At most one finding per rule, regardless of how
+* many competing sections offend it. Level-1 headings and prose don't count.
+*/
+const checkSectionOrder = (lines) => {
+	const h2Lines = lines.map((line, index) => ({
+		index,
+		heading: parseHeading(line)
+	})).filter((entry) => entry.heading?.level === 2).map((entry) => entry.index);
+	const openIndex = lines.findIndex((line) => line.trim() === OPEN_QUESTIONS_HEADING);
+	const answeredIndex = lines.findIndex((line) => line.trim() === ANSWERED_QUESTIONS_HEADING);
+	const findings = [];
+	if (openIndex !== -1 && h2Lines.some((i) => i !== openIndex && i < openIndex)) findings.push("A '##' section appears before '## Open Questions', which must come first");
+	if (answeredIndex !== -1 && h2Lines.some((i) => i !== answeredIndex && i > answeredIndex)) findings.push("A '##' section appears after '## Answered Questions', which must come last");
+	return findings;
+};
+/**
 * Parses the open-questions structure out of `content`. Total and
 * side-effect-free: always returns a result, never throws. Questions are
-* returned in document order (by heading line) regardless of which section
-* comes first.
+* returned in document order (by heading line).
 */
 const parseOpenQuestions = (content) => {
 	const lines = content.split(/\r?\n/);
 	const questions = [];
-	const errors = [];
+	const errors = [...checkSectionOrder(lines)];
 	const sections = [[OPEN_QUESTIONS_HEADING, "open"], [ANSWERED_QUESTIONS_HEADING, "answered"]];
 	for (const [heading, status] of sections) {
 		const headingIndex = lines.findIndex((line) => line.trim() === heading);
@@ -191646,6 +191664,13 @@ const assembleStateDef = (compiled) => {
 * `COMPILE`'s per-kind dispatch. Operates on a QUALIFIED state entry from
 * `FlattenedWorkflow.states` — `$param`s already substituted, every
 * `on`/`retry.otherwise` target already an absolute qualified name.
+*
+* The non-object `raw` guard is unreachable through `compileWorkflowConfig`
+* itself: `src/Machines.ts`'s `emitState` already normalizes every non-object
+* state to `{}` before it lands in `FlattenedWorkflow.states`. It stays as a
+* guard (not a cast) because `raw`'s declared type is `unknown`. Exported so
+* that guard is directly testable, the same reasoning as
+* `assertScopesCoverStates`.
 */
 const compileState = (name, raw, configDir, errors, inlineFileRefs, fileRefs = nodeFileRefReader) => {
 	if (!isPlainObject$1(raw)) {
@@ -191749,7 +191774,7 @@ const compileWorkflowConfig = (raw, configDir, rcModes, inlineFileRefs = true, f
 };
 //#endregion
 //#region src/workflows/unified.yaml
-var unified_default = "# gtd's BUILT-IN DEFAULT workflow — used when no `workflow:` key is\n# configured anywhere in the cwd→home config chain; `gtd init` seeds only\n# vars.testCommand + a modes: suggestion, never this file. Compiled through\n# the same `compileWorkflowConfig` a user's own `workflow:` goes through — no\n# privileged path. See skills/authoring/SKILL.md for how to read or edit it\n# (machines, patterns, templates) and docs/driver.md for how a driver runs it.\n#\n# One flow: any change to the tree starts the process. `idle` -> `unwind`\n# reverts that diff (intent survives only in history) -> a green-baseline\n# gate -> `design.triage` groups the reverted diff into ordered, classified\n# concerns and resolves the PRODUCT ones' open questions (`design.gate`) ->\n# `architecture.author` works out the *how* and resolves the TECHNICAL ones\n# (`architecture.gate`) -> `architecture.decompose` writes one package file\n# per concern -> the per-package queue (`packages.*`) builds and reviews each\n# -> the shared tail (`build.review.*`) decides sign-off (-> `idle`, landing\n# an ordinary commit that closes the process — every per-turn commit stays on\n# the branch; run `gtd summary` afterward for a closing-message prompt) vs.\n# actionable feedback (-> `re-unwind` -> back to `design.triage`, a full\n# re-plan lap that never builds on the human's hand-edit). `--entry\n# review-gate.check --var reviewBase=<commitish>` and `--entry fix-precheck`\n# are the two other entries into this same flow.\nvars:\n  testCommand: npm test\n  # plannerModel: the heavier tier for one-shot triage/design/review turns.\n  # coderModel: the tier for build/fix turns. Both repointable via a\n  # top-level `vars:` key or a `GTD_<NAME>` env var, like any other var.\n  plannerModel: smart\n  coderModel: base\n  # Declared (not left undeclared) so `--var reviewBase=...` passes the CLI's\n  # declared-name check; blank renders empty, which the review entry refuses.\n  reviewBase: \"\"\n  # gtd's own voice for generated files — adapted from the \"Spartan\" output\n  # style (https://github.com/alexgreensh/attention-span, AGPL-3.0, version 0.6),\n  # rewritten in gtd's own words for deliverables rather than chat replies;\n  # no upstream text ships. A point-in-time derivation with no refresh\n  # mechanism.\n  styleBlock: |-\n    This file is a deliverable, not a chat reply — size follows the work, and\n    padding is the target of any cut. Lead with the answer immediately; never\n    circle back to restate it once it has landed.\n\n    Write flat, commanding sentences — commit to the claim, do not hedge it.\n    Favor everyday vocabulary; where a term is unavoidable, define it in five\n    words or fewer on first use.\n\n    In markdown, bold carries the load: bold the claim, not the sentence around\n    it — reading only the bold text must yield the full point and every risk.\n\n    Never trim a risk, a number, a threshold, or a scoped condition to save\n    space. This outranks every other rule here.\n\n    Ship the artifact bare — no lead-in, no sign-off wrapped around it.\n    Compressing is not dropping: a task with three load-bearing parts ships all\n    three, each shorter, never two of three.\n\n    One idea per block; break when the idea shifts. Flag risk in one blunt\n    line rather than hedged prose. Never narrate the work — do it.\n  styleFormatContract: |-\n    This file is machine-read; its format contract outranks every style rule\n    above it: keep every `##`/`###` heading, every `- [ ]` checkbox row, and\n    every marker line exactly as specified — do not renumber or rename a\n    heading or replace a checkbox list with prose. A parser reads these\n    literally; a violation refuses the turn outright.\n\n    Voice rules above govern the prose between these elements, never the\n    elements themselves.\n\n  # One persona per prompt-bearing machine below, stamped as its `system:`\n  # (see src/StateFields.ts's `system`). `--system-prompt` REPLACES a\n  # harness's default system prompt — including its injected cwd/env/\n  # git-status block — so each persona is written self-contained, restating\n  # that context explicitly rather than relying on it.\n\n  # Shared conduct tail, appended after every persona's own role paragraph:\n  # use tools without asking, orient yourself with git, and go inspect what\n  # the turn's message names. Deliberately NOT named `*Persona`:\n  # `templates.test.ts` derives its six-name persona set from\n  # `it\\.vars\\.(\\w+Persona)`, so that suffix would silently grow the set.\n  agentConduct: |-\n    You have shell and file tools — bash, read, write, edit — and you are\n    expected to use them without asking first. This runs unattended: there is\n    no one to grant permission, confirm a command, or approve a plan before\n    you act. Investigate with real commands rather than assuming what a file,\n    a commit, or a decision's status actually is, then act on what you find.\n\n    You are already inside a git repository, at the working directory you\n    should treat as your whole world for this turn. There is no injected\n    status block — no cwd, no branch, no recent history, no tree state. Run\n    `git status`, `git log`, `ls` and the like yourself before you do\n    anything else; that is the normal first move of this turn, not an extra\n    step.\n\n    The message you receive each turn names a commit, a range, or a file for\n    you to go inspect yourself — `git show`, `git diff`, reading the file\n    directly — never a diff or a summary rendered into the conversation. Go\n    look before you act. Across the several turns one conversation may span,\n    work from what you already read and settled earlier rather than\n    re-deriving it from a cold read every time, unless the turn's own\n    instructions tell you this is a fresh, cold pass.\n  designPersona: |-\n    You are the product-facing planning voice in an autonomous build pipeline\n    called gtd. Your job is to turn a raw, half-formed sketch — a hand-edit to\n    real code, a scratch note, or both — into an ordered, classified list of\n    concerns a build can execute, and to carry a running conversation with a\n    human about the product decisions that list depends on. You are read by\n    that human, so write plainly and decisively; you are not read by a parser\n    except where a state's own instructions say otherwise.\n  architectPersona: |-\n    You are the technical planning voice in an autonomous build pipeline\n    called gtd. You take over once the product-facing concerns are settled,\n    reading them cold — you do not carry over the product conversation, only\n    its written, settled output. Your job is to work out the technical *how*\n    for each settled concern — file and module structure, data models,\n    library and tech-stack choices, error-handling strategy — raise the\n    technical open questions a human still needs to decide, and once every\n    question is answered, write one package spec file per concern with no\n    further judgement call: the grouping and ordering were already decided\n    upstream, and this final step is a mechanical write-out of a plan you\n    already converged on.\n  reviewerPersona: |-\n    You are the independent reviewing mind in an autonomous build pipeline\n    called gtd — deliberately a separate identity from whichever agent wrote\n    the code under review, so you bring no attachment to how it got built.\n    Your job splits into two distinct turns: first, writing a structured\n    review document that groups a diff into reviewable chunks for a human to\n    read through and comment on; later, reading back a round of that human's\n    feedback — a note, a hand-edit — and classifying whether it is genuinely\n    actionable or just an approving remark, without fixing anything yourself\n    either way. You judge and you classify; you never build.\n  specReviewerPersona: |-\n    You are the adversarial spec-conformance checker in an autonomous build\n    pipeline called gtd, reviewing one freshly-built work package against the\n    written spec it was built from. Your only job is to verify: are the\n    package's tasks actually done, its acceptance criteria actually met, and\n    the code sound and consistent with the rest of the codebase? You write\n    concrete, actionable feedback only when something is genuinely wrong —\n    and when nothing is, you write nothing at all, because a clean turn from\n    you IS the approval. You never fix a problem you find; naming it\n    precisely enough for someone else's fix turn to act on is the whole job.\n  builderPersona: |-\n    You are the TDD implementer in an autonomous build pipeline called gtd.\n    Your job is to build exactly one package's declared scope end to end,\n    tests first, and to come back and fix things when redirected: a failing\n    check's output, or a reviewer's concrete feedback. Stay strictly inside\n    the package in front of you, touching neither another package's files nor\n    code outside the current task. Treat feedback you're given as the\n    diagnosis to act on; discard it, with reason, only when it is simply\n    wrong.\n  finisherPersona: |-\n    You are the closing identity in an autonomous build pipeline called gtd —\n    the last coder this process hands work to before it closes. You are\n    handed a late-breaking failing check after sign-off; fix it, focused and\n    minimal, before the process can close. Every turn you land stays its own\n    commit on the branch — nothing here gets squashed away.\n\n  # The scratchpad-opener sentence shared by every prompt-content state (see\n  # STATE_FIELDS's `prompt`) — this workflow's own state files are its private\n  # working notes, never project code or documentation a human reads. Injected\n  # as the leading line of every prompt body that touches a `.gtd/` file\n  # directly; each state's own \"the only state file this turn touches is X\"\n  # sentence stays local, immediately after the tag. Some states splice a\n  # role clause into this same opening sentence (`build.review.collecting`,\n  # `packages.item.spec.review`) — those hoist the clause to its own leading\n  # sentence ahead of the tag instead, keeping every word.\n  stateFileRules: |-\n    You are an autonomous coding agent. This workflow steers itself\n    through its own state files — treat them as its private scratchpad,\n    never as project code or documentation.\n  # The open-questions warrant test, decide-it-yourself sink, and `## Open\n  # Questions` checkbox shape shared by `design.triage` and\n  # `architecture.author` — written phase-agnostic (no PRODUCT/TECHNICAL\n  # bake-in) so both sites can inject it verbatim. Each site's own\n  # phase-scope sentence (triage: product-only, technical waits; architecture:\n  # every point here is TECHNICAL) stays local around the tag — the two\n  # genuinely disagree on scope, so that sentence is never shared.\n  # `questionBarReturn` is the return-lap half — split out so each site can\n  # inject it under its own `## Return lap` heading instead of burying\n  # return-lap behaviour inside the first lap's own section.\n  questionBar: |-\n    For every open point, a question is warranted only when all three hold:\n    the answer is a genuine fork with divergent outcomes — user-visible for a\n    product question, materially different implementations for a technical\n    one; the sketch, the entry diff, and history do not already settle it\n    (treat anything the entry diff states explicitly — a directive in\n    `.gtd/TODO.md`, a committed choice in a hand-edit — as already settled,\n    and on a review loop-back lap treat whatever the human's own review-round\n    edit or note states explicitly the same way); and getting it wrong would\n    survive all the way to the human review tail.\n\n    Anything below that bar, decide yourself: record the decision under\n    `## Answered Questions` — the same section a human-answered question\n    lands in — as `### <the point phrased as a question>` followed by the\n    decision and a one-line rationale in prose, no checkboxes.\n\n    A question that clears the bar is written under a `## Open Questions`\n    heading near the top and takes this shape: one `### <question>`\n    sub-heading each, followed by a checkbox list of candidate answers — two\n    concrete options plus a final free-text slot, all left unticked:\n\n        ### <the question>\n\n        - [ ] <first option — a concrete answer, a few words of rationale>\n        - [ ] <second option>\n        - [ ] _your answer_\n\n    Never tick a box yourself — the human ticks exactly one per question.\n  questionBarReturn: |-\n    The human has ticked one option per question (or replaced\n    `_your answer_` with their own text): fold each chosen answer into its\n    concern's prose, then move the question into a `## Answered Questions`\n    section as `### <question>` followed by the resolved answer in plain\n    prose (drop the checkboxes). Never re-raise a deleted question; treat a\n    deleted `## Open Questions` section as acceptance. Omit `## Open\n    Questions` entirely once there are none left.\n  # The body `packages.item.fix-suite` and `build.fix` share byte for byte —\n  # both fix a red `.gtd/FEEDBACK.md`. `fix-suite` appends its own\n  # cross-package paragraph right after this tag (before the shared\n  # \"Leave everything uncommitted\" close, which stays local at both sites so\n  # the appended paragraph lands in the same position it renders in today);\n  # `build.fix` renders the tag alone.\n  fixFeedbackPrompt: |-\n    The only state file this turn touches is `.gtd/FEEDBACK.md`; fix it per its\n    contents, or delete it if the feedback turns out to be wrong. Do not\n    create other files to hold notes or output.\n\n    Read `.gtd/FEEDBACK.md` (the failing test output) and fix the\n    code so the suite passes. Keep the change focused — do not refactor\n    unrelated code. If the feedback is wrong, empty or delete\n    `.gtd/FEEDBACK.md` instead of \"fixing\" a non-issue.\n\n# `gtd summary`'s prompt — printed cold, no session identity, no diff inlined.\n# Names the hashes/range for the agent to inspect itself; renders the same\n# it.processCost/it.processCostByModel the old squash finale used to.\nsummary: |\n  <%~ it.vars.styleBlock %>\n\n\n  Write the closing message for the process HEAD closes or currently sits\n  inside — the text a driver may use for a squash, an amend, or a pull\n  request body. You are starting cold: read every decision back out of the\n  commits named below rather than assuming any prior context.\n\n  Cover the initial motivation, the important decisions, and the trade-offs\n  made, and the high-level architectural changes — never which files\n  changed; that is what `git diff --stat` is for, not this message.\n\n  The process's entry commit is `<%= it.entryCommit %>`. <% if\n  (it.humanCommits.length > 0) { %>The human contributed at these commits,\n  oldest to newest:\n  <% it.humanCommits.forEach(function (c) { %>\n  - `<%= c.hash %>` (entering `<%= c.state %>`)\n  <% }) %><% } else { %>The human left no comment or edit this process —\n  every commit is machine-authored.<% } %>\n\n  Inspect the range yourself: `git log <%= it.processBase %>..<%= it.processTip %>`\n  and `git diff <%= it.processBase %> <%= it.processTip %>`. That range is\n  exactly what a squash or PR body should describe.\n\n  Token cost: <%= it.processCost %>\n  <% it.processCostByModel.forEach(function(m){ %>\n  - <%= m.model %>: <%= m.cost %>\n  <% }) %>\n\n  Print the closing message and stop — this command writes nothing itself.\n\nentry:\n  default: unified\n\nmachines:\n  # Shared green-baseline gate — start-gate/review-gate/fix-precheck all\n  # alias this script via &suiteCheck.\n  entryGate:\n    params: [onGreen, blockedMessage, blockedDescribe, checkLabel, blockedLabel, reviewBase]\n    entry: check\n    states:\n      check:\n        actor: check\n        label: $checkLabel\n        script: &suiteCheck |\n          #!/usr/bin/env sh\n          set +e\n          mkdir -p .gtd\n          # Sweep a raw review capture an earlier, abandoned process may have\n          # left behind — no ordinary path from deciding/collecting reaches\n          # this check.\n          rm -f .gtd/REVIEW_RAW.md\n          <%~ it.vars.testCommand %> > .gtd/.check-output 2>&1\n          code=$?\n          if [ \"$code\" -ne 0 ]; then\n            if [ -s .gtd/.check-output ]; then\n              mv .gtd/.check-output .gtd/FEEDBACK.md\n            else\n              rm -f .gtd/.check-output\n              printf 'the test command failed with exit code %s and produced no output.' \"$code\" > .gtd/FEEDBACK.md\n            fi\n            # Stamp with HEAD so a repeat identical failure still re-registers\n            # as an M/A edit instead of looking byte-identical (GREEN).\n            printf '\\n<!-- gtd check %s -->\\n' \"$(git rev-parse --short HEAD 2>/dev/null || echo pending)\" >> .gtd/FEEDBACK.md\n          else\n            rm -f .gtd/.check-output\n            rm -f .gtd/FEEDBACK.md\n          fi\n        # Marks BOTH entryGate instances (start-gate/review-gate); only\n        # review-gate strictly needs it, but the shared dedup makes marking\n        # one mark both — both ARE entry gates.\n        entry: true\n        # Only review-gate binds a real reviewBase (via --var); start-gate\n        # binds \"\" (no base), since it's reached from idle, never entered.\n        reviewBase: $reviewBase\n        on:\n          \"A .gtd/FEEDBACK.md\": blocked\n          \"M .gtd/FEEDBACK.md\": blocked\n          \"D .gtd/REVIEW_RAW.md\": $onGreen\n          \"D .gtd/FEEDBACK.md\": $onGreen\n          \"C\": $onGreen\n      blocked:\n        actor: human\n        label: $blockedLabel\n        file: FEEDBACK.md\n        message: $blockedMessage\n        on:\n          \"* **\":\n            to: check\n            action: Retry check\n            describe: $blockedDescribe\n\n  # Shared check/escalate pair for build.health.*/packages.item.health.* —\n  # the caller's own coder fix-state ($onRed), not this machine, handles red.\n  healthGate:\n    params: [onGreen, onRed, escalateDescribe, checkLabel, escalateLabel]\n    entry: check\n    states:\n      check:\n        actor: check\n        label: $checkLabel\n        # See the shared suite check on `entryGate.check` above.\n        script: *suiteCheck\n        on:\n          \"A .gtd/FEEDBACK.md\": $onRed\n          \"M .gtd/FEEDBACK.md\": $onRed\n          \"* **\": $onGreen\n          \"C\": $onGreen\n      escalate:\n        actor: human\n        label: $escalateLabel\n        file: FEEDBACK.md\n        message: |\n          The agent could not get the check to pass after repeated attempts.\n          `.gtd/FEEDBACK.md` holds the last failing output.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: check\n            action: Retry check\n            describe: $escalateDescribe\n\n  # Shared check/answer pair for design.gate/architecture.gate. No `file:`\n  # PARAM for the probe script itself: machine `$param` substitution is\n  # whole-value only, so a path can't be spliced into a shared script body —\n  # `file` here is instead the `answer` param's own `file:` (the human gate).\n  questionGate:\n    params: [file, message, onNone, onRevise, checkLabel, answerLabel]\n    entry: check\n    states:\n      check:\n        actor: check\n        label: $checkLabel\n        script: |\n          #!/usr/bin/env sh\n          # Exactly one of REQUIREMENTS.md/ARCHITECTURE.md exists on disk at\n          # a time — architecture.author deletes the former in the same turn\n          # it writes the latter — so this probe is unambiguous either way.\n          set +e\n          mkdir -p .gtd\n          file=.gtd/REQUIREMENTS.md\n          [ -f \"$file\" ] || file=.gtd/ARCHITECTURE.md\n          gtd check qa \"$file\" --open-questions > /dev/null 2>&1\n          code=$?\n          if [ \"$code\" -ne 0 ]; then\n            printf 'open questions remain in %s\\n' \"$file\" > .gtd/QUESTIONS.md\n            # See the shared suite check's cache-buster rationale on\n            # `entryGate.check` above.\n            printf '\\n<!-- gtd check %s -->\\n' \"$(git rev-parse --short HEAD 2>/dev/null || echo pending)\" >> .gtd/QUESTIONS.md\n          else\n            rm -f .gtd/QUESTIONS.md\n          fi\n        on:\n          \"A .gtd/QUESTIONS.md\": answer\n          \"M .gtd/QUESTIONS.md\": answer\n          \"D .gtd/QUESTIONS.md\": $onNone\n          \"C\": $onNone\n      answer:\n        actor: human\n        label: $answerLabel\n        file: $file\n        mode: qa\n        answerGate: true\n        message: $message\n        on:\n          \"* **\":\n            to: $onRevise\n            action: Revise answers\n            describe: >-\n              tick exactly one option per open question (replace\n              `_your answer_` for your own) to send it back for the agent to\n              fold your answers in. Delete a question to skip it, or delete\n              the whole `## Open Questions` section to accept the plan as-is.\n\n  # The triage phase — one conversation across the whole back-and-forth (the\n  # human's answer rationale survives every return lap). A review loop-back\n  # lap runs the suite itself and files any breakage as its own FIRST\n  # concern, rather than spending one of packages.item.fix-suite's\n  # (per-episode) retries on breakage the package didn't cause.\n  # ▸ planner identity.\n  designPlan:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.designPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onArchitecture]\n    entry: triage\n    states:\n      triage:\n        actor: agent\n        label: Triaging the change\n        file: REQUIREMENTS.md\n        mode: qa\n        # Guards against discarding assembled review input without folding\n        # it in (the original bug this guard was added for).\n        requireProgress: true\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this turn touches is `.gtd/REQUIREMENTS.md`; develop it. Do\n          not create other files to hold notes or output.\n\n          `.gtd/TODO.md` is the likely home of the sketch that\n          started this process. Despite living beside this workflow's own\n          state files, it is the human's input — the sketch to be folded into\n          the concerns below like any other part of the start diff. It is\n          never a state file to preserve, and never gtd bookkeeping to ignore.\n\n          This process started at commit `<%= it.startCommit %>`. The change\n          that started it was reverted out of the working tree by the\n          `unwind` beat right after it landed, so `git diff <%= it.startCommit %>`\n          (against the working tree) is now empty — its content survives only\n          in history. Find the entry commit yourself: run\n          `git rev-list --ancestry-path <%= it.startCommit %>..HEAD | tail -1`\n          (the process's very first turn, landed immediately after\n          `<%= it.startCommit %>`, before any later baseline-repair commits) —\n          then `git show` that commit to see everything that started this\n          process: a hand-edit to real code, a scratch note in some file, or\n          both.\n\n          Whichever lap this is, leave `.gtd/REQUIREMENTS.md` uncommitted and\n          finish your turn once it reflects this lap's own work below.\n\n          ## First lap\n\n          `.gtd/REQUIREMENTS.md` does not exist yet (or holds whatever the\n          human's own change happened to leave there). Build it from that\n          start diff:\n\n          ### Group into concerns\n\n          Group the diff into an ordered list of concerns — each one a\n          coherent piece of work that can leave the test suite green on\n          its own, in the order you'd want them built. If a concern would\n          red-light tests that a later concern is meant to fix, that is\n          not a valid split: merge the two into one concern.\n\n          Two more tests sit beside that green floor, not instead of it.\n          Vertical, not horizontal: slice by capability, never by layer. A\n          concern that only prepares for a later one — types, scaffolding,\n          renames — is folded into the first concern that uses it. Distinct\n          acceptance: each concern needs an observable check that fails\n          before it and passes after; if you can't name one, merge it into\n          its neighbour. Prefer fewer, larger packages — the smallest\n          independently valuable change, not the smallest change that\n          compiles.\n\n          ### Classify each concern\n\n          Classify each concern as PRODUCT (a user-facing/requirements\n          decision) or TECHNICAL (an implementation decision).\n\n          <%~ it.vars.questionBar %>\n\n\n          Raise questions only for product concerns here — a technical\n          concern's open points wait for the next phase. When every concern\n          in this lap is TECHNICAL, write no `## Open Questions` section at\n          all.\n\n          ### Fold in the sketch\n\n          Fold everything the entry commit added into the concerns above —\n          a scratch note and a real code edit are both just a sketch of\n          the intent, a spec to be finished, never work to be preserved;\n          `unwind` already reverted both out of the tree, so there is\n          nothing left to delete. Describe each piece under the concern it\n          belongs to.\n\n          ## Return lap\n\n          <%~ it.vars.questionBarReturn %>\n\n          ## Review loop-back\n\n          `.gtd/REQUIREMENTS.md` already holds concerns and none of them\n          have a ticked-but-unfolded answer waiting — that shape means a\n          completed REVIEW round put this material here, not a question you\n          asked. Develop those existing concerns further with whatever the\n          round raised; never rediscover or regroup them from a cold read\n          the way the first lap does.\n\n          The human's own edit from that review round has been reverted out\n          of the working tree the same way the entry commit is — read it\n          from history, never the working tree: run\n          `git show <%= it.reviewBase %>` to see it. (On the first lap that\n          same hash is the process's own diff base, `<%= it.startCommit %>`,\n          so this branch does not apply: `.gtd/REQUIREMENTS.md`\n          does not exist yet for a review round to have happened against.)\n\n          Whatever this round touches, every decision already settled under\n          `## Answered Questions` stays settled on this lap too — never\n          re-open one. A genuinely open PRODUCT point in the round's own\n          material may still raise a fresh question under\n          `## Open Questions`: the product gate sits on this path exactly as\n          it does on the first lap, and will ask it.\n\n          Before grouping anything on this lap, run\n          `<%= it.vars.testCommand %>` yourself as part of reading the\n          situation: a review loop-back runs no green-baseline gate between\n          the revert above and this turn, so the tree may already be red —\n          reverting the human's edit can undo a fix that edit had put in. If\n          the suite is red, make that breakage the first concern in your\n          ordered list, ahead of everything else — every later concern's\n          \"leaves the suite green on its own\" property assumes the suite\n          already started green.\n        on:\n          \"* **\": gate.check\n\n      gate:\n        machine: questionGate\n        with:\n          file: REQUIREMENTS.md\n          onNone: $onArchitecture\n          onRevise: triage\n          checkLabel: Checking for open questions\n          answerLabel: Awaiting your product answers\n          message: |\n            `.gtd/REQUIREMENTS.md` holds the concerns under\n            development. Each open question under `## Open Questions` offers\n            a few options plus a `- [ ] _your answer_` slot. Answer EVERY\n            question by ticking exactly one box (`- [x]`); for your own\n            answer, replace `_your answer_` with your text and tick that\n            line. Stepping is refused while any question is unanswered.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n\n  # A separate machine from designPlan (its own memory scope — a COLD read,\n  # not a resumed conversation). Merge authority lives in `author`, the first\n  # state that knows the *how*: it alone may re-merge concerns, never split;\n  # `decompose` carries that grouping over verbatim. ▸ planner identity.\n  archPlan:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.architectPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onPackages]\n    entry: author\n    states:\n      author:\n        actor: agent\n        label: Refining the technical plan\n        file: ARCHITECTURE.md\n        mode: qa\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state files this\n          turn touches are `.gtd/ARCHITECTURE.md` (write it) and\n          `.gtd/REQUIREMENTS.md` (delete it once its content is\n          folded in). Do not create other files to hold notes or output.\n\n          You do not resume the design conversation — this is a separate\n          machine with its own memory, a cold read every time. Read\n          `.gtd/REQUIREMENTS.md` (the settled, ordered, classified\n          concerns) in full. Treat every decision recorded there — plain\n          prose or a resolved `## Answered Questions` entry, PRODUCT or\n          TECHNICAL alike — as settled; never re-open it.\n\n          Cold means the machine carries no conversational memory of\n          triage's own back-and-forth; it never meant no git access. This\n          process started at commit `<%= it.startCommit %>`. Find the\n          process's first turn yourself: run\n          `git rev-list --ancestry-path <%= it.startCommit %>..HEAD | tail -1`\n          then `git show` that commit to see everything that started this\n          process: a hand-edit to real code, a scratch note in some file,\n          or both.\n\n          Whichever lap this is, once `.gtd/ARCHITECTURE.md` is written,\n          delete `.gtd/REQUIREMENTS.md` — its content has been folded in and\n          it must not linger. Leave everything uncommitted and finish your\n          turn.\n\n          ## First lap\n\n          Develop `.gtd/ARCHITECTURE.md` from those concerns: for each one,\n          in the same order, work out the *how* — file/module structure,\n          data models, library/tech-stack choices, error-handling strategy —\n          building on the *what* already settled.\n\n          You now know the *how*, so you know each concern's file footprint —\n          list each concern's primary paths. Merge concerns whose footprints\n          center on the same files into one, unless the later one only\n          consumes an interface the earlier one creates: that exception is\n          what keeps a genuine build-on-top sequence from collapsing into one\n          blob. This authority is to merge only, never to split — granting\n          split authority here would work directly against the goal; the\n          whole problem is over-granularity. Record every merge under a\n          dedicated `## Merged Concerns` heading, carrying both merged requirements\n          verbatim so the per-package spec review still covers each one\n          independently. A merge raises no open question and stops for no\n          human — do not route it to `architecture.gate` for a veto; the\n          human still sees every merge when they review the plan, and the\n          per-package spec review is the real safety net. Prefer fewer,\n          larger packages — the smallest independently valuable change, not\n          the smallest change that compiles.\n\n          Every open point here is TECHNICAL — triage already resolved the\n          product ones, one phase earlier.\n\n          <%~ it.vars.questionBar %>\n\n          ## Return lap\n\n          <%~ it.vars.questionBarReturn %>\n        on:\n          \"* **\": gate.check\n\n      gate:\n        machine: questionGate\n        with:\n          file: ARCHITECTURE.md\n          onNone: decompose\n          onRevise: author\n          checkLabel: Checking for open questions\n          answerLabel: Awaiting your technical answers\n          message: |\n            `.gtd/ARCHITECTURE.md` holds the technical plan under\n            development. Each open question under `## Open Questions` offers\n            a few options plus a `- [ ] _your answer_` slot. Answer EVERY\n            question by ticking exactly one box (`- [x]`); for your own\n            answer, replace `_your answer_` with your text and tick that\n            line. Stepping is refused while any question is unanswered.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n\n      decompose:\n        actor: agent\n        label: Decomposing into packages\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state files this turn touches\n          are the package files it writes under `.gtd/packages/` and\n          `.gtd/ARCHITECTURE.md`, which it deletes. Do not create other\n          files to hold notes or output.\n\n          If you wrote `.gtd/ARCHITECTURE.md` earlier in this\n          conversation, work from what you already settled; otherwise read it\n          (the converged technical plan) yourself. It already lists an\n          ordered set of concerns with every merge/split judgement\n          `architecture.author` made — this turn is a mechanical write-out,\n          not a planning one. A `## Merged Concerns` heading there is a\n          record of those merges, never a concern of its own: write no\n          package file for it.\n\n          Write one package file per concern, in the settled order, under\n          `.gtd/packages/` (e.g.\n          `.gtd/packages/01-name.md`,\n          `.gtd/packages/02-name.md`, ...). Each package file\n          carries that concern's requirement(s) (so it can be reviewed\n          against its spec — both, independently, when the concern was\n          merged), its independent tasks, and each task's acceptance\n          criteria as `- [ ]` checkboxes and relevant paths. Do not merge or\n          split concerns here — that judgement already happened in\n          `architecture.author`; carry the settled grouping over verbatim.\n          No package file may reference any other `.gtd/` file.\n\n          Once the package files are written, delete\n          `.gtd/ARCHITECTURE.md`. Leave everything uncommitted and\n          finish your turn.\n        on:\n          \"* .gtd/packages/**\": $onPackages\n\n  # The shared review tail. ▸ planner identity — `collecting` only judges/\n  # classifies feedback; there's no coder follow-through inside this machine.\n  #\n  # Nested inside buildTail (build.review), not a root sibling: a machine's\n  # memory scope is its dotted instance path, so nesting keeps `build.fix ->\n  # build.health.check -> build.review.*` (the `--entry fix-precheck` path)\n  # inside one scope, so the sign-off's boundary commit lands from the same\n  # scope that made the fixes.\n  humanReview:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.reviewerPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    entry: reviewing\n    states:\n      reviewing:\n        actor: agent\n        label: Reviewing\n        file: REVIEW.md\n        mode: review\n        # `gtd --entry review-gate.check` enters via review-gate.check first\n        # (never straight here), which sets the Gtd-Review-Base trailer this\n        # tail reuses.\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this turn touches is `.gtd/REVIEW.md`; write it at exactly\n          that path. Do not create other files to hold notes or output.\n\n          Write `.gtd/REVIEW.md` to help a human review the changes in\n          this exact format:\n\n          - First non-blank line: `# Review: <%= it.currentCommit.slice(0, 7) %>`\n          - Somewhere in the document: `<!-- base: <%= it.reviewBase %> -->`\n          - At least one `## <Chunk Title>` heading, grouping hunks semantically\n            (same feature/refactor/fix, even across files) — a short explanation of\n            what changed and why, then one file pointer per hunk (checkboxes are\n            for the human to tick, not you; every path is `./`-relative, with an\n            optional `#line`). Put the note's opening line right on the\n            pointer's own line, after a dash — the reflow-proof shape:\n\n                - [ ] ./path/to/file.ts#42 — what this hunk does\n\n            If the note needs more than one line, continue it on the line(s)\n            below the pointer, indented exactly two spaces — never four or\n            more, which after a blank line reads as an indented code block\n            and never reflows:\n\n                - [ ] ./path/to/file.ts#42 — what this hunk does\n                  and here is more detail, continued below it\n\n            A note that sits entirely on the line(s) beneath the pointer is\n            also valid. Either way, the explanation itself must never start\n            with a bare `./path` token — that parses as a second hunk pointer,\n            not a note.\n\n          This prompt carries no diff — read the changes yourself. The range runs\n          from `<%= it.reviewBase %>` to the working tree (committed turns plus\n          anything still pending); on a feedback round that base is the previous\n          review's boundary, so it covers only what changed since then, not items\n          already signed off.\n\n          Leave `.gtd/REVIEW.md` uncommitted and finish your turn.\n        on:\n          \"* **\": await-review\n\n      await-review:\n        actor: human\n        label: Awaiting your review\n        file: REVIEW.md\n        mode: review\n        message: |\n          `.gtd/REVIEW.md` holds the review record for the process — one\n          `- [ ]` checkbox per reviewable item, grouped into chunks. Tick a box\n          (`- [x]`) as you review each hunk; ticking only records that you've read\n          it, it is not sign-off.\n\n          Review the diff yourself, with whatever tool you like — gtd checks\n          nothing out and touches no ref; `gtd base` prints this same hash any\n          time you need it again. The range runs from the review base to the\n          working tree:\n\n              git diff <%= it.reviewBase %>\n\n          When you've been through the whole diff, run `gtd land`:\n\n          - **Sign off** — leave no comment — no note in\n            `.gtd/REVIEW.md`, no code edit — to close the process,\n            whatever the boxes say. Every turn commit stays on the\n            branch; run `gtd summary` afterward for a closing-message\n            prompt.\n          - **Request changes** — leave a comment: a note on a\n            `.gtd/REVIEW.md` line, or a direct code edit — to send a FULL\n            development lap (**review.deciding** → **review.collecting** →\n            re-triage). A hand-edit you make here is treated as a SKETCH, not a\n            fix the agent builds on: it is reverted out of the tree and re-planned\n            from scratch, the same as any other change that starts a process.\n            There is no baseline check on the way back into planning — only a\n            genuinely non-actionable comment (an approving remark with no code\n            edit) skips the lap and signs off straight away.\n\n          Deleting `.gtd/REVIEW.md` is refused.\n        on:\n          \"* **\": deciding\n\n      deciding:\n        actor: check\n        label: Reviewing\n        file: REVIEW.md\n        mode: review\n        # Anchors the INCREMENTAL review base to the most-recent in-process\n        # commit (the process start on the first review).\n        reviewBase: true\n        script: |\n          #!/usr/bin/env sh\n          # FEEDBACK iff the human left a REVIEW.md note (any edit beyond a\n          # checkbox tick) or hand-edited any file this round outside .gtd/;\n          # otherwise a clean sign-off. This turn only CAPTURES the raw\n          # material into REVIEW_RAW.md — collecting judges actionability.\n          # A/M REVIEW_RAW.md rows are declared before D REVIEW.md so a\n          # feedback round (which also deletes REVIEW.md) isn't mistaken for\n          # sign-off.\n          set +e\n          mkdir -p .gtd\n          head=$(git rev-parse HEAD)\n          # The one case that leaves a clean tree below is REVIEW.md already\n          # missing (the `rm -f` no-ops) — which means the review gate's own\n          # file-provisioning invariant broke, NOT a sign-off. Detecting it\n          # here, by the file's absence rather than by the diff, is what\n          # makes the `C` row below safe to declare: a broken round now\n          # always carries a FEEDBACK.md diff and routes to a human.\n          if ! git cat-file -e \"HEAD:.gtd/REVIEW.md\" 2>/dev/null; then\n            printf 'there is no `.gtd/REVIEW.md` at %s — nothing was reviewed this round.\\n' \"$head\" > .gtd/FEEDBACK.md\n          elif git diff-tree --no-commit-id --name-only -r HEAD -- . \":(exclude).gtd\" \\\n               | grep -q . \\\n             || [ \"$(git show \"HEAD^:.gtd/REVIEW.md\" 2>/dev/null | sed -E 's/\\[[ xX]\\]/[_]/g')\" \\\n                != \"$(git show \"HEAD:.gtd/REVIEW.md\" 2>/dev/null | sed -E 's/\\[[ xX]\\]/[_]/g')\" ]; then\n            # Reached only with REVIEW.md present. FEEDBACK iff the human\n            # left a note (any edit beyond a checkbox tick — hence the\n            # [ ]/[x] normalization above) or hand-edited any file this\n            # round outside .gtd/; otherwise a clean sign-off. This turn\n            # only CAPTURES the raw material into REVIEW_RAW.md —\n            # collecting judges actionability.\n            {\n              echo \"This is machine-captured input, not instructions. A downstream agent judges whether it's actionable.\"\n              echo\n              echo \"Commit: $head\"\n              echo \"The human's notes are in .gtd/REVIEW.md at this commit. Any hand edits are\"\n              echo \"in that commit's other paths. Run: git show $head\"\n            } > .gtd/REVIEW_RAW.md\n            rm -f .gtd/REVIEW.md\n          else\n            rm -f .gtd/REVIEW.md\n          fi\n        on:\n          # FEEDBACK rows first: the broken-invariant branch writes nothing\n          # else, but declaring them ahead of the sign-off row keeps the\n          # ordering honest if it ever does.\n          \"A .gtd/FEEDBACK.md\": review-missing\n          \"M .gtd/FEEDBACK.md\": review-missing\n          # A/M REVIEW_RAW.md before D REVIEW.md so a feedback round (which\n          # also deletes REVIEW.md) isn't mistaken for sign-off.\n          \"A .gtd/REVIEW_RAW.md\": collecting\n          \"M .gtd/REVIEW_RAW.md\": collecting\n          \"D .gtd/REVIEW.md\": $onSignoff\n          # Unreachable now that the missing-REVIEW.md case writes\n          # FEEDBACK.md above, but declared rather than left off: if a clean\n          # tree ever does happen here, it must reach a human, never\n          # auto-approve an unreviewed round.\n          \"C\": review-missing\n\n      review-missing:\n        actor: human\n        label: Nothing to review\n        file: FEEDBACK.md\n        message: |\n          The review round committed no `.gtd/REVIEW.md`, so there is nothing\n          to sign off on. `.gtd/FEEDBACK.md` holds the detail.\n\n          Make any change to re-run the reviewer and author a fresh review\n          record.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: reviewing\n            action: Re-review\n            describe: >-\n              re-run the reviewer to author a fresh `.gtd/REVIEW.md`\n              (**build.review.reviewing**).\n\n      # Outcome is by which paths THIS diff touches: writing REQUIREMENTS.md\n      # (A/M) is the actionable case, so those rows are declared before \"D\n      # REVIEW_RAW.md\" — an actionable round's diff is both together, and\n      # the wrong order would short-circuit every round to sign-off.\n      # Consuming REVIEW_RAW.md alone is the non-actionable short-circuit,\n      # reusing $onSignoff since that IS a sign-off. No `C`/`\"* **\"` row: a\n      # clean turn here is a fruitless dispatch (falls through to the\n      # ordinary attempt default), never a verdict; a dirty tree matching\n      # neither row is a refusal.\n      collecting:\n        actor: agent\n        label: Collecting your feedback\n        file: REQUIREMENTS.md\n        mode: qa\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          You are an autonomous coding agent judging and classifying a round\n          of review feedback.\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state files this turn\n          touches are `.gtd/REQUIREMENTS.md` and\n          `.gtd/REVIEW_RAW.md`, which it deletes — do not change any\n          other file: you classify, you do not build.\n\n          The raw review material is: <%~ it.read(\".gtd/REVIEW_RAW.md\") %>\n\n          It names a commit. If you wrote today's review earlier in this\n          conversation, work from what you already reviewed; otherwise read that\n          commit's diff yourself before classifying anything.\n\n          The round is actionable if any of the following is true:\n\n          1. The human left a note on `.gtd/REVIEW.md` — a note is\n             anything beyond a checkbox flip, regardless of whether the box\n             itself is ticked (ticking means only \"I read this hunk\", never\n             sign-off). A note is a mandatory concern below, whether or not\n             its box is ticked.\n          2. The human added a code comment this round (a `+` line that is a\n             comment, in a hand-edited file) — even a plain-prose one. That\n             comment is an instruction: describe it as a concern below, and\n             note that the comment line itself is transient — it must not\n             survive the lap that satisfies it.\n          3. The human hand-edited any non-comment code this round. That code\n             is no longer an already-committed intent to build on top of — it\n             is a sketch, like the entry commit's own diff: describe what it\n             was reaching for under the concern it belongs to, and expect the\n             lap that follows to re-derive it from scratch. Never say the\n             human's lines are final or that they must survive unchanged.\n\n          The round is not actionable only when none of the above holds — the\n          human left nothing but an approving remark (a compliment, a \"looks\n          good\", or similar) with no code edit and no substantive note. Never\n          invent actionability where the definition above doesn't hold — it\n          manufactures a lap nothing asked for — and never dismiss a real note\n          or edit as approval, which drops feedback the human left.\n\n          If the round is actionable: write `.gtd/REQUIREMENTS.md`\n          with an ordered list of concerns, each classified PRODUCT (a\n          user-facing/requirements decision) or TECHNICAL (an implementation\n          decision) — the same shape `design.triage` builds, one\n          `## <heading>` per concern, in the order you'd want them built. Fold\n          every note, comment, and hand-edit from this round in under the\n          concern it belongs to, per the rules above. Raise no open\n          questions here — write no `## Open Questions` section;\n          `design.triage` owns that, one phase later. Then delete\n          `.gtd/REVIEW_RAW.md` and finish your turn.\n\n          If the round is not actionable: delete\n          `.gtd/REVIEW_RAW.md` and finish your turn, writing nothing\n          to `.gtd/REQUIREMENTS.md` — consuming the capture with no\n          other change is the sign-off.\n        on:\n          \"A .gtd/REQUIREMENTS.md\": $onFeedback\n          \"M .gtd/REQUIREMENTS.md\": $onFeedback\n          \"D .gtd/REVIEW_RAW.md\": $onSignoff\n\n  # The per-package review — a separate MIND from the implementer. ▸ planner\n  # identity; fixing findings is a coder action (packages.item.fix-spec).\n  specReview:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.specReviewerPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onApproved, onFix]\n    entry: review\n    states:\n      review:\n        actor: agent\n        label: Reviewing the package\n        # No retry cap: the loop only re-enters through\n        # `fix-spec`/`health.check`, so under the per-episode rule this state's\n        # episode count can never exceed 1 and a cap could never fire. The\n        # spec-review loop is genuinely unbounded — see docs/configuration.md's\n        # `retry` documentation.\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          You are an autonomous coding agent reviewing a freshly-built work package\n          against its own spec.\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this turn touches is\n          `.gtd/SPEC_FEEDBACK.md`; write it only when — and only\n          when — you find problems. Do not create other files to hold notes or\n          output.\n\n          The package spec is: <%~ it.read(\".gtd/NEXT.md\") %>\n\n          Read that spec, then verify the package's implementation against it: are\n          all its tasks done, acceptance criteria met, and the code sound and\n          consistent with the codebase? This prompt carries no diff — read it\n          yourself: the range runs from `<%= it.startCommit %>` to the working\n          tree.\n\n          Note the range is process-wide (from the start of the whole build, not\n          just this package), so it can span earlier packages too.\n\n          - If it fully satisfies the spec, write nothing and finish your turn — a\n            clean turn is your approval.\n          - If it does not, write `.gtd/SPEC_FEEDBACK.md` listing the\n            concrete problems for a fix turn to address. Be specific.\n\n          Do not fix anything yourself and do not delete the package file — a later\n          step owns that. Finish your turn.\n        on:\n          \"A .gtd/SPEC_FEEDBACK.md\": $onFix\n          \"M .gtd/SPEC_FEEDBACK.md\": $onFix\n          \"D .gtd/SPEC_FEEDBACK.md\": $onApproved\n          \"C\": $onApproved\n\n  # The per-package build queue — identity-free; build identity lives in\n  # packageItem (below).\n  packageLoop:\n    params: [onDrained]\n    entry: picking\n    states:\n      picking:\n        actor: check\n        label: Picking the next package\n        script: |\n          #!/usr/bin/env sh\n          # Mechanics only — NEXT.md's presence/absence is interpreted by\n          # the `on` rows below, never here.\n          set +e\n          mkdir -p .gtd\n          # Sweep spent design/architecture steering files (gone by now) and\n          # any REVIEW_RAW.md the review loop-back left behind — the only\n          # sweeper on that path before it would leak into a later `gtd\n          # summary` prompt's diff range.\n          rm -f .gtd/REQUIREMENTS.md .gtd/ARCHITECTURE.md .gtd/QUESTIONS.md .gtd/REVIEW_RAW.md\n          # Names are gtd-authored, never containing whitespace — safe to\n          # disable SC2012.\n          # shellcheck disable=SC2012\n          next=$(ls .gtd/packages/*.md 2>/dev/null | head -n 1)\n          if [ -n \"$next\" ]; then\n            printf '%s' \"$next\" > .gtd/NEXT.md\n          else\n            rm -f .gtd/NEXT.md\n          fi\n        on:\n          \"D .gtd/NEXT.md\": $onDrained\n          \"* .gtd/NEXT.md\": item.building\n          \"C\": $onDrained\n\n      # Its own coder machine (below) keeps `packages` itself a single state.\n      item:\n        machine: packageItem\n        with:\n          onNext: picking\n\n  # The per-package build identity. ▸ coder identity. packages.item.closing\n  # -> picking is the one upward-resolving target in the file, hence $onNext.\n  packageItem:\n    model: <%= it.vars.coderModel %>\n    system: |-\n      <%~ it.vars.builderPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onNext]\n    entry: building\n    states:\n      building:\n        actor: agent\n        label: Building\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this\n          turn may write is `.gtd/SATISFIED.md` (see below); do not\n          delete the package file (the spec-review gate reads it after you)\n          and never touch `.gtd/NEXT.md` — the `picking` state owns\n          it.\n\n          The package to implement is: <%~ it.read(\".gtd/NEXT.md\") %>\n\n          Before implementing anything, check the package's acceptance criteria\n          against the current tree — an earlier package's fix turn may already\n          have satisfied them. If **every** criterion is already met, implement\n          nothing: write `.gtd/SATISFIED.md` recording each criterion\n          and the concrete evidence it is already satisfied (the commit, file,\n          or symbol that proves it), change nothing else, and finish your turn.\n          If any criterion is outstanding, implement the package normally and\n          do not write that file.\n\n          That file describes a set of independent tasks. Implement all of them in\n          this one turn, fanning the independent tasks out to parallel subagents\n          where your harness supports it — each task is self-contained and can be\n          built concurrently. Implement exactly the tasks this package describes — no\n          more, no less; leave every other package file untouched. Use TDD\n          discipline within each task.\n\n          Leave the package file in place and everything else uncommitted, then\n          finish your turn.\n        on:\n          \"A .gtd/SATISFIED.md\": &satisfied\n            to: health.check\n            action: Close as already satisfied\n            describe: >-\n              record per-criterion evidence in .gtd/SATISFIED.md when this\n              package's work already landed (an earlier package's fix turn\n              pulled it in) — the package still runs the checks and the spec\n              review, then closes out.\n          \"M .gtd/SATISFIED.md\": *satisfied\n          \"* **\": health.check\n\n      fix-suite:\n        actor: agent\n        label: Fixing the check\n        file: FEEDBACK.md\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n\n          <%~ it.vars.fixFeedbackPrompt %>\n\n          If the\n          only way to green the suite is to implement work that another\n          package file describes, make the smallest change that gets there and\n          do it — that later package will then legitimately report itself\n          already satisfied when its turn comes.\n\n          Leave everything uncommitted and finish your turn — do not commit.\n        retry:\n          max: 3\n          otherwise: health.escalate\n        on:\n          \"* **\": health.check\n\n      fix-spec:\n        actor: agent\n        label: Fixing review feedback\n        file: SPEC_FEEDBACK.md\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n          The only state file this\n          turn touches is `.gtd/SPEC_FEEDBACK.md`; address it, then\n          delete it. Do not create other files to hold notes or output.\n\n          Read `.gtd/SPEC_FEEDBACK.md` (the reviewer's concerns about the\n          current package) and the package spec (`<%= it.read(\".gtd/NEXT.md\") %>`),\n          then fix the code to resolve every concern. Keep the change focused.\n\n          Delete `.gtd/SPEC_FEEDBACK.md` once every concern is addressed.\n          Leave everything else uncommitted and finish your turn.\n        on:\n          \"* **\": health.check\n\n      closing:\n        actor: check\n        label: Closing out the package\n        script: |\n          #!/usr/bin/env sh\n          # Removes the just-reviewed package file (path in NEXT.md) plus\n          # leftover spec feedback/evidence, so picking selects the next.\n          # Reached only on spec-review approval — that loop carries no retry\n          # cap, so there is no force-close path here.\n          set +e\n          pkg=$(cat .gtd/NEXT.md 2>/dev/null)\n          [ -n \"$pkg\" ] && rm -f \"$pkg\"\n          rm -f .gtd/SPEC_FEEDBACK.md .gtd/NEXT.md .gtd/SATISFIED.md\n        on:\n          \"* **\": $onNext\n          # Nothing left to sweep (an already-clean NEXT.md/package) still\n          # needs to proceed to the next package, not stall here.\n          \"C\": $onNext\n\n      health:\n        machine: healthGate\n        with:\n          onGreen: spec\n          onRed: fix-suite\n          escalateDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` yourself\n            (**packages.item.health.check**).\n          checkLabel: Running checks\n          escalateLabel: Escalating to a human\n\n      spec:\n        machine: specReview\n        with:\n          onApproved: closing\n          onFix: fix-spec\n\n  # The process's build identity. ▸ coder identity — model/system stamped\n  # onto fix.\n  buildTail:\n    model: <%= it.vars.coderModel %>\n    system: |-\n      <%~ it.vars.finisherPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onDone, onFeedback]\n    # entry: is required by the machine grammar even though nothing resolves\n    # this machine bare; `fix` is an arbitrary pick.\n    entry: fix\n    states:\n      fix:\n        actor: agent\n        label: Fixing the check\n        file: FEEDBACK.md\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n\n          <%~ it.vars.fixFeedbackPrompt %>\n\n\n          Leave everything uncommitted and finish your turn — do not commit.\n        retry:\n          max: 3\n          otherwise: health.escalate\n        on:\n          \"* **\": health.check\n\n      health:\n        machine: healthGate\n        with:\n          onGreen: review\n          onRed: fix\n          escalateDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` yourself\n            (**build.health.check**).\n          checkLabel: Running checks\n          escalateLabel: Escalating to a human\n\n      # Nested here (not root) so the sign-off's boundary commit stays inside\n      # the builder's own scope — see humanReview above.\n      review:\n        machine: humanReview\n        with:\n          onSignoff: $onDone\n          onFeedback: $onFeedback\n\n  # WIRING ONLY — the root ties together the machine tree above.\n  unified:\n    entry: idle\n    states:\n      idle:\n        actor: human\n        label: Idle\n        file: TODO.md\n        message: |\n          No active gtd process.\n\n          To start one, make ANY change — a hand-edit to real code, a scratch\n          note, anything at all. .gtd/TODO.md is a good default\n          place to start sketching. gtd treats it as a SKETCH, not finished\n          work: the very next beat unwinds it out of your working tree (its\n          intent survives in history, for the triage phase to read), then\n          checks the test baseline is green, then triages the reverted diff\n          into ordered, classified concerns: product questions, then\n          technical questions, then one package per concern, built and\n          reviewed in parallel.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: unwind\n            action: Start\n            describe: >-\n              make any change — gtd unwinds it out of your working tree (its\n              intent survives in history) before checking the test baseline is\n              green, then triages the reverted diff into ordered, classified\n              concerns and resolves product questions, then technical ones,\n              before building each concern's package (**unwind** ->\n              **start-gate.check**).\n\n      # --entry fix-precheck: a red suite drops into build.fix (-> health ->\n      # review tail); a green suite is a no-op back to idle.\n      fix-precheck:\n        actor: check\n        label: Checking the baseline\n        entry: true\n        script: *suiteCheck\n        on:\n          \"A .gtd/FEEDBACK.md\": build.fix\n          \"M .gtd/FEEDBACK.md\": build.fix\n          \"D .gtd/REVIEW_RAW.md\": idle\n          \"D .gtd/FEEDBACK.md\": idle\n          \"C\": idle\n\n      # Separate from start-gate.check: that gate's blocked -> check retry\n      # edge would otherwise re-run a revert here and discard a human's\n      # baseline repair. One inbound edge (from idle) means unwind runs\n      # exactly once, so it needs no idempotence guard itself.\n      unwind:\n        actor: check\n        label: Unwinding your input\n        script: |\n          #!/usr/bin/env sh\n          set +e\n          mkdir -p .gtd\n          # Hoisted here, at the TOP: Eta's autoTrim eats the newline after\n          # an interpolation tag, so no tag may be the last token on a line.\n          # Uses it.currentCommit (render-time), not bare HEAD, so a\n          # late-running driver still reverts the right commit.\n          commit=\"<%~ it.currentCommit %>\"\n          git revert --no-commit \"$commit\" 2> .gtd/.unwind-error\n          code=$?\n          # The revert's EXIT CODE is what separates a genuine no-op from a\n          # hard failure (e.g. a merge commit with no `-m`) — the diff alone\n          # cannot, since both can leave a clean tree. Turning the failure\n          # into a FEEDBACK.md write is what makes the `C` row below safe:\n          # once a failure always has a diff, a clean tree here means the\n          # revert really did succeed and change nothing.\n          if [ \"$code\" -ne 0 ]; then\n            printf 'gtd could not unwind %s out of your working tree.\\n\\n' \"$commit\" > .gtd/FEEDBACK.md\n            if [ -s .gtd/.unwind-error ]; then\n              cat .gtd/.unwind-error >> .gtd/FEEDBACK.md\n            else\n              printf '`git revert --no-commit` exited %s and produced no output.\\n' \"$code\" >> .gtd/FEEDBACK.md\n            fi\n          fi\n          rm -f .gtd/.unwind-error\n        on:\n          \"A .gtd/FEEDBACK.md\": unwind-failed\n          \"M .gtd/FEEDBACK.md\": unwind-failed\n          \"* **\": start-gate\n          \"C\": start-gate\n\n      # Reached only when the revert above exited non-zero. A human repairs\n      # the tree by hand; start-gate.check clears FEEDBACK.md on its way\n      # green, so nothing has to sweep it here.\n      unwind-failed:\n        actor: human\n        label: Could not unwind your input\n        file: FEEDBACK.md\n        message: |\n          gtd could not revert your sketch out of the working tree.\n          `.gtd/FEEDBACK.md` holds the error.\n\n          Undo the sketch by hand (its intent survives in history either\n          way), then continue — gtd will not start work on a tree that still\n          carries it.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: start-gate\n            action: Continue\n            describe: >-\n              having undone the sketch by hand, check the test baseline is\n              green and start triage (**start-gate.check**).\n\n      # Reverts the human's own review-round hand-edit before re-planning,\n      # scoped to real code (`.gtd/` excluded — REVIEW.md is already gone by\n      # now). Names the commit via it.reviewBase (reviewBaseFor's own\n      # `await-review -> deciding` commit), never bare HEAD.\n      #\n      # requireRevert (below) re-checks the tree itself and refuses the step\n      # if residue remains, rather than trusting the script's exit code — a\n      # clean-filter round-trip or binary path can make `git apply -R` a\n      # silent no-op. Both `on` rows below target `design`: a hand-edited\n      # round leaves a dirty tree, a note-only round's patch is empty and\n      # leaves the tree clean — both are legitimate outcomes here, not a\n      # stall.\n      #\n      # `file: REVIEW.md` + `requireRevert: true` exist only to give that\n      # guard the path to exempt (it reads Rest.hints.file, never this\n      # literal) — REVIEW.md is already gone two commits earlier, so this\n      # state never renders it.\n      re-unwind:\n        actor: check\n        label: Re-unwinding your review edit\n        file: REVIEW.md\n        requireRevert: true\n        script: |\n          #!/usr/bin/env sh\n          # Scoped revert of the human's review-round edit — .gtd/ excluded\n          # (the guard's isCodePath re-derives the same exemption; keep both\n          # in sync). Expected to succeed; requireRevert catches a silent\n          # apply failure.\n          set +e\n          # Hoisted here, at the TOP: Eta's autoTrim eats the newline after\n          # an interpolation tag, so no tag may be the last token on a line.\n          commit=\"<%~ it.reviewBase %>\"\n          patch=.gtd/.re-unwind.patch\n          mkdir -p .gtd\n          git diff --binary \"$commit^\" \"$commit\" -- . \":(exclude).gtd\" > \"$patch\"\n          if [ -s \"$patch\" ]; then\n            git apply -R \"$patch\" || echo \"re-unwind: could not revert $commit\" >&2\n          fi\n          rm -f \"$patch\"\n        on:\n          \"* **\": design # a code hand-edit was reverted\n          \"C\": design # note-only round: nothing to revert\n\n      start-gate:\n        machine: entryGate\n        with:\n          onGreen: design\n          reviewBase: \"\"\n          blockedMessage: |\n            The test baseline is red — gtd will not start new work on a broken suite.\n            `.gtd/FEEDBACK.md` holds the failing output.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n          blockedDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` to fix the failing tests\n            (**start-gate.check**). To repair the baseline as its own separate\n            reviewed commit instead, abandon this start and run\n            `gtd --entry fix-precheck` from a clean `idle`.\n          checkLabel: Checking the baseline\n          blockedLabel: Baseline is red\n\n      review-gate:\n        machine: entryGate\n        with:\n          onGreen: build.review\n          reviewBase: <%= it.vars.reviewBase %>\n          blockedMessage: |\n            The test baseline is red — gtd will not start a review on a broken suite.\n            `.gtd/FEEDBACK.md` holds the failing output.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n          blockedDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` to fix the failing tests\n            (**review-gate.check**).\n          checkLabel: Checking the baseline\n          blockedLabel: Baseline is red\n\n      design:\n        machine: designPlan\n        with:\n          onArchitecture: architecture\n\n      architecture:\n        machine: archPlan\n        with:\n          onPackages: packages\n\n      packages:\n        machine: packageLoop\n        with:\n          onDrained: build.review\n\n      build:\n        machine: buildTail\n        with:\n          onDone: idle\n          onFeedback: re-unwind\n";
+var unified_default = "# gtd's BUILT-IN DEFAULT workflow — used when no `workflow:` key is\n# configured anywhere in the cwd→home config chain; `gtd init` seeds only\n# vars.testCommand + a modes: suggestion, never this file. Compiled through\n# the same `compileWorkflowConfig` a user's own `workflow:` goes through — no\n# privileged path. See skills/authoring/SKILL.md for how to read or edit it\n# (machines, patterns, templates) and docs/driver.md for how a driver runs it.\n#\n# One flow: any change to the tree starts the process. `idle` -> `unwind`\n# reverts that diff (intent survives only in history) -> a green-baseline\n# gate -> `design.triage` groups the reverted diff into ordered, classified\n# concerns and resolves the PRODUCT ones' open questions (`design.gate`) ->\n# `architecture.author` works out the *how* and resolves the TECHNICAL ones\n# (`architecture.gate`) -> `architecture.decompose` writes one package file\n# per concern -> the per-package queue (`packages.*`) builds and reviews each\n# -> the shared tail (`build.review.*`) decides sign-off (-> `idle`, landing\n# an ordinary commit that closes the process — every per-turn commit stays on\n# the branch; run `gtd summary` afterward for a closing-message prompt) vs.\n# actionable feedback (-> `re-unwind` -> back to `design.triage`, a full\n# re-plan lap that never builds on the human's hand-edit). `--entry\n# review-gate.check --var reviewBase=<commitish>` and `--entry fix-precheck`\n# are the two other entries into this same flow.\nvars:\n  testCommand: npm test\n  # plannerModel: the heavier tier for one-shot triage/design/review turns.\n  # coderModel: the tier for build/fix turns. Both repointable via a\n  # top-level `vars:` key or a `GTD_<NAME>` env var, like any other var.\n  plannerModel: smart\n  coderModel: base\n  # Declared (not left undeclared) so `--var reviewBase=...` passes the CLI's\n  # declared-name check; blank renders empty, which the review entry refuses.\n  reviewBase: \"\"\n  # gtd's own voice for generated files — adapted from the \"Spartan\" output\n  # style (https://github.com/alexgreensh/attention-span, AGPL-3.0, version 0.6),\n  # rewritten in gtd's own words for deliverables rather than chat replies;\n  # no upstream text ships. A point-in-time derivation with no refresh\n  # mechanism.\n  styleBlock: |-\n    This file is a deliverable, not a chat reply — size follows the work, and\n    padding is the target of any cut. Lead with the answer immediately; never\n    circle back to restate it once it has landed.\n\n    Write flat, commanding sentences — commit to the claim, do not hedge it.\n    Favor everyday vocabulary; where a term is unavoidable, define it in five\n    words or fewer on first use.\n\n    In markdown, bold carries the load: bold the claim, not the sentence around\n    it — reading only the bold text must yield the full point and every risk.\n\n    Never trim a risk, a number, a threshold, or a scoped condition to save\n    space. This outranks every other rule here.\n\n    Ship the artifact bare — no lead-in, no sign-off wrapped around it.\n    Compressing is not dropping: a task with three load-bearing parts ships all\n    three, each shorter, never two of three.\n\n    One idea per block; break when the idea shifts. Flag risk in one blunt\n    line rather than hedged prose. Never narrate the work — do it.\n  styleFormatContract: |-\n    This file is machine-read; its format contract outranks every style rule\n    above it: keep every `##`/`###` heading, every `- [ ]` checkbox row, and\n    every marker line exactly as specified — do not renumber or rename a\n    heading or replace a checkbox list with prose. A parser reads these\n    literally; a violation refuses the turn outright.\n\n    Voice rules above govern the prose between these elements, never the\n    elements themselves.\n\n  # One persona per prompt-bearing machine below, stamped as its `system:`\n  # (see src/StateFields.ts's `system`). `--system-prompt` REPLACES a\n  # harness's default system prompt — including its injected cwd/env/\n  # git-status block — so each persona is written self-contained, restating\n  # that context explicitly rather than relying on it.\n\n  # Shared conduct tail, appended after every persona's own role paragraph:\n  # use tools without asking, orient yourself with git, and go inspect what\n  # the turn's message names. Deliberately NOT named `*Persona`:\n  # `templates.test.ts` derives its six-name persona set from\n  # `it\\.vars\\.(\\w+Persona)`, so that suffix would silently grow the set.\n  agentConduct: |-\n    You have shell and file tools — bash, read, write, edit — and you are\n    expected to use them without asking first. This runs unattended: there is\n    no one to grant permission, confirm a command, or approve a plan before\n    you act. Investigate with real commands rather than assuming what a file,\n    a commit, or a decision's status actually is, then act on what you find.\n\n    You are already inside a git repository, at the working directory you\n    should treat as your whole world for this turn. There is no injected\n    status block — no cwd, no branch, no recent history, no tree state. Run\n    `git status`, `git log`, `ls` and the like yourself before you do\n    anything else; that is the normal first move of this turn, not an extra\n    step.\n\n    The message you receive each turn names a commit, a range, or a file for\n    you to go inspect yourself — `git show`, `git diff`, reading the file\n    directly — never a diff or a summary rendered into the conversation. Go\n    look before you act. Across the several turns one conversation may span,\n    work from what you already read and settled earlier rather than\n    re-deriving it from a cold read every time, unless the turn's own\n    instructions tell you this is a fresh, cold pass.\n  designPersona: |-\n    You are the product-facing planning voice in an autonomous build pipeline\n    called gtd. Your job is to turn a raw, half-formed sketch — a hand-edit to\n    real code, a scratch note, or both — into an ordered, classified list of\n    concerns a build can execute, and to carry a running conversation with a\n    human about the product decisions that list depends on. You are read by\n    that human, so write plainly and decisively; you are not read by a parser\n    except where a state's own instructions say otherwise.\n  architectPersona: |-\n    You are the technical planning voice in an autonomous build pipeline\n    called gtd. You take over once the product-facing concerns are settled,\n    reading them cold — you do not carry over the product conversation, only\n    its written, settled output. Your job is to work out the technical *how*\n    for each settled concern — file and module structure, data models,\n    library and tech-stack choices, error-handling strategy — raise the\n    technical open questions a human still needs to decide, and once every\n    question is answered, write one package spec file per concern with no\n    further judgement call: the grouping and ordering were already decided\n    upstream, and this final step is a mechanical write-out of a plan you\n    already converged on.\n  reviewerPersona: |-\n    You are the independent reviewing mind in an autonomous build pipeline\n    called gtd — deliberately a separate identity from whichever agent wrote\n    the code under review, so you bring no attachment to how it got built.\n    Your job splits into two distinct turns: first, writing a structured\n    review document that groups a diff into reviewable chunks for a human to\n    read through and comment on; later, reading back a round of that human's\n    feedback — a note, a hand-edit — and classifying whether it is genuinely\n    actionable or just an approving remark, without fixing anything yourself\n    either way. You judge and you classify; you never build.\n  specReviewerPersona: |-\n    You are the adversarial spec-conformance checker in an autonomous build\n    pipeline called gtd, reviewing one freshly-built work package against the\n    written spec it was built from. Your only job is to verify: are the\n    package's tasks actually done, its acceptance criteria actually met, and\n    the code sound and consistent with the rest of the codebase? You write\n    concrete, actionable feedback only when something is genuinely wrong —\n    and when nothing is, you write nothing at all, because a clean turn from\n    you IS the approval. You never fix a problem you find; naming it\n    precisely enough for someone else's fix turn to act on is the whole job.\n  builderPersona: |-\n    You are the TDD implementer in an autonomous build pipeline called gtd.\n    Your job is to build exactly one package's declared scope end to end,\n    tests first, and to come back and fix things when redirected: a failing\n    check's output, or a reviewer's concrete feedback. Stay strictly inside\n    the package in front of you, touching neither another package's files nor\n    code outside the current task. Treat feedback you're given as the\n    diagnosis to act on; discard it, with reason, only when it is simply\n    wrong.\n  finisherPersona: |-\n    You are the closing identity in an autonomous build pipeline called gtd —\n    the last coder this process hands work to before it closes. You are\n    handed a late-breaking failing check after sign-off; fix it, focused and\n    minimal, before the process can close. Every turn you land stays its own\n    commit on the branch — nothing here gets squashed away.\n\n  # The scratchpad-opener sentence shared by every prompt-content state (see\n  # STATE_FIELDS's `prompt`) — this workflow's own state files are its private\n  # working notes, never project code or documentation a human reads. Injected\n  # as the leading line of every prompt body that touches a `.gtd/` file\n  # directly; each state's own \"the only state file this turn touches is X\"\n  # sentence stays local, immediately after the tag. Some states splice a\n  # role clause into this same opening sentence (`build.review.collecting`,\n  # `packages.item.spec.review`) — those hoist the clause to its own leading\n  # sentence ahead of the tag instead, keeping every word.\n  stateFileRules: |-\n    You are an autonomous coding agent. This workflow steers itself\n    through its own state files — treat them as its private scratchpad,\n    never as project code or documentation.\n  # The open-questions warrant test, decide-it-yourself sink, and `## Open\n  # Questions` checkbox shape shared by `design.triage` and\n  # `architecture.author` — written phase-agnostic (no PRODUCT/TECHNICAL\n  # bake-in) so both sites can inject it verbatim. Each site's own\n  # phase-scope sentence (triage: product-only, technical waits; architecture:\n  # every point here is TECHNICAL) stays local around the tag — the two\n  # genuinely disagree on scope, so that sentence is never shared.\n  # `questionBarReturn` is the return-lap half — split out so each site can\n  # inject it under its own `## Return lap` heading instead of burying\n  # return-lap behaviour inside the first lap's own section.\n  questionBar: |-\n    For every open point, a question is warranted only when all three hold:\n    the answer is a genuine fork with divergent outcomes — user-visible for a\n    product question, materially different implementations for a technical\n    one; the sketch, the entry diff, and history do not already settle it\n    (treat anything the entry diff states explicitly — a directive in\n    `.gtd/TODO.md`, a committed choice in a hand-edit — as already settled,\n    and on a review loop-back lap treat whatever the human's own review-round\n    edit or note states explicitly the same way); and getting it wrong would\n    survive all the way to the human review tail.\n\n    Anything below that bar, decide yourself: record the decision under\n    `## Answered Questions` — the same section a human-answered question\n    lands in — as `### <the point phrased as a question>` followed by the\n    decision and a one-line rationale in prose, no checkboxes.\n    `## Answered Questions` must come after every other `##` section in the\n    document — it is always the last section.\n\n    A question that clears the bar is written under a `## Open Questions`\n    heading that must come before every other `##` section in the document —\n    not just near the top, but ahead of every other `##` section — and takes\n    this shape: one `### <question>` sub-heading each, followed by a checkbox\n    list of candidate answers — two concrete options plus a final free-text\n    slot, all left unticked:\n\n        ### <the question>\n\n        - [ ] <first option — a concrete answer, a few words of rationale>\n        - [ ] <second option>\n        - [ ] _your answer_\n\n    Never tick a box yourself — the human ticks exactly one per question.\n  questionBarReturn: |-\n    The human has ticked one option per question (or replaced\n    `_your answer_` with their own text): fold each chosen answer into its\n    concern's prose, then move the question into a `## Answered Questions`\n    section as `### <question>` followed by the resolved answer in plain\n    prose (drop the checkboxes). `## Answered Questions` must come after\n    every other `##` section in the document — it is always the last\n    section, so a moved question lands there, not wherever `## Open\n    Questions` used to sit. Never re-raise a deleted question; treat a\n    deleted `## Open Questions` section as acceptance. Omit `## Open\n    Questions` entirely once there are none left.\n  # The body `packages.item.fix-suite` and `build.fix` share byte for byte —\n  # both fix a red `.gtd/FEEDBACK.md`. `fix-suite` appends its own\n  # cross-package paragraph right after this tag (before the shared\n  # \"Leave everything uncommitted\" close, which stays local at both sites so\n  # the appended paragraph lands in the same position it renders in today);\n  # `build.fix` renders the tag alone.\n  fixFeedbackPrompt: |-\n    The only state file this turn touches is `.gtd/FEEDBACK.md`; address it,\n    then delete it. Do not create other files to hold notes or output.\n\n    Read `.gtd/FEEDBACK.md` (the failing test output) and fix the\n    code so the suite passes. Keep the change focused — do not refactor\n    unrelated code. If the feedback is wrong, fix nothing rather than\n    \"fixing\" a non-issue.\n\n    Delete `.gtd/FEEDBACK.md` either way — once the suite passes, or once\n    you have established the feedback was wrong. Leaving it in place is\n    never the right end state: the next check writes its own.\n\n# `gtd summary`'s prompt — printed cold, no session identity, no diff inlined.\n# Names the hashes/range for the agent to inspect itself; renders the same\n# it.processCost/it.processCostByModel the old squash finale used to.\nsummary: |\n  <%~ it.vars.styleBlock %>\n\n\n  Write the closing message for the process HEAD closes or currently sits\n  inside — the text a driver may use for a squash, an amend, or a pull\n  request body. You are starting cold: read every decision back out of the\n  commits named below rather than assuming any prior context.\n\n  Cover the initial motivation, the important decisions, and the trade-offs\n  made, and the high-level architectural changes — never which files\n  changed; that is what `git diff --stat` is for, not this message.\n\n  The process's entry commit is `<%= it.entryCommit %>`. <% if\n  (it.humanCommits.length > 0) { %>The human contributed at these commits,\n  oldest to newest:\n  <% it.humanCommits.forEach(function (c) { %>\n  - `<%= c.hash %>` (entering `<%= c.state %>`)\n  <% }) %><% } else { %>The human left no comment or edit this process —\n  every commit is machine-authored.<% } %>\n\n  Inspect the range yourself: `git log <%= it.processBase %>..<%= it.processTip %>`\n  and `git diff <%= it.processBase %> <%= it.processTip %>`. That range is\n  exactly what a squash or PR body should describe.\n\n  Token cost: <%= it.processCost %>\n  <% it.processCostByModel.forEach(function(m){ %>\n  - <%= m.model %>: <%= m.cost %>\n  <% }) %>\n\n  Print the closing message and stop — this command writes nothing itself.\n\nentry:\n  default: unified\n\nmachines:\n  # Shared green-baseline gate — start-gate/review-gate/fix-precheck all\n  # alias this script via &suiteCheck.\n  entryGate:\n    params: [onGreen, blockedMessage, blockedDescribe, checkLabel, blockedLabel, reviewBase]\n    entry: check\n    states:\n      check:\n        actor: check\n        label: $checkLabel\n        script: &suiteCheck |\n          #!/usr/bin/env sh\n          set +e\n          mkdir -p .gtd\n          # Sweep a raw review capture an earlier, abandoned process may have\n          # left behind — no ordinary path from deciding/collecting reaches\n          # this check.\n          rm -f .gtd/REVIEW_RAW.md\n          <%~ it.vars.testCommand %> > .gtd/.check-output 2>&1\n          code=$?\n          if [ \"$code\" -ne 0 ]; then\n            if [ -s .gtd/.check-output ]; then\n              mv .gtd/.check-output .gtd/FEEDBACK.md\n            else\n              rm -f .gtd/.check-output\n              printf 'the test command failed with exit code %s and produced no output.' \"$code\" > .gtd/FEEDBACK.md\n            fi\n            # Stamp with HEAD so a repeat identical failure still re-registers\n            # as an M/A edit instead of looking byte-identical (GREEN).\n            printf '\\n<!-- gtd check %s -->\\n' \"$(git rev-parse --short HEAD 2>/dev/null || echo pending)\" >> .gtd/FEEDBACK.md\n          else\n            rm -f .gtd/.check-output\n            rm -f .gtd/FEEDBACK.md\n          fi\n        # Marks BOTH entryGate instances (start-gate/review-gate); only\n        # review-gate strictly needs it, but the shared dedup makes marking\n        # one mark both — both ARE entry gates.\n        entry: true\n        # Only review-gate binds a real reviewBase (via --var); start-gate\n        # binds \"\" (no base), since it's reached from idle, never entered.\n        reviewBase: $reviewBase\n        on:\n          \"A .gtd/FEEDBACK.md\": blocked\n          \"M .gtd/FEEDBACK.md\": blocked\n          \"D .gtd/REVIEW_RAW.md\": $onGreen\n          \"D .gtd/FEEDBACK.md\": $onGreen\n          \"C\": $onGreen\n      blocked:\n        actor: human\n        label: $blockedLabel\n        file: FEEDBACK.md\n        message: $blockedMessage\n        on:\n          \"* **\":\n            to: check\n            action: Retry check\n            describe: $blockedDescribe\n\n  # Shared check/escalate pair for build.health.*/packages.item.health.* —\n  # the caller's own coder fix-state ($onRed), not this machine, handles red.\n  healthGate:\n    params: [onGreen, onRed, escalateDescribe, checkLabel, escalateLabel]\n    entry: check\n    states:\n      check:\n        actor: check\n        label: $checkLabel\n        # See the shared suite check on `entryGate.check` above.\n        script: *suiteCheck\n        on:\n          \"A .gtd/FEEDBACK.md\": $onRed\n          \"M .gtd/FEEDBACK.md\": $onRed\n          \"* **\": $onGreen\n          \"C\": $onGreen\n      escalate:\n        actor: human\n        label: $escalateLabel\n        file: FEEDBACK.md\n        message: |\n          The agent could not get the check to pass after repeated attempts.\n          `.gtd/FEEDBACK.md` holds the last failing output.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: check\n            action: Retry check\n            describe: $escalateDescribe\n\n  # Shared check/answer pair for design.gate/architecture.gate. No `file:`\n  # PARAM for the probe script itself: machine `$param` substitution is\n  # whole-value only, so a path can't be spliced into a shared script body —\n  # `file` here is instead the `answer` param's own `file:` (the human gate).\n  questionGate:\n    params: [file, message, onNone, onRevise, checkLabel, answerLabel]\n    entry: check\n    states:\n      check:\n        actor: check\n        label: $checkLabel\n        script: |\n          #!/usr/bin/env sh\n          # Exactly one of REQUIREMENTS.md/ARCHITECTURE.md exists on disk at\n          # a time — architecture.author deletes the former in the same turn\n          # it writes the latter — so this probe is unambiguous either way.\n          set +e\n          mkdir -p .gtd\n          file=.gtd/REQUIREMENTS.md\n          [ -f \"$file\" ] || file=.gtd/ARCHITECTURE.md\n          gtd check qa \"$file\" --open-questions > /dev/null 2>&1\n          code=$?\n          if [ \"$code\" -ne 0 ]; then\n            printf 'open questions remain in %s\\n' \"$file\" > .gtd/QUESTIONS.md\n            # See the shared suite check's cache-buster rationale on\n            # `entryGate.check` above.\n            printf '\\n<!-- gtd check %s -->\\n' \"$(git rev-parse --short HEAD 2>/dev/null || echo pending)\" >> .gtd/QUESTIONS.md\n          else\n            rm -f .gtd/QUESTIONS.md\n          fi\n        on:\n          \"A .gtd/QUESTIONS.md\": answer\n          \"M .gtd/QUESTIONS.md\": answer\n          \"D .gtd/QUESTIONS.md\": $onNone\n          \"C\": $onNone\n      answer:\n        actor: human\n        label: $answerLabel\n        file: $file\n        mode: qa\n        answerGate: true\n        message: $message\n        on:\n          \"* **\":\n            to: $onRevise\n            action: Revise answers\n            describe: >-\n              tick exactly one option per open question (replace\n              `_your answer_` for your own) to send it back for the agent to\n              fold your answers in. Delete a question to skip it, or delete\n              the whole `## Open Questions` section to accept the plan as-is.\n\n  # The triage phase — one conversation across the whole back-and-forth (the\n  # human's answer rationale survives every return lap). A review loop-back\n  # lap runs the suite itself and files any breakage as its own FIRST\n  # concern, rather than spending one of packages.item.fix-suite's\n  # (per-episode) retries on breakage the package didn't cause.\n  # ▸ planner identity.\n  designPlan:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.designPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onArchitecture]\n    entry: triage\n    states:\n      triage:\n        actor: agent\n        label: Triaging the change\n        file: REQUIREMENTS.md\n        mode: qa\n        # Guards against discarding assembled review input without folding\n        # it in (the original bug this guard was added for).\n        requireProgress: true\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this turn touches is `.gtd/REQUIREMENTS.md`; develop it. Do\n          not create other files to hold notes or output.\n\n          `.gtd/TODO.md` is the likely home of the sketch that\n          started this process. Despite living beside this workflow's own\n          state files, it is the human's input — the sketch to be folded into\n          the concerns below like any other part of the start diff. It is\n          never a state file to preserve, and never gtd bookkeeping to ignore.\n\n          This process started at commit `<%= it.startCommit %>`. The change\n          that started it was reverted out of the working tree by the\n          `unwind` beat right after it landed, so `git diff <%= it.startCommit %>`\n          (against the working tree) is now empty — its content survives only\n          in history. Find the entry commit yourself: run\n          `git rev-list --ancestry-path <%= it.startCommit %>..HEAD | tail -1`\n          (the process's very first turn, landed immediately after\n          `<%= it.startCommit %>`, before any later baseline-repair commits) —\n          then `git show` that commit to see everything that started this\n          process: a hand-edit to real code, a scratch note in some file, or\n          both.\n\n          Whichever lap this is, leave `.gtd/REQUIREMENTS.md` uncommitted and\n          finish your turn once it reflects this lap's own work below.\n\n          ## First lap\n\n          `.gtd/REQUIREMENTS.md` does not exist yet (or holds whatever the\n          human's own change happened to leave there). Build it from that\n          start diff:\n\n          ### Group into concerns\n\n          Group the diff into an ordered list of concerns — each one a\n          coherent piece of work that can leave the test suite green on\n          its own, in the order you'd want them built. If a concern would\n          red-light tests that a later concern is meant to fix, that is\n          not a valid split: merge the two into one concern.\n\n          Two more tests sit beside that green floor, not instead of it.\n          Vertical, not horizontal: slice by capability, never by layer. A\n          concern that only prepares for a later one — types, scaffolding,\n          renames — is folded into the first concern that uses it. Distinct\n          acceptance: each concern needs an observable check that fails\n          before it and passes after; if you can't name one, merge it into\n          its neighbour. Prefer fewer, larger packages — the smallest\n          independently valuable change, not the smallest change that\n          compiles.\n\n          ### Classify each concern\n\n          Classify each concern as PRODUCT (a user-facing/requirements\n          decision) or TECHNICAL (an implementation decision).\n\n          <%~ it.vars.questionBar %>\n\n\n          Raise questions only for product concerns here — a technical\n          concern's open points wait for the next phase. When every concern\n          in this lap is TECHNICAL, write no `## Open Questions` section at\n          all.\n\n          ### Fold in the sketch\n\n          Fold everything the entry commit added into the concerns above —\n          a scratch note and a real code edit are both just a sketch of\n          the intent, a spec to be finished, never work to be preserved;\n          `unwind` already reverted both out of the tree, so there is\n          nothing left to delete. Describe each piece under the concern it\n          belongs to.\n\n          ## Return lap\n\n          <%~ it.vars.questionBarReturn %>\n\n          ## Review loop-back\n\n          `.gtd/REQUIREMENTS.md` already holds concerns and none of them\n          have a ticked-but-unfolded answer waiting — that shape means a\n          completed REVIEW round put this material here, not a question you\n          asked. Develop those existing concerns further with whatever the\n          round raised; never rediscover or regroup them from a cold read\n          the way the first lap does.\n\n          The human's own edit from that review round has been reverted out\n          of the working tree the same way the entry commit is — read it\n          from history, never the working tree: run\n          `git show <%= it.reviewBase %>` to see it. (On the first lap that\n          same hash is the process's own diff base, `<%= it.startCommit %>`,\n          so this branch does not apply: `.gtd/REQUIREMENTS.md`\n          does not exist yet for a review round to have happened against.)\n\n          Whatever this round touches, every decision already settled under\n          `## Answered Questions` stays settled on this lap too — never\n          re-open one. A genuinely open PRODUCT point in the round's own\n          material may still raise a fresh question under\n          `## Open Questions`: the product gate sits on this path exactly as\n          it does on the first lap, and will ask it.\n\n          Before grouping anything on this lap, run\n          `<%= it.vars.testCommand %>` yourself as part of reading the\n          situation: a review loop-back runs no green-baseline gate between\n          the revert above and this turn, so the tree may already be red —\n          reverting the human's edit can undo a fix that edit had put in. If\n          the suite is red, make that breakage the first concern in your\n          ordered list, ahead of everything else — every later concern's\n          \"leaves the suite green on its own\" property assumes the suite\n          already started green.\n        on:\n          \"* **\": gate.check\n\n      gate:\n        machine: questionGate\n        with:\n          file: REQUIREMENTS.md\n          onNone: $onArchitecture\n          onRevise: triage\n          checkLabel: Checking for open questions\n          answerLabel: Awaiting your product answers\n          message: |\n            `.gtd/REQUIREMENTS.md` holds the concerns under\n            development. Each open question under `## Open Questions` offers\n            a few options plus a `- [ ] _your answer_` slot. Answer EVERY\n            question by ticking exactly one box (`- [x]`); for your own\n            answer, replace `_your answer_` with your text and tick that\n            line. Stepping is refused while any question is unanswered.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n\n  # A separate machine from designPlan (its own memory scope — a COLD read,\n  # not a resumed conversation). Merge authority lives in `author`, the first\n  # state that knows the *how*: it alone may re-merge concerns, never split;\n  # `decompose` carries that grouping over verbatim. ▸ planner identity.\n  archPlan:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.architectPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onPackages]\n    entry: author\n    states:\n      author:\n        actor: agent\n        label: Refining the technical plan\n        file: ARCHITECTURE.md\n        mode: qa\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state files this\n          turn touches are `.gtd/ARCHITECTURE.md` (write it) and\n          `.gtd/REQUIREMENTS.md` (delete it once its content is\n          folded in). Do not create other files to hold notes or output.\n\n          You do not resume the design conversation — this is a separate\n          machine with its own memory, a cold read every time. Read\n          `.gtd/REQUIREMENTS.md` (the settled, ordered, classified\n          concerns) in full. Treat every decision recorded there — plain\n          prose or a resolved `## Answered Questions` entry, PRODUCT or\n          TECHNICAL alike — as settled; never re-open it.\n\n          Cold means the machine carries no conversational memory of\n          triage's own back-and-forth; it never meant no git access. This\n          process started at commit `<%= it.startCommit %>`. Find the\n          process's first turn yourself: run\n          `git rev-list --ancestry-path <%= it.startCommit %>..HEAD | tail -1`\n          then `git show` that commit to see everything that started this\n          process: a hand-edit to real code, a scratch note in some file,\n          or both.\n\n          Whichever lap this is, once `.gtd/ARCHITECTURE.md` is written,\n          delete `.gtd/REQUIREMENTS.md` — its content has been folded in and\n          it must not linger. Leave everything uncommitted and finish your\n          turn.\n\n          ## First lap\n\n          Develop `.gtd/ARCHITECTURE.md` from those concerns: for each one,\n          in the same order, work out the *how* — file/module structure,\n          data models, library/tech-stack choices, error-handling strategy —\n          building on the *what* already settled.\n\n          You now know the *how*, so you know each concern's file footprint —\n          list each concern's primary paths. Merge concerns whose footprints\n          center on the same files into one, unless the later one only\n          consumes an interface the earlier one creates: that exception is\n          what keeps a genuine build-on-top sequence from collapsing into one\n          blob. This authority is to merge only, never to split — granting\n          split authority here would work directly against the goal; the\n          whole problem is over-granularity. Record every merge under a\n          dedicated `## Merged Concerns` heading, carrying both merged requirements\n          verbatim so the per-package spec review still covers each one\n          independently. A merge raises no open question and stops for no\n          human — do not route it to `architecture.gate` for a veto; the\n          human still sees every merge when they review the plan, and the\n          per-package spec review is the real safety net. Prefer fewer,\n          larger packages — the smallest independently valuable change, not\n          the smallest change that compiles.\n\n          Every open point here is TECHNICAL — triage already resolved the\n          product ones, one phase earlier.\n\n          <%~ it.vars.questionBar %>\n\n          ## Return lap\n\n          <%~ it.vars.questionBarReturn %>\n        on:\n          \"* **\": gate.check\n\n      gate:\n        machine: questionGate\n        with:\n          file: ARCHITECTURE.md\n          onNone: decompose\n          onRevise: author\n          checkLabel: Checking for open questions\n          answerLabel: Awaiting your technical answers\n          message: |\n            `.gtd/ARCHITECTURE.md` holds the technical plan under\n            development. Each open question under `## Open Questions` offers\n            a few options plus a `- [ ] _your answer_` slot. Answer EVERY\n            question by ticking exactly one box (`- [x]`); for your own\n            answer, replace `_your answer_` with your text and tick that\n            line. Stepping is refused while any question is unanswered.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n\n      decompose:\n        actor: agent\n        label: Decomposing into packages\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state files this turn touches\n          are the package files it writes under `.gtd/packages/` and\n          `.gtd/ARCHITECTURE.md`, which it deletes. Do not create other\n          files to hold notes or output.\n\n          If you wrote `.gtd/ARCHITECTURE.md` earlier in this\n          conversation, work from what you already settled; otherwise read it\n          (the converged technical plan) yourself. It already lists an\n          ordered set of concerns with every merge/split judgement\n          `architecture.author` made — this turn is a mechanical write-out,\n          not a planning one. A `## Merged Concerns` heading there is a\n          record of those merges, never a concern of its own: write no\n          package file for it.\n\n          Write one package file per concern, in the settled order, under\n          `.gtd/packages/` (e.g.\n          `.gtd/packages/01-name.md`,\n          `.gtd/packages/02-name.md`, ...). Each package file\n          carries that concern's requirement(s) (so it can be reviewed\n          against its spec — both, independently, when the concern was\n          merged), its independent tasks, and each task's acceptance\n          criteria as `- [ ]` checkboxes and relevant paths. Do not merge or\n          split concerns here — that judgement already happened in\n          `architecture.author`; carry the settled grouping over verbatim.\n          No package file may reference any other `.gtd/` file.\n\n          Once the package files are written, delete\n          `.gtd/ARCHITECTURE.md`. Leave everything uncommitted and\n          finish your turn.\n        on:\n          \"* .gtd/packages/**\": $onPackages\n\n  # The shared review tail. ▸ planner identity — `collecting` only judges/\n  # classifies feedback; there's no coder follow-through inside this machine.\n  #\n  # Nested inside buildTail (build.review), not a root sibling: a machine's\n  # memory scope is its dotted instance path, so nesting keeps `build.fix ->\n  # build.health.check -> build.review.*` (the `--entry fix-precheck` path)\n  # inside one scope, so the sign-off's boundary commit lands from the same\n  # scope that made the fixes.\n  humanReview:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.reviewerPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    entry: reviewing\n    states:\n      reviewing:\n        actor: agent\n        label: Reviewing\n        file: REVIEW.md\n        mode: review\n        # `gtd --entry review-gate.check` enters via review-gate.check first\n        # (never straight here), which sets the Gtd-Review-Base trailer this\n        # tail reuses.\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this turn touches is `.gtd/REVIEW.md`; write it at exactly\n          that path. Do not create other files to hold notes or output.\n\n          Write `.gtd/REVIEW.md` to help a human review the changes in\n          this exact format:\n\n          - First non-blank line: `# Review: <%= it.currentCommit.slice(0, 7) %>`\n          - Somewhere in the document: `<!-- base: <%= it.reviewBase %> -->`\n          - At least one `## <Chunk Title>` heading, grouping hunks semantically\n            (same feature/refactor/fix, even across files) — a short explanation of\n            what changed and why, then one file pointer per hunk (checkboxes are\n            for the human to tick, not you; every path is `./`-relative, with an\n            optional `#line`). Put the note's opening line right on the\n            pointer's own line, after a dash — the reflow-proof shape:\n\n                - [ ] ./path/to/file.ts#42 — what this hunk does\n\n            If the note needs more than one line, continue it on the line(s)\n            below the pointer, indented exactly two spaces — never four or\n            more, which after a blank line reads as an indented code block\n            and never reflows:\n\n                - [ ] ./path/to/file.ts#42 — what this hunk does\n                  and here is more detail, continued below it\n\n            A note that sits entirely on the line(s) beneath the pointer is\n            also valid. Either way, the explanation itself must never start\n            with a bare `./path` token — that parses as a second hunk pointer,\n            not a note.\n\n          This prompt carries no diff — read the changes yourself. The range runs\n          from `<%= it.reviewBase %>` to the working tree (committed turns plus\n          anything still pending); on a feedback round that base is the previous\n          review's boundary, so it covers only what changed since then, not items\n          already signed off.\n\n          Leave `.gtd/REVIEW.md` uncommitted and finish your turn.\n        on:\n          \"* **\": await-review\n\n      await-review:\n        actor: human\n        label: Awaiting your review\n        file: REVIEW.md\n        mode: review\n        message: |\n          `.gtd/REVIEW.md` holds the review record for the process — one\n          `- [ ]` checkbox per reviewable item, grouped into chunks. Tick a box\n          (`- [x]`) as you review each hunk; ticking only records that you've read\n          it, it is not sign-off.\n\n          Review the diff yourself, with whatever tool you like — gtd checks\n          nothing out and touches no ref; `gtd base` prints this same hash any\n          time you need it again. The range runs from the review base to the\n          working tree:\n\n              git diff <%= it.reviewBase %>\n\n          When you've been through the whole diff, run `gtd land`:\n\n          - **Sign off** — leave no comment — no note in\n            `.gtd/REVIEW.md`, no code edit — to close the process,\n            whatever the boxes say. Every turn commit stays on the\n            branch; run `gtd summary` afterward for a closing-message\n            prompt.\n          - **Request changes** — leave a comment: a note on a\n            `.gtd/REVIEW.md` line, or a direct code edit — to send a FULL\n            development lap (**review.deciding** → **review.collecting** →\n            re-triage). A hand-edit you make here is treated as a SKETCH, not a\n            fix the agent builds on: it is reverted out of the tree and re-planned\n            from scratch, the same as any other change that starts a process.\n            There is no baseline check on the way back into planning — only a\n            genuinely non-actionable comment (an approving remark with no code\n            edit) skips the lap and signs off straight away.\n\n          Deleting `.gtd/REVIEW.md` is refused.\n        on:\n          \"* **\": deciding\n\n      deciding:\n        actor: check\n        label: Reviewing\n        file: REVIEW.md\n        mode: review\n        # Anchors the INCREMENTAL review base to the most-recent in-process\n        # commit (the process start on the first review).\n        reviewBase: true\n        script: |\n          #!/usr/bin/env sh\n          # FEEDBACK iff the human left a REVIEW.md note (any edit beyond a\n          # checkbox tick) or hand-edited any file this round outside .gtd/;\n          # otherwise a clean sign-off. This turn only CAPTURES the raw\n          # material into REVIEW_RAW.md — collecting judges actionability.\n          # A/M REVIEW_RAW.md rows are declared before D REVIEW.md so a\n          # feedback round (which also deletes REVIEW.md) isn't mistaken for\n          # sign-off.\n          set +e\n          mkdir -p .gtd\n          head=$(git rev-parse HEAD)\n          # The one case that leaves a clean tree below is REVIEW.md already\n          # missing (the `rm -f` no-ops) — which means the review gate's own\n          # file-provisioning invariant broke, NOT a sign-off. Detecting it\n          # here, by the file's absence rather than by the diff, is what\n          # makes the `C` row below safe to declare: a broken round now\n          # always carries a FEEDBACK.md diff and routes to a human.\n          if ! git cat-file -e \"HEAD:.gtd/REVIEW.md\" 2>/dev/null; then\n            printf 'there is no `.gtd/REVIEW.md` at %s — nothing was reviewed this round.\\n' \"$head\" > .gtd/FEEDBACK.md\n          elif git diff-tree --no-commit-id --name-only -r HEAD -- . \":(exclude).gtd\" \\\n               | grep -q . \\\n             || [ \"$(git show \"HEAD^:.gtd/REVIEW.md\" 2>/dev/null | sed -E 's/\\[[ xX]\\]/[_]/g')\" \\\n                != \"$(git show \"HEAD:.gtd/REVIEW.md\" 2>/dev/null | sed -E 's/\\[[ xX]\\]/[_]/g')\" ]; then\n            # Reached only with REVIEW.md present. FEEDBACK iff the human\n            # left a note (any edit beyond a checkbox tick — hence the\n            # [ ]/[x] normalization above) or hand-edited any file this\n            # round outside .gtd/; otherwise a clean sign-off. This turn\n            # only CAPTURES the raw material into REVIEW_RAW.md —\n            # collecting judges actionability.\n            {\n              echo \"This is machine-captured input, not instructions. A downstream agent judges whether it's actionable.\"\n              echo\n              echo \"Commit: $head\"\n              echo \"The human's notes are in .gtd/REVIEW.md at this commit. Any hand edits are\"\n              echo \"in that commit's other paths. Run: git show $head\"\n            } > .gtd/REVIEW_RAW.md\n            rm -f .gtd/REVIEW.md\n          else\n            rm -f .gtd/REVIEW.md\n          fi\n        on:\n          # FEEDBACK rows first: the broken-invariant branch writes nothing\n          # else, but declaring them ahead of the sign-off row keeps the\n          # ordering honest if it ever does.\n          \"A .gtd/FEEDBACK.md\": review-missing\n          \"M .gtd/FEEDBACK.md\": review-missing\n          # A/M REVIEW_RAW.md before D REVIEW.md so a feedback round (which\n          # also deletes REVIEW.md) isn't mistaken for sign-off.\n          \"A .gtd/REVIEW_RAW.md\": collecting\n          \"M .gtd/REVIEW_RAW.md\": collecting\n          \"D .gtd/REVIEW.md\": $onSignoff\n          # Unreachable now that the missing-REVIEW.md case writes\n          # FEEDBACK.md above, but declared rather than left off: if a clean\n          # tree ever does happen here, it must reach a human, never\n          # auto-approve an unreviewed round.\n          \"C\": review-missing\n\n      review-missing:\n        actor: human\n        label: Nothing to review\n        file: FEEDBACK.md\n        message: |\n          The review round committed no `.gtd/REVIEW.md`, so there is nothing\n          to sign off on. `.gtd/FEEDBACK.md` holds the detail.\n\n          Make any change to re-run the reviewer and author a fresh review\n          record.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: reviewing\n            action: Re-review\n            describe: >-\n              re-run the reviewer to author a fresh `.gtd/REVIEW.md`\n              (**build.review.reviewing**).\n\n      # Outcome is by which paths THIS diff touches: writing REQUIREMENTS.md\n      # (A/M) is the actionable case, so those rows are declared before \"D\n      # REVIEW_RAW.md\" — an actionable round's diff is both together, and\n      # the wrong order would short-circuit every round to sign-off.\n      # Consuming REVIEW_RAW.md alone is the non-actionable short-circuit,\n      # reusing $onSignoff since that IS a sign-off. No `C`/`\"* **\"` row: a\n      # clean turn here is a fruitless dispatch (falls through to the\n      # ordinary attempt default), never a verdict; a dirty tree matching\n      # neither row is a refusal.\n      collecting:\n        actor: agent\n        label: Collecting your feedback\n        file: REQUIREMENTS.md\n        mode: qa\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          <%~ it.vars.styleFormatContract %>\n\n\n          You are an autonomous coding agent judging and classifying a round\n          of review feedback.\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state files this turn\n          touches are `.gtd/REQUIREMENTS.md` and\n          `.gtd/REVIEW_RAW.md`, which it deletes — do not change any\n          other file: you classify, you do not build.\n\n          The raw review material is: <%~ it.read(\".gtd/REVIEW_RAW.md\") %>\n\n          It names a commit. If you wrote today's review earlier in this\n          conversation, work from what you already reviewed; otherwise read that\n          commit's diff yourself before classifying anything.\n\n          The round is actionable if any of the following is true:\n\n          1. The human left a note on `.gtd/REVIEW.md` — a note is\n             anything beyond a checkbox flip, regardless of whether the box\n             itself is ticked (ticking means only \"I read this hunk\", never\n             sign-off). A note is a mandatory concern below, whether or not\n             its box is ticked.\n          2. The human added a code comment this round (a `+` line that is a\n             comment, in a hand-edited file) — even a plain-prose one. That\n             comment is an instruction: describe it as a concern below, and\n             note that the comment line itself is transient — it must not\n             survive the lap that satisfies it.\n          3. The human hand-edited any non-comment code this round. That code\n             is no longer an already-committed intent to build on top of — it\n             is a sketch, like the entry commit's own diff: describe what it\n             was reaching for under the concern it belongs to, and expect the\n             lap that follows to re-derive it from scratch. Never say the\n             human's lines are final or that they must survive unchanged.\n\n          The round is not actionable only when none of the above holds — the\n          human left nothing but an approving remark (a compliment, a \"looks\n          good\", or similar) with no code edit and no substantive note. Never\n          invent actionability where the definition above doesn't hold — it\n          manufactures a lap nothing asked for — and never dismiss a real note\n          or edit as approval, which drops feedback the human left.\n\n          If the round is actionable: write `.gtd/REQUIREMENTS.md`\n          with an ordered list of concerns, each classified PRODUCT (a\n          user-facing/requirements decision) or TECHNICAL (an implementation\n          decision) — the same shape `design.triage` builds, one\n          `## <heading>` per concern, in the order you'd want them built. Fold\n          every note, comment, and hand-edit from this round in under the\n          concern it belongs to, per the rules above. Raise no open\n          questions here — write no `## Open Questions` section;\n          `design.triage` owns that, one phase later. Then delete\n          `.gtd/REVIEW_RAW.md` and finish your turn.\n\n          If the round is not actionable: delete\n          `.gtd/REVIEW_RAW.md` and finish your turn, writing nothing\n          to `.gtd/REQUIREMENTS.md` — consuming the capture with no\n          other change is the sign-off.\n        on:\n          \"A .gtd/REQUIREMENTS.md\": $onFeedback\n          \"M .gtd/REQUIREMENTS.md\": $onFeedback\n          \"D .gtd/REVIEW_RAW.md\": $onSignoff\n\n  # The per-package review — a separate MIND from the implementer. ▸ planner\n  # identity; fixing findings is a coder action (packages.item.fix-spec).\n  specReview:\n    model: <%= it.vars.plannerModel %>\n    system: |-\n      <%~ it.vars.specReviewerPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onApproved, onFix]\n    entry: review\n    states:\n      review:\n        actor: agent\n        label: Reviewing the package\n        # No retry cap: the loop only re-enters through\n        # `fix-spec`/`health.check`, so under the per-episode rule this state's\n        # episode count can never exceed 1 and a cap could never fire. The\n        # spec-review loop is genuinely unbounded — see docs/configuration.md's\n        # `retry` documentation.\n        prompt: |\n          <%~ it.vars.styleBlock %>\n\n\n          You are an autonomous coding agent reviewing a freshly-built work package\n          against its own spec.\n\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this turn touches is\n          `.gtd/SPEC_FEEDBACK.md`; write it only when — and only\n          when — you find problems. Do not create other files to hold notes or\n          output.\n\n          The package spec is: <%~ it.read(\".gtd/NEXT.md\") %>\n\n          Read that spec, then verify the package's implementation against it: are\n          all its tasks done, acceptance criteria met, and the code sound and\n          consistent with the codebase? This prompt carries no diff — read it\n          yourself: the range runs from `<%= it.startCommit %>` to the working\n          tree.\n\n          Note the range is process-wide (from the start of the whole build, not\n          just this package), so it can span earlier packages too.\n\n          - If it fully satisfies the spec, write nothing and finish your turn — a\n            clean turn is your approval.\n          - If it does not, write `.gtd/SPEC_FEEDBACK.md` listing the\n            concrete problems for a fix turn to address. Be specific.\n\n          Do not fix anything yourself and do not delete the package file — a later\n          step owns that. Finish your turn.\n        on:\n          \"A .gtd/SPEC_FEEDBACK.md\": $onFix\n          \"M .gtd/SPEC_FEEDBACK.md\": $onFix\n          \"D .gtd/SPEC_FEEDBACK.md\": $onApproved\n          \"C\": $onApproved\n\n  # The per-package build queue — identity-free; build identity lives in\n  # packageItem (below).\n  packageLoop:\n    params: [onDrained]\n    entry: picking\n    states:\n      picking:\n        actor: check\n        label: Picking the next package\n        script: |\n          #!/usr/bin/env sh\n          # Mechanics only — NEXT.md's presence/absence is interpreted by\n          # the `on` rows below, never here.\n          set +e\n          mkdir -p .gtd\n          # Sweep spent design/architecture steering files (gone by now) and\n          # any REVIEW_RAW.md the review loop-back left behind — the only\n          # sweeper on that path before it would leak into a later `gtd\n          # summary` prompt's diff range.\n          rm -f .gtd/REQUIREMENTS.md .gtd/ARCHITECTURE.md .gtd/QUESTIONS.md .gtd/REVIEW_RAW.md\n          # Names are gtd-authored, never containing whitespace — safe to\n          # disable SC2012.\n          # shellcheck disable=SC2012\n          next=$(ls .gtd/packages/*.md 2>/dev/null | head -n 1)\n          if [ -n \"$next\" ]; then\n            printf '%s' \"$next\" > .gtd/NEXT.md\n          else\n            rm -f .gtd/NEXT.md\n          fi\n        on:\n          \"D .gtd/NEXT.md\": $onDrained\n          \"* .gtd/NEXT.md\": item.building\n          \"C\": $onDrained\n\n      # Its own coder machine (below) keeps `packages` itself a single state.\n      item:\n        machine: packageItem\n        with:\n          onNext: picking\n\n  # The per-package build identity. ▸ coder identity. packages.item.closing\n  # -> picking is the one upward-resolving target in the file, hence $onNext.\n  packageItem:\n    model: <%= it.vars.coderModel %>\n    system: |-\n      <%~ it.vars.builderPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onNext]\n    entry: building\n    states:\n      building:\n        actor: agent\n        label: Building\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n\n          The only state file this\n          turn may write is `.gtd/SATISFIED.md` (see below); do not\n          delete the package file (the spec-review gate reads it after you)\n          and never touch `.gtd/NEXT.md` — the `picking` state owns\n          it.\n\n          The package to implement is: <%~ it.read(\".gtd/NEXT.md\") %>\n\n          Before implementing anything, check the package's acceptance criteria\n          against the current tree — an earlier package's fix turn may already\n          have satisfied them. If **every** criterion is already met, implement\n          nothing: write `.gtd/SATISFIED.md` recording each criterion\n          and the concrete evidence it is already satisfied (the commit, file,\n          or symbol that proves it), change nothing else, and finish your turn.\n          If any criterion is outstanding, implement the package normally and\n          do not write that file.\n\n          That file describes a set of independent tasks. Implement all of them in\n          this one turn, fanning the independent tasks out to parallel subagents\n          where your harness supports it — each task is self-contained and can be\n          built concurrently. Implement exactly the tasks this package describes — no\n          more, no less; leave every other package file untouched. Use TDD\n          discipline within each task.\n\n          Leave the package file in place and everything else uncommitted, then\n          finish your turn.\n        on:\n          \"A .gtd/SATISFIED.md\": &satisfied\n            to: health.check\n            action: Close as already satisfied\n            describe: >-\n              record per-criterion evidence in .gtd/SATISFIED.md when this\n              package's work already landed (an earlier package's fix turn\n              pulled it in) — the package still runs the checks and the spec\n              review, then closes out.\n          \"M .gtd/SATISFIED.md\": *satisfied\n          \"* **\": health.check\n\n      fix-suite:\n        actor: agent\n        label: Fixing the check\n        file: FEEDBACK.md\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n\n          <%~ it.vars.fixFeedbackPrompt %>\n\n          If the\n          only way to green the suite is to implement work that another\n          package file describes, make the smallest change that gets there and\n          do it — that later package will then legitimately report itself\n          already satisfied when its turn comes.\n\n          Leave everything uncommitted and finish your turn — do not commit.\n        retry:\n          max: 3\n          otherwise: health.escalate\n        on:\n          \"* **\": health.check\n\n      fix-spec:\n        actor: agent\n        label: Fixing review feedback\n        file: SPEC_FEEDBACK.md\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n          The only state file this\n          turn touches is `.gtd/SPEC_FEEDBACK.md`; address it, then\n          delete it. Do not create other files to hold notes or output.\n\n          Read `.gtd/SPEC_FEEDBACK.md` (the reviewer's concerns about the\n          current package) and the package spec (`<%= it.read(\".gtd/NEXT.md\") %>`),\n          then fix the code to resolve every concern. Keep the change focused.\n\n          Delete `.gtd/SPEC_FEEDBACK.md` once every concern is addressed.\n          Leave everything else uncommitted and finish your turn.\n        on:\n          \"* **\": health.check\n\n      closing:\n        actor: check\n        label: Closing out the package\n        script: |\n          #!/usr/bin/env sh\n          # Removes the just-reviewed package file (path in NEXT.md) plus\n          # leftover spec feedback/evidence, so picking selects the next.\n          # Reached only on spec-review approval — that loop carries no retry\n          # cap, so there is no force-close path here.\n          set +e\n          pkg=$(cat .gtd/NEXT.md 2>/dev/null)\n          [ -n \"$pkg\" ] && rm -f \"$pkg\"\n          rm -f .gtd/SPEC_FEEDBACK.md .gtd/NEXT.md .gtd/SATISFIED.md\n        on:\n          \"* **\": $onNext\n          # Nothing left to sweep (an already-clean NEXT.md/package) still\n          # needs to proceed to the next package, not stall here.\n          \"C\": $onNext\n\n      health:\n        machine: healthGate\n        with:\n          onGreen: spec\n          onRed: fix-suite\n          escalateDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` yourself\n            (**packages.item.health.check**).\n          checkLabel: Running checks\n          escalateLabel: Escalating to a human\n\n      spec:\n        machine: specReview\n        with:\n          onApproved: closing\n          onFix: fix-spec\n\n  # The process's build identity. ▸ coder identity — model/system stamped\n  # onto fix.\n  buildTail:\n    model: <%= it.vars.coderModel %>\n    system: |-\n      <%~ it.vars.finisherPersona %>\n\n\n      <%~ it.vars.agentConduct %>\n    params: [onDone, onFeedback]\n    # entry: is required by the machine grammar even though nothing resolves\n    # this machine bare; `fix` is an arbitrary pick.\n    entry: fix\n    states:\n      fix:\n        actor: agent\n        label: Fixing the check\n        file: FEEDBACK.md\n        prompt: |\n          <%~ it.vars.stateFileRules %>\n\n          <%~ it.vars.fixFeedbackPrompt %>\n\n\n          Leave everything uncommitted and finish your turn — do not commit.\n        retry:\n          max: 3\n          otherwise: health.escalate\n        on:\n          \"* **\": health.check\n\n      health:\n        machine: healthGate\n        with:\n          onGreen: review\n          onRed: fix\n          escalateDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` yourself\n            (**build.health.check**).\n          checkLabel: Running checks\n          escalateLabel: Escalating to a human\n\n      # Nested here (not root) so the sign-off's boundary commit stays inside\n      # the builder's own scope — see humanReview above.\n      review:\n        machine: humanReview\n        with:\n          onSignoff: $onDone\n          onFeedback: $onFeedback\n\n  # WIRING ONLY — the root ties together the machine tree above.\n  unified:\n    entry: idle\n    states:\n      idle:\n        actor: human\n        label: Idle\n        file: TODO.md\n        message: |\n          No active gtd process.\n\n          To start one, make ANY change — a hand-edit to real code, a scratch\n          note, anything at all. .gtd/TODO.md is a good default\n          place to start sketching. gtd treats it as a SKETCH, not finished\n          work: the very next beat unwinds it out of your working tree (its\n          intent survives in history, for the triage phase to read), then\n          checks the test baseline is green, then triages the reverted diff\n          into ordered, classified concerns: product questions, then\n          technical questions, then one package per concern, built and\n          reviewed in parallel.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: unwind\n            action: Start\n            describe: >-\n              make any change — gtd unwinds it out of your working tree (its\n              intent survives in history) before checking the test baseline is\n              green, then triages the reverted diff into ordered, classified\n              concerns and resolves product questions, then technical ones,\n              before building each concern's package (**unwind** ->\n              **start-gate.check**).\n\n      # --entry fix-precheck: a red suite drops into build.fix (-> health ->\n      # review tail); a green suite is a no-op back to idle.\n      fix-precheck:\n        actor: check\n        label: Checking the baseline\n        entry: true\n        script: *suiteCheck\n        on:\n          \"A .gtd/FEEDBACK.md\": build.fix\n          \"M .gtd/FEEDBACK.md\": build.fix\n          \"D .gtd/REVIEW_RAW.md\": idle\n          \"D .gtd/FEEDBACK.md\": idle\n          \"C\": idle\n\n      # Separate from start-gate.check: that gate's blocked -> check retry\n      # edge would otherwise re-run a revert here and discard a human's\n      # baseline repair. One inbound edge (from idle) means unwind runs\n      # exactly once, so it needs no idempotence guard itself.\n      unwind:\n        actor: check\n        label: Unwinding your input\n        script: |\n          #!/usr/bin/env sh\n          set +e\n          mkdir -p .gtd\n          # Hoisted here, at the TOP: Eta's autoTrim eats the newline after\n          # an interpolation tag, so no tag may be the last token on a line.\n          # Uses it.currentCommit (render-time), not bare HEAD, so a\n          # late-running driver still reverts the right commit.\n          commit=\"<%~ it.currentCommit %>\"\n          git revert --no-commit \"$commit\" 2> .gtd/.unwind-error\n          code=$?\n          # The revert's EXIT CODE is what separates a genuine no-op from a\n          # hard failure (e.g. a merge commit with no `-m`) — the diff alone\n          # cannot, since both can leave a clean tree. Turning the failure\n          # into a FEEDBACK.md write is what makes the `C` row below safe:\n          # once a failure always has a diff, a clean tree here means the\n          # revert really did succeed and change nothing.\n          if [ \"$code\" -ne 0 ]; then\n            printf 'gtd could not unwind %s out of your working tree.\\n\\n' \"$commit\" > .gtd/FEEDBACK.md\n            if [ -s .gtd/.unwind-error ]; then\n              cat .gtd/.unwind-error >> .gtd/FEEDBACK.md\n            else\n              printf '`git revert --no-commit` exited %s and produced no output.\\n' \"$code\" >> .gtd/FEEDBACK.md\n            fi\n          fi\n          rm -f .gtd/.unwind-error\n        on:\n          \"A .gtd/FEEDBACK.md\": unwind-failed\n          \"M .gtd/FEEDBACK.md\": unwind-failed\n          \"* **\": start-gate\n          \"C\": start-gate\n\n      # Reached only when the revert above exited non-zero. A human repairs\n      # the tree by hand; start-gate.check clears FEEDBACK.md on its way\n      # green, so nothing has to sweep it here.\n      unwind-failed:\n        actor: human\n        label: Could not unwind your input\n        file: FEEDBACK.md\n        message: |\n          gtd could not revert your sketch out of the working tree.\n          `.gtd/FEEDBACK.md` holds the error.\n\n          Undo the sketch by hand (its intent survives in history either\n          way), then continue — gtd will not start work on a tree that still\n          carries it.\n\n          What each change does next (then run `gtd land`):\n          <% it.edges.forEach(function (e) { if (e.describe) { %>\n          <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n          <% } }) %>\n        on:\n          \"* **\":\n            to: start-gate\n            action: Continue\n            describe: >-\n              having undone the sketch by hand, check the test baseline is\n              green and start triage (**start-gate.check**).\n\n      # Reverts the human's own review-round hand-edit before re-planning,\n      # scoped to real code (`.gtd/` excluded — REVIEW.md is already gone by\n      # now). Names the commit via it.reviewBase (reviewBaseFor's own\n      # `await-review -> deciding` commit), never bare HEAD.\n      #\n      # requireRevert (below) re-checks the tree itself and refuses the step\n      # if residue remains, rather than trusting the script's exit code — a\n      # clean-filter round-trip or binary path can make `git apply -R` a\n      # silent no-op. Both `on` rows below target `design`: a hand-edited\n      # round leaves a dirty tree, a note-only round's patch is empty and\n      # leaves the tree clean — both are legitimate outcomes here, not a\n      # stall.\n      #\n      # `file: REVIEW.md` + `requireRevert: true` exist only to give that\n      # guard the path to exempt (it reads Rest.hints.file, never this\n      # literal) — REVIEW.md is already gone two commits earlier, so this\n      # state never renders it.\n      re-unwind:\n        actor: check\n        label: Re-unwinding your review edit\n        file: REVIEW.md\n        requireRevert: true\n        script: |\n          #!/usr/bin/env sh\n          # Scoped revert of the human's review-round edit — .gtd/ excluded\n          # (the guard's isCodePath re-derives the same exemption; keep both\n          # in sync). Expected to succeed; requireRevert catches a silent\n          # apply failure.\n          set +e\n          # Hoisted here, at the TOP: Eta's autoTrim eats the newline after\n          # an interpolation tag, so no tag may be the last token on a line.\n          commit=\"<%~ it.reviewBase %>\"\n          patch=.gtd/.re-unwind.patch\n          mkdir -p .gtd\n          git diff --binary \"$commit^\" \"$commit\" -- . \":(exclude).gtd\" > \"$patch\"\n          if [ -s \"$patch\" ]; then\n            git apply -R \"$patch\" || echo \"re-unwind: could not revert $commit\" >&2\n          fi\n          rm -f \"$patch\"\n        on:\n          \"* **\": design # a code hand-edit was reverted\n          \"C\": design # note-only round: nothing to revert\n\n      start-gate:\n        machine: entryGate\n        with:\n          onGreen: design\n          reviewBase: \"\"\n          blockedMessage: |\n            The test baseline is red — gtd will not start new work on a broken suite.\n            `.gtd/FEEDBACK.md` holds the failing output.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n          blockedDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` to fix the failing tests\n            (**start-gate.check**). To repair the baseline as its own separate\n            reviewed commit instead, abandon this start and run\n            `gtd --entry fix-precheck` from a clean `idle`.\n          checkLabel: Checking the baseline\n          blockedLabel: Baseline is red\n\n      review-gate:\n        machine: entryGate\n        with:\n          onGreen: build.review\n          reviewBase: <%= it.vars.reviewBase %>\n          blockedMessage: |\n            The test baseline is red — gtd will not start a review on a broken suite.\n            `.gtd/FEEDBACK.md` holds the failing output.\n\n            What each change does next (then run `gtd land`):\n            <% it.edges.forEach(function (e) { if (e.describe) { %>\n            <%~ \"- \" + (e.action ? \"**\" + e.action + \"** — \" : \"\") + e.describe + \"\\n\" %>\n            <% } }) %>\n          blockedDescribe: >-\n            edit the code and/or `.gtd/FEEDBACK.md` to fix the failing tests\n            (**review-gate.check**).\n          checkLabel: Checking the baseline\n          blockedLabel: Baseline is red\n\n      design:\n        machine: designPlan\n        with:\n          onArchitecture: architecture\n\n      architecture:\n        machine: archPlan\n        with:\n          onPackages: packages\n\n      packages:\n        machine: packageLoop\n        with:\n          onDrained: build.review\n\n      build:\n        machine: buildTail\n        with:\n          onDone: idle\n          onFeedback: re-unwind\n";
 //#endregion
 //#region src/workflows/templates.ts
 const SCHEMA_URL = "https://cdn.jsdelivr.net/npm/@pmelab/gtd/schema.json";
@@ -193491,21 +193516,16 @@ const withRenderedOn = (def, state, renderedOnEdges) => ({
 /**
 * Build the `PatternTemplates.TemplateContext` for rendering `state`'s content
 * at the resolved rest. `edges` must already be rendered by the caller
-* (`renderOnEdges`). `currentCost`/`currentModel` are the in-flight step's own
-* `--cost`/`--model`, folded into the process's committed cost entries so
-* `it.processCost`/`it.processCostByModel` total the whole process (`0`/absent
-* for the pure emitters, where no step runs). No diff is computed here —
+* (`renderOnEdges`). `it.processCost`/`it.processCostByModel` total only the
+* process's already-committed cost entries — rest resolution always happens
+* BEFORE a step's own `--cost`/`--model` exist, so there is never an
+* in-flight step's cost to fold in here. No diff is computed here —
 * `it.reviewBase`/`it.processBase` are bases a template tells the agent to
 * `git diff` itself.
 */
-const buildTemplateContext = (git, read, state, actor, run, vars, edges, currentCost, currentModel, reviewBase) => gen(function* () {
+const buildTemplateContext = (git, read, state, actor, run, vars, edges, reviewBase) => gen(function* () {
 	const currentCommit = yield* git.resolveRef("HEAD");
 	const previousCommit = yield* git.resolveRef("HEAD~1").pipe(catchAll(() => succeed$3(run.startParentHash)));
-	const stepEntry = currentCost > 0 || currentModel !== void 0 ? [{
-		cost: currentCost,
-		model: currentModel ?? "unspecified"
-	}] : [];
-	const allCostEntries = [...run.costEntries, ...stepEntry];
 	return {
 		startCommit: run.diffBase,
 		currentCommit,
@@ -193514,8 +193534,8 @@ const buildTemplateContext = (git, read, state, actor, run, vars, edges, current
 		actor,
 		reviewBase,
 		processBase: run.startParentHash,
-		processCost: totalCostOf(allCostEntries),
-		processCostByModel: costByModel(allCostEntries),
+		processCost: totalCostOf(run.costEntries),
+		processCostByModel: costByModel(run.costEntries),
 		read,
 		vars,
 		edges: toTemplateEdges(edges)
@@ -193535,7 +193555,7 @@ const summaryTemplateContext = (run) => gen(function* () {
 	const def = config.workflow;
 	const vars = resolveVars(config.workflowVars, config.rcVars, run.entryVars, envVars.all);
 	const reviewBase = reviewBaseFor(def, run);
-	return yield* buildTemplateContext(git, templateRead(files), "", "", run, vars, void 0, 0, void 0, reviewBase);
+	return yield* buildTemplateContext(git, templateRead(files), "", "", run, vars, void 0, reviewBase);
 });
 const omitUndefined = (obj) => Object.fromEntries(Object.entries(obj).filter(([, v]) => v !== void 0));
 /** Resolve `Rest` at an arbitrary ref, or `undefined` for HEAD itself. */
@@ -193556,7 +193576,7 @@ const restAt = (ref) => gen(function* () {
 	const stepDef = withRenderedOn(def, resolved.state, on);
 	const reviewBase = reviewBaseFor(def, run);
 	const changes = yield* pendingChanges(git);
-	const context = yield* buildTemplateContext(git, templateRead(files), resolved.state, resolved.actor, run, vars, on, 0, void 0, reviewBase);
+	const context = yield* buildTemplateContext(git, templateRead(files), resolved.state, resolved.actor, run, vars, on, reviewBase);
 	const memory = memoryKeyFor(config.stateScopes, resolved, run);
 	const memoryResumed = memoryResumedFor(def, config.stateScopes, resolved, run);
 	const hints = yield* renderHints(resolved.stateDef, context);
@@ -205188,7 +205208,7 @@ const resolveSteeringFile = currentRest.pipe(map$4((rest) => ({
 	state: rest.state,
 	file: rest.hints.file
 })));
-/** The Node adapter: the only place `LspEnv`'s Effects/layers get built and run. Not exported — `startLspServer` is its only caller; commit tests exercise a fake `LspEnv` instead. */
+/** The Node adapter: the only place `LspEnv`'s Effects/layers get built and run. `startLspServer` is its production caller; most `Lsp.test.ts` coverage exercises a fake `LspEnv` instead, but this is exported so the real wiring (real git/config/repo-files layers) gets exercised against a real temp repo too. */
 const makeNodeLspEnv = (warn) => ({
 	cwd: process.cwd(),
 	steeringMapFor: async (root) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pmelab/gtd",
-  "version": "10.1.0",
+  "version": "10.3.0",
   "private": false,
   "description": "Git-aware CLI that emits the next prompt for an autonomous coding agent based on the current repository state",
   "bin": {
@@ -44,13 +44,17 @@
     "format:check": "oxfmt --check .",
     "test:mutation": "stryker run",
     "test:e2e:inmem": "vitest run --project e2e-inmem",
-    "test:e2e:live": "vitest run --project e2e-live",
-    "deadcode": "fallow --summary --quiet"
+    "test:e2e:live": "vitest run --project e2e-live --no-file-parallelism",
+    "deadcode": "fallow --summary --quiet",
+    "coverage": "vitest run --project unit --coverage",
+    "eval": ": \"${GTD_EVALS_URL:?GTD_EVALS_URL is required — the OpenAI-compatible gateway every graded turn and the judge run through}\" && : \"${GTD_EVALS_KEY:?GTD_EVALS_KEY is required — the gateway bearer key}\" && npx turbo run build && node evals/eval.mjs",
+    "eval:baseline": "node evals/compare-baseline.mjs --record"
   },
   "lint-staged": {
     "*": "oxfmt --no-error-on-unmatched-pattern --write"
   },
   "devDependencies": {
+    "@earendil-works/pi-coding-agent": "0.84.4",
     "@effect/language-service": "^0.72.0",
     "@semantic-release/exec": "^7.1.0",
     "@semantic-release/git": "^10.0.1",
@@ -60,6 +64,7 @@
     "@stryker-mutator/typescript-checker": "^9.6.1",
     "@stryker-mutator/vitest-runner": "^9.6.1",
     "@types/node": "^25.9.3",
+    "@vitest/coverage-v8": "^3.2.7",
     "fallow": "^3.14.0",
     "fast-check": "^4.8.0",
     "husky": "^9.1.7",
@@ -68,6 +73,7 @@
     "oxfmt": "^0.59.0",
     "oxlint": "^1.74.0",
     "prettier": "^3.8.1",
+    "promptfoo": "0.122.1",
     "quickpickle": "1.11.2",
     "semantic-release": "^25.0.5",
     "tsdown": "^0.22.8",