@pmelab/gtd 47.0.0 → 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/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@pmelab/gtd",
-  "version": "47.0.0",
+  "version": "48.0.0",
   "private": false,
   "description": "Git-aware CLI that emits the next prompt for an autonomous coding agent based on the current repository state",
   "bin": {

package/src/cli/Cli.test.ts

@@ -1147,7 +1147,6 @@ describe("standaloneKinds / needsOf", () => {
       "check",
       "uncheck",
       "install",
-      "loop",
       "contextHook",
       "judgeRun",
       "strays",
@@ -1159,7 +1158,6 @@ describe("standaloneKinds / needsOf", () => {
     expect(needsOf("check")).toBe("fs")
     expect(needsOf("uncheck")).toBe("fs")
     expect(needsOf("install")).toBe("none")
-    expect(needsOf("loop")).toBe("none")
     expect(needsOf("contextHook")).toBe("none")
     expect(needsOf("judgeRun")).toBe("none")
     expect(needsOf("strays")).toBe("none")

package/src/cli/Cli.ts

@@ -642,20 +642,6 @@ const COMMAND_ROWS: readonly CommandRow[] = [
       "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",
@@ -1528,7 +1514,6 @@ export const parseArgv = (argv: readonly string[]): CliPlan => {
       | "next"
       | "validate"
       | "install"
-      | "loop"
       | "doors"
       | "slices"
       | "base"

package/src/cli/Program.test.ts

@@ -1685,7 +1685,6 @@ describe("runCommand — refuses in a repository with no commits", () => {
     check: { kind: "check", mode: "qa", file: ".gtd/TODO.md" },
     uncheck: { kind: "uncheck", file: ".gtd/REVIEW.md" },
     install: { kind: "install" },
-    loop: { kind: "loop" },
     contextHook: { kind: "contextHook" },
     summary: { kind: "summary" },
     base: { kind: "base" },

package/src/cli/Program.ts

@@ -163,7 +163,6 @@ export type Command =
     }
   | { readonly kind: "uncheck"; readonly file: string }
   | { readonly kind: "install" }
-  | { readonly kind: "loop" }
   | {
       readonly kind: "contextHook"
       readonly soft?: number
@@ -239,9 +238,6 @@ const runInstallCommand = (
     )
   })
 
-const runLoopCommand = (): Effect.Effect<void, Error, ToolCommands> =>
-  Effect.flatMap(ToolCommands, (tools) => tools.loop(selfGtd()))
-
 /**
  * `planLanding`'s result — a preview of what `gtd land` would do, plus the
  * combined `script` the driver runs to do it. `script` is
@@ -1221,7 +1217,6 @@ export const needsOf = (kind: Command["kind"]): Needs => {
   switch (kind) {
     case "lsp":
     case "install":
-    case "loop":
     case "contextHook":
     case "judgeRun":
     case "strays":
@@ -1240,7 +1235,6 @@ export const standaloneKinds = (): readonly Command["kind"][] => [
   "check",
   "uncheck",
   "install",
-  "loop",
   "contextHook",
   "judgeRun",
   "strays",
@@ -1311,8 +1305,6 @@ const dispatchVoidCommand = (
       return runUncheckCommand(command.file)
     case "install":
       return runInstallCommand(out)
-    case "loop":
-      return runLoopCommand()
     case "contextHook":
       return runContextHookCommand(command, out)
     case "summary":

package/src/cli/ToolCommands.ts

@@ -3,7 +3,7 @@ import type { Effect } from "effect"
 import type { UiSettings } from "../flows/index.js"
 import type { ContextHookRequest } from "../wire/index.js"
 
-/** How to run THIS gtd — a driver (the loop, the LSP) never reads a different version's state than the one it ships with. */
+/** How to run THIS gtd — a tool (the LSP) never reads a different version's state than the one it ships with. */
 export type SelfGtd = readonly [string, ...string[]]
 
 export interface UiRequest {
@@ -38,7 +38,6 @@ export class ToolCommands extends Context.Tag("ToolCommands")<
       out: { readonly write: (chunk: string) => void; readonly flush: () => void },
     ) => Effect.Effect<void, Error>
     readonly lsp: (gtd: SelfGtd) => Effect.Effect<void, Error>
-    readonly loop: (gtd: SelfGtd) => Effect.Effect<void, Error>
     /** The hook's stdout: a JSON document, or empty to let the turn go on. */
     readonly contextHook: (request: ContextHookRequest) => Effect.Effect<string, Error>
     readonly judge: (request: JudgeRequest) => Promise<string>

package/src/install/Install.ts

@@ -11,8 +11,15 @@ 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 "$(git rev-parse --show-toplevel)"
+
 beat=1
 while :; do
+  cost=
+  model=
+  judged_before="\${judged:-}"
+  judged=
   # One value per invocation: \`gtd next --json=<path>\` is a pure read (peek
   # or dispatch, indistinguishable, always safe to call again), so nothing is
   # lost by reading \`kind\`/\`idle\` as two separate calls rather than one
@@ -40,14 +47,22 @@ while :; do
     # \`gtd judge run\` and land the verdict via \`gtd judge answer\`. POSIX sh has
     # no pipefail, so each step is captured first and only piped on when it
     # succeeded; if \`gtd judge run\` fails or \`gtd judge answer\` rejects the
-    # verdict, its stderr reaches you and we fall through to the human hand-off.
+    # verdict, its stderr reaches you and the gate is handed to you — on the
+    # opening beat too, so it never lands unanswered — as it is when the gate
+    # we just answered comes straight back.
     message)
-      if [ -n "$(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
@@ -66,45 +81,83 @@ 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\`:
@@ -114,7 +167,8 @@ 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

package/src/testkit/Layers.ts

@@ -205,7 +205,6 @@ const missingTool = (name: string) => () =>
 const missingToolCommands: ToolCommands["Type"] = {
   ui: missingTool("ui"),
   lsp: missingTool("lsp"),
-  loop: missingTool("loop"),
   contextHook: missingTool("context-hook"),
   judge: () =>
     Promise.reject(

package/src/wire/ContextLimit.ts

@@ -1,4 +1,4 @@
-// The context-limit decision every driver shares: the loop's `gtd context-hook`
+// The context-limit decision every driver shares: `gtd context-hook`
 // and the Claude Code mod both answer a hook event through `contextHook`, so
 // the two cannot drift on thresholds, wording or which tokens count.