@pmelab/gtd 3.4.0 → 3.6.0

package/bin/gtd-loop

@@ -7,9 +7,10 @@ set -euo pipefail
 #   - "script"  (a check rest) -> `gtd run` executes it and steps the check
 #     actor in one command; zero new commits = green with nothing owed (the
 #     terminal signal at idle).
-#   - "prompt"  (an agent rest) -> feed `content` to the agent, then
-#     `gtd step <actor>` ourselves — the harness owns ending the turn, not
-#     the prompt text.
+#   - "prompt"  (an agent rest) -> feed `content` to the agent, run the
+#     self-validation gate (`gtd validate`, re-prompting on findings — see the
+#     gate's own comment below), then `gtd step <actor>` ourselves — the
+#     harness owns ending the turn, not the prompt text.
 # `.model` (optional, opaque) is exported as $GTD_LOOP_MODEL for the agent
 # adapter invoked at a "prompt" rest; the default adapter appends `--model
 # "$GTD_LOOP_MODEL"` only when it's non-empty.
@@ -34,6 +35,34 @@ prev_content=""
 # across iterations and across restarts. In the git dir, never the work tree.
 memory_marker="$(git rev-parse --git-dir 2>/dev/null || echo .git)/gtd-loop-memory"
 
+# Swappable agent adapter: GTD_LOOP_AGENT_CMD lets any coding agent CLI stand
+# in for the default, receiving the prompt via $GTD_LOOP_PROMPT, the state's
+# opaque model hint (if any) via $GTD_LOOP_MODEL, and the memory scope via
+# $GTD_LOOP_MEMORY / $GTD_LOOP_SESSION_ID / $GTD_LOOP_MEMORY_RESUME ("1" when
+# this turn should continue the prior session). An adapter that ignores any of
+# these keeps working — they're exported alongside $GTD_LOOP_PROMPT, never
+# required. The default adapter maps the memory signal onto claude's own
+# session flags: --resume to continue the scope's session, --session-id to pin
+# a fresh one so the next same-scope turn can resume it. $1 is the prompt, $2
+# the resume flag; the memory scope ($memory/$session_id) is read from the
+# enclosing iteration, so a same-turn re-prompt (the validate gate) resumes the
+# same session.
+run_agent_turn() {
+  local prompt="$1" do_resume="$2" cmd="${GTD_LOOP_AGENT_CMD:-}"
+  if [[ -z "$cmd" ]]; then
+    cmd='claude -p "$GTD_LOOP_PROMPT"'
+    [[ -n "$model" ]] && cmd="$cmd"' --model "$GTD_LOOP_MODEL"'
+    if [[ "$do_resume" == "1" ]]; then
+      cmd="$cmd"' --resume "$GTD_LOOP_SESSION_ID"'
+    elif [[ -n "$session_id" ]]; then
+      cmd="$cmd"' --session-id "$GTD_LOOP_SESSION_ID"'
+    fi
+    cmd="$cmd"' --dangerously-skip-permissions'
+  fi
+  GTD_LOOP_PROMPT="$prompt" GTD_LOOP_MODEL="$model" GTD_LOOP_MEMORY="$memory" \
+    GTD_LOOP_SESSION_ID="$session_id" GTD_LOOP_MEMORY_RESUME="$do_resume" bash -c "$cmd"
+}
+
 while true; do
   if ! next_json="$(gtd next --json 2>&1)"; then
     echo "gtd-loop: could not determine the next step:" >&2
@@ -104,28 +133,32 @@ while true; do
     rm -f "$memory_marker"
   fi
 
-  # Swappable agent adapter: GTD_LOOP_AGENT_CMD lets any coding agent CLI
-  # stand in for the default, receiving the prompt via $GTD_LOOP_PROMPT, the
-  # state's opaque model hint (if any) via $GTD_LOOP_MODEL, and the memory
-  # scope via $GTD_LOOP_MEMORY / $GTD_LOOP_SESSION_ID / $GTD_LOOP_MEMORY_RESUME
-  # ("1" when this turn should continue the prior session). An adapter that
-  # ignores any of these keeps working — they're exported alongside
-  # $GTD_LOOP_PROMPT, never required. The default adapter maps the memory
-  # signal onto claude's own session flags: --resume to continue the scope's
-  # session, --session-id to pin a fresh one so the next same-scope turn can
-  # resume it.
-  cmd="${GTD_LOOP_AGENT_CMD:-}"
-  if [[ -z "$cmd" ]]; then
-    cmd='claude -p "$GTD_LOOP_PROMPT"'
-    [[ -n "$model" ]] && cmd="$cmd"' --model "$GTD_LOOP_MODEL"'
-    if [[ "$resume" == "1" ]]; then
-      cmd="$cmd"' --resume "$GTD_LOOP_SESSION_ID"'
-    elif [[ -n "$session_id" ]]; then
-      cmd="$cmd"' --session-id "$GTD_LOOP_SESSION_ID"'
+  # Run the agent turn (see run_agent_turn's own comment for the adapter
+  # contract). Then the self-validation gate: a producing state may declare a
+  # `file:`/`mode:` pair (e.g. the default workflow's grilling/reviewing),
+  # making its output's format checkable. `gtd validate` is a no-op (exit 0)
+  # when the resolved rest has nothing to validate, so run it after EVERY agent
+  # turn; on findings, re-prompt the SAME agent session to fix them (up to a
+  # cap) before stepping — the plain-`gtd next` self-validation instruction's
+  # driver-side equivalent, so the next rest is only ever handed a well-formed
+  # file. gtd itself never validates at step time — this is the loop's job.
+  run_agent_turn "$content" "$resume"
+
+  fix_attempt=0
+  while ! validate_out="$(gtd validate 2>&1)"; do
+    fix_attempt=$((fix_attempt + 1))
+    if ((fix_attempt > 3)); then
+      echo "gtd-loop: '$state' still fails \`gtd validate\` after 3 fix attempts:" >&2
+      echo "$validate_out" >&2
+      echo "Stopping rather than stepping with a malformed steering file." >&2
+      exit 1
     fi
-    cmd="$cmd"' --dangerously-skip-permissions'
-  fi
-  GTD_LOOP_PROMPT="$content" GTD_LOOP_MODEL="$model" GTD_LOOP_MEMORY="$memory" \
-    GTD_LOOP_SESSION_ID="$session_id" GTD_LOOP_MEMORY_RESUME="$resume" bash -c "$cmd"
+    # Resume the same session when there is one (the agent just wrote the file);
+    # otherwise a fresh invocation still gets the findings via the prompt.
+    fix_resume=0
+    [[ -n "$session_id" ]] && fix_resume=1
+    run_agent_turn "$content"$'\n\n'"Your last turn does not pass \`gtd validate\`. Fix these format violations, then finish:"$'\n'"$validate_out" "$fix_resume"
+  done
+
   gtd step "$actor" >/dev/null
 done