@pmelab/gtd 46.0.0 → 48.0.0

package/dist/testkit.mjs

@@ -6154,8 +6154,15 @@ export CLAUDE_CODE_DISABLE_BACKGROUND_TASKS=1
 export BASH_DEFAULT_TIMEOUT_MS=1800000
 export BASH_MAX_TIMEOUT_MS=7200000
 
+# gtd's paths, scripts and checks are relative to the repository root.
+cd "$(git rev-parse --show-toplevel)"
+
 beat=1
 while :; do
+  cost=
+  model=
+  judged_before="\${judged:-}"
+  judged=
   # One value per invocation: \`gtd next --json=<path>\` is a pure read (peek
   # or dispatch, indistinguishable, always safe to call again), so nothing is
   # lost by reading \`kind\`/\`idle\` as two separate calls rather than one
@@ -6183,14 +6190,22 @@ while :; do
     # \`gtd judge run\` and land the verdict via \`gtd judge answer\`. POSIX sh has
     # no pipefail, so each step is captured first and only piped on when it
     # succeeded; if \`gtd judge run\` fails or \`gtd judge answer\` rejects the
-    # verdict, its stderr reaches you and we fall through to the human hand-off.
+    # verdict, its stderr reaches you and the gate is handed to you — on the
+    # opening beat too, so it never lands unanswered — as it is when the gate
+    # we just answered comes straight back.
     message)
-      if [ -n "$(gtd next --json=judge)" ] &&
-        verdict="$(gtd judge --json | gtd judge run)" &&
-        script="$(printf '%s\\n' "$verdict" | gtd judge answer --json=script)"; then
-        printf '%s\\n' "$script" | sh
-        beat=$((beat + 1))
-        continue
+      if [ -n "$(gtd next --json=judge)" ]; then
+        state="$(gtd next --json=state)"
+        if [ "$state" != "$judged_before" ] &&
+          verdict="$(gtd judge --json | gtd judge run)" &&
+          script="$(printf '%s\\n' "$verdict" | gtd judge answer --json=script)"; then
+          printf '%s\\n' "$script" | sh
+          judged="$state"
+          beat=$((beat + 1))
+          continue
+        fi
+        gtd next
+        exit 0
       fi
       [ "$beat" = 1 ] || {
         gtd next
@@ -6209,45 +6224,83 @@ while :; do
       model="$(gtd next --json=model)"
       system="$(gtd next --json=system)"
       validate="$(gtd next --json=validate)"
+      context="$(gtd next --json=context)"
       log="$(gtd next --json=log)"
+      # Absolute, since the hook runs wherever the agent last cd'ed to.
+      case "$log" in /*) ;; *) log="$(git rev-parse --show-toplevel)/$log" ;; esac
       # \`model\` is a hint, not a model name — the bundled workflow ships
       # \`smart\` and \`base\`, which \`claude --model\` rejects. Resolve them
-      # with the variables \`gtd loop\` reads; unset, the hint is dropped and
-      # the agent runs its own default. Any other value is a real model name.
+      # from your environment; unset, the hint is dropped and the agent runs
+      # its own default. Any other value is a real model name.
       case "$model" in
         smart) model="\${GTD_LOOP_MODEL_SMART:-}" ;;
         base) model="\${GTD_LOOP_MODEL_BASE:-}" ;;
       esac
-      # The prompt embeds a full diff, so it goes to the agent over stdin,
-      # never as an argv positional — argv is capped (~1 MB on macOS, and
-      # POSIX guarantees only 4 KB, ARG_MAX), and a diff crosses that far
-      # sooner than you'd expect. Piping plain \`gtd next\` (not
-      # --json=content) means the diff is rendered once, for this input,
-      # rather than twice.
-      agent_turn() { gtd next | claude -p "$1" "$session_id" \\
-        \${model:+--model "$model"} \\
-        \${system:+--system-prompt "$system"} \\
-        --dangerously-skip-permissions >>"$log" 2>&1; }
+      # The step's context limit, as a Claude Code hook after every tool
+      # call and before compacting. Past \`hard\` it stops the turn and
+      # appends why to the stop file, so we land nothing and stop too.
+      soft="$(printf '%s' "$context" | sed -n 's/.*"soft":\\([0-9.]*\\).*/\\1/p')"
+      hard="$(printf '%s' "$context" | sed -n 's/.*"hard":\\([0-9.]*\\).*/\\1/p')"
+      window="\${GTD_LOOP_CONTEXT_WINDOW:-}"
+      case "$model" in *'[1m]'* | *'[1M]'*) window="\${window:-1000000}" ;; esac
+      hook="gtd context-hook\${soft:+ --soft $soft}\${hard:+ --hard $hard}\${window:+ --window $window}"
+      hooks="[{\\"hooks\\":[{\\"type\\":\\"command\\",\\"command\\":\\"$hook\\"}]}]"
+      settings="{\\"hooks\\":{\\"PostToolUse\\":$hooks,\\"PreCompact\\":$hooks}}"
+      stop="$log.context-stop"
+      rm -f "$stop"
+      # One turn, its prompt on stdin. The prompt embeds a full diff, so it
+      # never goes on argv — argv is capped (~1 MB on macOS, and POSIX
+      # guarantees only 4 KB, ARG_MAX), and a diff crosses that far sooner
+      # than you'd expect. Every turn passes the same $model/$system:
+      # whether \`claude --resume\` re-applies the original session's is a
+      # harness detail gtd cannot verify from outside. Its JSON envelope
+      # goes to the log, and its \`total_cost_usd\` adds up into $cost. An
+      # \`is_error\` envelope (max turns, an error mid-turn) or one another
+      # hook stopped fails the turn even at exit 0: landing it would land
+      # the failure as work. Only a turn that never ran (no envelope: the
+      # session flag was refused) is worth retrying under the other flag.
+      turn() {
+        out="$(GTD_LOOP_CONTEXT_STOP_FILE="$stop" claude -p "$@" \\
+          \${model:+--model "$model"} \\
+          \${system:+--system-prompt "$system"} \\
+          --settings "$settings" --output-format json \\
+          --dangerously-skip-permissions 2>>"$log")" && ok=0 || ok=$?
+        printf '%s\\n' "$out" >>"$log"
+        case "$out" in *'"type":"result"'*) ran=true ;; *) ran=false ;; esac
+        case "$out" in
+          *'"is_error":true'* | *'"terminal_reason":"hook_stopped"'*) ok=1 ;;
+        esac
+        usd="$(printf '%s\\n' "$out" | sed -n 's/.*"total_cost_usd":\\([0-9.eE+-]*\\).*/\\1/p' | tail -n 1)"
+        [ -z "$usd" ] || cost="$(awk -v a="\${cost:-0}" -v b="$usd" 'BEGIN { printf "%.6f", a + b }')"
+        [ "\${cost:-}" != 0.000000 ] || cost=
+        if [ -s "$stop" ]; then
+          cat "$stop" >&2
+          exit 1
+        fi
+        return "$ok"
+      }
+      # Plain \`gtd next\` (not --json=content) renders the diff once.
+      gtd next >"$log.prompt"
       if [ "\${resume:-}" = true ]
-      then agent_turn --resume || agent_turn --session-id
-      else agent_turn --session-id || agent_turn --resume
+      then first=--resume second=--session-id
+      else first=--session-id second=--resume
       fi
+      turn "$first" "$session_id" <"$log.prompt" ||
+        { [ "$ran" = false ] && turn "$second" "$session_id" <"$log.prompt"; } || {
+        echo "agent turn failed — see $log" >&2
+        exit 1
+      }
       n=0
       while [ -n "\${validate:-}" ] && ! fix="$(sh -c "$validate" 2>&1)"; do
         n=$((n + 1)) && [ "$n" -gt 3 ] && { printf '%s\\n' "$fix" >&2; exit 1; }
-        # $fix IS the fix prompt, verbatim — piped for the same reason the
-        # agent's own turn above is. Whether \`claude --resume\` re-applies the
-        # original session's model/system prompt is a harness detail gtd
-        # cannot verify from outside, so this passes the identical
-        # $model/$system on both calls rather than assume it does —
-        # otherwise this fix turn might silently fall back to Claude Code's
-        # own defaults while the turn that produced the file ran under the
-        # workflow's own model and persona.
-        printf '%s' "$fix" | claude -p --resume "$session_id" \\
-          \${model:+--model "$model"} \\
-          \${system:+--system-prompt "$system"} \\
-          --dangerously-skip-permissions >>"$log" 2>&1
-      done ;;
+        # $fix IS the fix prompt, verbatim.
+        printf '%s' "$fix" >"$log.prompt"
+        turn --resume "$session_id" <"$log.prompt" || {
+          echo "fix turn failed — see $log" >&2
+          exit 1
+        }
+      done
+      rm -f "$log.prompt" ;;
   esac
 
   # Read every value the landing needs BEFORE piping the script to \`sh\`:
@@ -6257,7 +6310,8 @@ while :; do
   # before the script runs and changes the tree out from under them.
   settled="$(gtd land --json=settled)"
   idle="$(gtd land --json=idle)"
-  gtd land --json=script | sh
+  # A prompt beat's turns record their cost, tagged with the model that ran.
+  gtd land --json=script \${cost:+--cost="$cost"} \${cost:+\${model:+--model="$model"}} | sh
   if [ "\${settled:-}" = true ]; then
     [ "\${idle:-}" = true ] && gtd next
     exit 0
@@ -10893,7 +10947,6 @@ const runInstallCommand = (out) => gen(function* () {
 	yield* (yield* Workspace).write(WORKFLOW_MODULE, CONFIG_SCAFFOLD);
 	yield* (yield* Narrator).warn(`gtd install: wrote ${WORKFLOW_MODULE}, importing gtd's shipped workflows — commit it before your first \`gtd land\``);
 });
-const runLoopCommand = () => flatMap$4(ToolCommands, (tools) => tools.loop(selfGtd()));
 /**
 * Exactly one trailing newline — duplicated from `Cli.ts`'s own
 * `normalizeTrailingNewline` rather than imported, to avoid a circular value
@@ -11471,7 +11524,6 @@ const needsOf = (kind) => {
 	switch (kind) {
 		case "lsp":
 		case "install":
-		case "loop":
 		case "contextHook":
 		case "judgeRun":
 		case "strays": return "none";
@@ -11511,7 +11563,6 @@ const dispatchVoidCommand = (command, json, out) => {
 			return runCheckCommand(command.mode, command.file, command.openQuestions ?? false, command.openThreads ?? false);
 		case "uncheck": return runUncheckCommand(command.file);
 		case "install": return runInstallCommand(out);
-		case "loop": return runLoopCommand();
 		case "contextHook": return runContextHookCommand(command, out);
 		case "summary": return runSummaryCommand(out, command.slice);
 		case "base": return runBaseCommand(out);
@@ -12038,20 +12089,6 @@ const COMMAND_ROWS = [
 			"else it writes nothing."
 		]
 	},
-	{
-		token: "loop",
-		kind: "loop",
-		arity: "none",
-		details: [
-			"Drive this worktree's process until it rests on you: read",
-			"each beat, run the agent turn or script it demands, and",
-			"land the result. Exits 0 at a human gate, at idle and when",
-			"a landing settles; exits 1 on a stall or a failed turn or",
-			"landing. The agent CLI and the model each workflow hint",
-			"maps to are per-machine settings (GTD_LOOP_* variables, or",
-			"~/.config/gtd/loop.json); see docs/driver.md"
-		]
-	},
 	{
 		token: "context-hook",
 		kind: "contextHook",
@@ -13403,7 +13440,6 @@ const missingTool = (name) => () => fail$5(/* @__PURE__ */ new Error(`gtd ${name
 const missingToolCommands = {
 	ui: missingTool("ui"),
 	lsp: missingTool("lsp"),
-	loop: missingTool("loop"),
 	contextHook: missingTool("context-hook"),
 	judge: () => Promise.reject(/* @__PURE__ */ new Error("gtd judge run: ToolCommands has no test double wired into testLayers()"))
 };

package/dist/workflows/cleanup.mjs

@@ -1,6 +1,6 @@
 import { agent, changes, glob, read, refuse, scope, start, vars } from "../flows.mjs";
-import { Bt as requireFlag, C as coder, Lt as flagReferences, Tt as envDefaults, i as skills, n as summary, o as flagShipped, r as access, t as steering, wt as defaults, zt as registryEntry } from "../chunks/reserved-BENPy6IJ.mjs";
-import { a as refusePendingHandOff, r as handOffTail } from "../chunks/tail-CDNMZ-2E.mjs";
+import { Ct as defaults, It as flagReferences, Rt as registryEntry, S as coder, i as skills, n as summary, o as flagShipped, r as access, t as steering, wt as envDefaults, zt as requireFlag } from "../chunks/reserved-DF1HD6jL.mjs";
+import { a as refusePendingHandOff, r as handOffTail } from "../chunks/tail-D89fU_wf.mjs";
 //#region src/workflows/cleanup/cleanup.ts
 const removingPrompt = (name, registry) => `Clean up the release flag "${name}", which is default-on: delete its entry in ${registry}, and remove every \`GTD_FLAG_${name.toUpperCase()}\` check from the shipped sources. Keep the flag-on path; delete code only the off path reached, including tests that pin the off path.`;
 const cleanupFlow = async () => {

package/dist/workflows/feature.mjs

@@ -1,8 +1,8 @@
 import { agent, answered, applyScript, changes, changesSince, check, codeChanges, env, hasThreadFor, head, human, judge, judgeDiffShape, kept, mark, matches, quote, read, refuse, removeScript, requireAnswers, requireProgress, requireReplies, requireThreadsClosed, revertScript, run, scope, slice, start, vars, wrote } from "../flows.mjs";
 import { lensEntriesOf } from "../workflow-kit.mjs";
-import { $ as naiveApplyFailure, B as coverageMessage, C as coder, Ct as unwindFailure, E as fixNits, Et as testFileGlobs, G as designSystem, H as coveringPrompt, I as builderSystem, K as designTriagePrompt, L as clearCriteriaMessage, N as agentWithSkills, P as buildReviewCollectingPrompt, Q as implementFixPrompt, R as clearCriteriaQuestion, St as unwindFailedMessage, T as fixCheck, Tt as envDefaults, U as criterionLine, V as coverageQuestion, W as designGateAnswerMessage, X as idleMessage, Z as implementBuildingPrompt, _ as REFACTOR_NOTE, _t as seamByHandMessage, a as flagModeOn, at as qualityUnresolvedMessage, b as SLICES, bt as startGateBlockedMessage, ct as refactorFixPrompt, et as naivePassPrompt, f as healthy, g as REFACTOR, gt as reviewerSystem, h as QUALITY, i as skills, k as planner, lt as refactorNotePrompt, m as HANDOFF, n as summary, nt as oversizedMessage, ot as refactorBuildingPrompt, p as FEEDBACK, r as access, s as shipped, st as refactorDroppedRefusal, t as steering, tt as normaliseCriteriaScript, u as gate, ut as refactorRules, v as REQUIREMENTS, vt as seamSteerMessage, w as collecting, wt as defaults, x as answerReviewQuestions, xt as uncoveredMessage, yt as slicingPrompt } from "../chunks/reserved-BENPy6IJ.mjs";
-import { a as refusePendingHandOff, o as tail } from "../chunks/tail-CDNMZ-2E.mjs";
-import { i as reUnwind, n as humanReview, r as isCode$1, t as buildTail } from "../chunks/review-DcbgUwr7.mjs";
+import { $ as naivePassPrompt, B as coverageQuestion, C as collecting, Ct as defaults, F as builderSystem, G as designTriagePrompt, H as criterionLine, I as clearCriteriaMessage, L as clearCriteriaQuestion, M as agentWithSkills, N as buildReviewCollectingPrompt, O as planner, Q as naiveApplyFailure, S as coder, T as fixNits, Tt as testFileGlobs, U as designGateAnswerMessage, V as coveringPrompt, W as designSystem, X as implementBuildingPrompt, Y as idleMessage, Z as implementFixPrompt, _ as REQUIREMENTS, _t as slicingPrompt, a as flagModeOn, at as refactorDroppedRefusal, b as answerReviewQuestions, bt as unwindFailedMessage, ct as refactorRules, d as healthy, et as normaliseCriteriaScript, f as FEEDBACK, g as REFACTOR_NOTE, gt as seamSteerMessage, h as REFACTOR, ht as seamByHandMessage, i as skills, it as refactorBuildingPrompt, l as gate, mt as reviewerSystem, n as summary, ot as refactorFixPrompt, p as HANDOFF, r as access, s as shipped, st as refactorNotePrompt, t as steering, tt as oversizedMessage, vt as startGateBlockedMessage, wt as envDefaults, xt as unwindFailure, y as SLICES, yt as uncoveredMessage, z as coverageMessage } from "../chunks/reserved-DF1HD6jL.mjs";
+import { a as refusePendingHandOff, c as joinReports, o as tail } from "../chunks/tail-D89fU_wf.mjs";
+import { i as reUnwind, n as humanReview, r as isCode$1, t as buildTail } from "../chunks/review-Bv40kZu9.mjs";
 import { posix } from "node:path";
 //#region src/workflows/shared/fences.ts
 const markFenced = (text) => {
@@ -1103,17 +1103,20 @@ const buildFeature = async (firstBase) => {
 	const handedOut = new Set(listedIds());
 	const one = slices.length === 1;
 	const built = (i) => sliceIds(sliceCriteria(slices, i))?.length !== 0;
+	const reports = [];
 	for (const i of slices.keys()) {
 		if (!built(i)) continue;
 		const n = one ? "" : `-${i + 1}`;
 		await buildSlice(i === 0 ? firstBase : head(), {
+			number: `${i + 1}`,
 			tier,
 			criteria: sliceCriteria(slices, i),
 			refactor: `refactor${n}`,
 			feature: `feature${n}`,
 			more: () => slices.some((_, j) => j > i && built(j)),
 			reviewBase: one ? void 0 : firstBase,
-			handedOut
+			handedOut,
+			laps: reports
 		});
 	}
 };
@@ -1145,38 +1148,39 @@ const buildSlice = async (sliceBase, build) => {
 		const base = seam.refactored === true && !changesSince(seam.base).some((c) => isCode$1(c.path) && !carried.includes(c.path)) ? sliceBase : seam.base;
 		if (more) {
 			const guard = () => requireCriteriaFrozen(REQUIREMENTS);
-			await tail(settleQuality, {
+			const report = await tail(async (_, report) => report, {
 				green: true,
 				lensBase: base,
 				guard,
 				guardRests: false
 			});
+			build.laps.push({
+				slice: build.number,
+				report
+			});
 			return;
 		}
 		const reviewBase = build.reviewBase ?? base;
-		await featureFeedback(await reviewedTail(reviewBase, base), base, {
+		const outcome = await reviewedTail(reviewBase, base, {
+			slice: build.number,
+			laps: build.laps
+		});
+		await featureFeedback(outcome, base, {
 			closed,
 			used
 		});
 	});
 };
-const settleQuality = async () => {
-	let rested = false;
-	while ((read(".gtd/QUALITY.md") ?? "").length > 0) {
-		rested = true;
-		await human("quality-unresolved", {
-			message: qualityUnresolvedMessage(),
-			label: "Quality findings left unresolved",
-			file: QUALITY
-		});
-	}
-	if (rested) await healthy(() => fixCheck());
-};
-/** The triaged quality lap, then the human review since `base`. */
-const reviewedTail = (base, lensBase) => tail((escalations, report) => humanReview(base, escalations, { report }), {
+/** The quality lap of slice `last.slice`, then the human review since `base`, listing what earlier slices' `laps` left too. */
+const reviewedTail = (base, lensBase, last = {
+	slice: "",
+	laps: []
+}) => tail((escalations, report) => humanReview(base, escalations, { report: last.laps.length === 0 ? report : joinReports([...last.laps, {
+	slice: last.slice,
+	report
+}]) }), {
 	green: true,
-	lensBase,
-	triage: true
+	lensBase
 });
 const featureFeedback = async (first, lensBase, ids) => {
 	let outcome = first;

package/dist/workflows/fix.mjs

@@ -1,7 +1,7 @@
 import { read, refuse, removeScript, run, scope, start } from "../flows.mjs";
 import { g as reviewRisks, h as reviewNotes } from "../chunks/steering-D7hWeMgE.mjs";
-import { C as coder, J as fixReviewNotesPrompt, N as agentWithSkills, Tt as envDefaults, c as baseline, i as skills, n as summary, q as finisherSystem, r as access, s as shipped, t as steering, wt as defaults, y as REVIEW } from "../chunks/reserved-BENPy6IJ.mjs";
-import { r as handOffTail, t as REVIEW_ORIG } from "../chunks/tail-CDNMZ-2E.mjs";
+import { Ct as defaults, K as finisherSystem, M as agentWithSkills, S as coder, c as baseline, i as skills, n as summary, q as fixReviewNotesPrompt, r as access, s as shipped, t as steering, v as REVIEW, wt as envDefaults } from "../chunks/reserved-DF1HD6jL.mjs";
+import { r as handOffTail, t as REVIEW_ORIG } from "../chunks/tail-D89fU_wf.mjs";
 //#region src/workflows/fix/fix.ts
 /**
 * Fix every note a human left in a handed-off `.gtd/REVIEW.md` and every

package/dist/workflows/review.mjs

@@ -1,7 +1,7 @@
 import { read, refuse, removeScript, run, start } from "../flows.mjs";
-import { C as coder, N as agentWithSkills, Tt as envDefaults, i as skills, mt as reviewGateBlockedMessage, n as summary, pt as reviewFeedbackFixPrompt, q as finisherSystem, r as access, s as shipped, t as steering, u as gate, v as REQUIREMENTS, wt as defaults } from "../chunks/reserved-BENPy6IJ.mjs";
-import { a as refusePendingHandOff } from "../chunks/tail-CDNMZ-2E.mjs";
-import { i as reUnwind, t as buildTail } from "../chunks/review-DcbgUwr7.mjs";
+import { Ct as defaults, K as finisherSystem, M as agentWithSkills, S as coder, _ as REQUIREMENTS, dt as reviewFeedbackFixPrompt, ft as reviewGateBlockedMessage, i as skills, l as gate, n as summary, r as access, s as shipped, t as steering, wt as envDefaults } from "../chunks/reserved-DF1HD6jL.mjs";
+import { a as refusePendingHandOff } from "../chunks/tail-D89fU_wf.mjs";
+import { i as reUnwind, t as buildTail } from "../chunks/review-Bv40kZu9.mjs";
 //#region src/workflows/review/review.ts
 const fixFeedback = async () => {
 	await agentWithSkills("feedback.fixing", reviewFeedbackFixPrompt(), {

package/dist/workflows/rollout.mjs

@@ -1,5 +1,5 @@
 import { agent, changesSince, glob, head, mark, read, refuse, removeScript, run, scope, vars } from "../flows.mjs";
-import { Bt as requireFlag, C as coder, It as flagMentions, Rt as flipProblems, Tt as envDefaults, i as skills, n as summary, o as flagShipped, p as FEEDBACK, r as access, t as steering, u as gate, wt as defaults } from "../chunks/reserved-BENPy6IJ.mjs";
+import { Ct as defaults, Ft as flagMentions, Lt as flipProblems, S as coder, f as FEEDBACK, i as skills, l as gate, n as summary, o as flagShipped, r as access, t as steering, wt as envDefaults, zt as requireFlag } from "../chunks/reserved-DF1HD6jL.mjs";
 //#region src/workflows/rollout/rollout.ts
 const flippingPrompt = (name, registry) => `Roll out the release flag "${name}": in ${registry}, set \`defaultOn: true\` on its entry. Change nothing else — no other entry, no other field, no other file, no comments.`;
 const suiteBlockedMessage = () => "The suite is red after the rollout. Fix it by hand — a rollout may only touch the flag's registry entry — then land.";