fix: Migrate planner, predictor, supervisor to SDK (#6)

Replace tmux-based run_formula_and_monitor() with synchronous agent_run()
from lib/agent-sdk.sh, matching the pattern established in gardener-run.sh.

Key changes per agent:
- Drop agent-session.sh, use agent-sdk.sh (SID_FILE, LOGFILE)
- Remove SESSION_NAME, PHASE_FILE, PHASE_POLL_INTERVAL (tmux/phase artifacts)
- Strip phase protocol from prompt footer (SDK mode needs no phase signals)
- Preserve all prompt composition: context blocks, memory, journal, preflight

Shared helpers added to lib/formula-session.sh:
- build_sdk_prompt_footer(): build_prompt_footer minus phase protocol
- formula_worktree_setup(): fetch + cleanup + create worktree + EXIT trap

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

gardener/gardener-run.sh

@@ -92,11 +92,8 @@ Supported actions:
 The commit-and-pr step converts JSONL to JSON array. The orchestrator executes
 actions after the PR merges. Do NOT call mutation APIs directly during the run."
 
-# Reuse shared footer (API reference + environment), replace phase protocol
+build_sdk_prompt_footer "$GARDENER_API_EXTRA"
-# shellcheck disable=SC2034  # consumed by build_prompt_footer
+PROMPT_FOOTER="${PROMPT_FOOTER}## Completion protocol (REQUIRED)
-PHASE_FILE=""  # not used in SDK mode
-build_prompt_footer "$GARDENER_API_EXTRA"
-PROMPT_FOOTER="${PROMPT_FOOTER%%## Phase protocol*}## Completion protocol (REQUIRED)
 When the commit-and-pr step creates a PR, write the PR number and stop:
   echo \"\$PR_NUMBER\" > '${GARDENER_PR_FILE}'
 Then STOP. Do NOT write PHASE: signals — the orchestrator handles CI, review, and merge.

lib/formula-session.sh

@@ -291,6 +291,33 @@ build_graph_section() {
   fi
 }
 
+# ── SDK helpers ───────────────────────────────────────────────────────────
+
+# build_sdk_prompt_footer [EXTRA_API_LINES]
+# Like build_prompt_footer but omits the phase protocol section (SDK mode).
+# Sets PROMPT_FOOTER.
+build_sdk_prompt_footer() {
+  # shellcheck disable=SC2034  # consumed by build_prompt_footer
+  PHASE_FILE=""  # not used in SDK mode
+  build_prompt_footer "${1:-}"
+  PROMPT_FOOTER="${PROMPT_FOOTER%%## Phase protocol*}"
+}
+
+# formula_worktree_setup WORKTREE
+# Creates an isolated worktree for synchronous formula execution.
+# Fetches primary branch, cleans stale worktree, creates new one, and
+# sets an EXIT trap for cleanup.
+# Requires globals: PROJECT_REPO_ROOT, PRIMARY_BRANCH.
+formula_worktree_setup() {
+  local worktree="$1"
+  cd "$PROJECT_REPO_ROOT"
+  git fetch origin "$PRIMARY_BRANCH" 2>/dev/null || true
+  worktree_cleanup "$worktree"
+  git worktree add "$worktree" "origin/${PRIMARY_BRANCH}" --detach 2>/dev/null
+  # shellcheck disable=SC2064  # expand worktree now, not at trap time
+  trap "worktree_cleanup '$worktree'" EXIT
+}
+
 # ── Prompt + monitor helpers ──────────────────────────────────────────────
 
 # build_prompt_footer [EXTRA_API_LINES]

planner/planner-run.sh

@@ -1,10 +1,16 @@
 #!/usr/bin/env bash
 # =============================================================================
-# planner-run.sh — Cron wrapper: direct planner execution via Claude + formula
+# planner-run.sh — Cron wrapper: planner execution via SDK + formula
 #
-# Runs daily (or on-demand). Guards against concurrent runs and low memory.
+# Synchronous bash loop using claude -p (one-shot invocation).
-# Creates a tmux session with Claude (opus) reading formulas/run-planner.toml.
+# No tmux sessions, no phase files — the bash script IS the state machine.
-# No action issues — the planner is a nervous system component, not work.
+#
+# Flow:
+#   1. Guards: cron lock, memory check
+#   2. Load formula (formulas/run-planner.toml)
+#   3. Context: VISION.md, AGENTS.md, ops:RESOURCES.md, structural graph,
+#      planner memory, journal entries
+#   4. agent_run(worktree, prompt) → Claude plans, may push knowledge updates
 #
 # Usage:
 #   planner-run.sh [projects/disinto.toml]   # project config (default: disinto)
@@ -20,24 +26,22 @@ export PROJECT_TOML="${1:-$FACTORY_ROOT/projects/disinto.toml}"
 source "$FACTORY_ROOT/lib/env.sh"
 # Use planner-bot's own Forgejo identity (#747)
 FORGE_TOKEN="${FORGE_PLANNER_TOKEN:-${FORGE_TOKEN}}"
-# shellcheck source=../lib/agent-session.sh
-source "$FACTORY_ROOT/lib/agent-session.sh"
 # shellcheck source=../lib/formula-session.sh
 source "$FACTORY_ROOT/lib/formula-session.sh"
 # shellcheck source=../lib/worktree.sh
 source "$FACTORY_ROOT/lib/worktree.sh"
 # shellcheck source=../lib/guard.sh
 source "$FACTORY_ROOT/lib/guard.sh"
+# shellcheck source=../lib/agent-sdk.sh
+source "$FACTORY_ROOT/lib/agent-sdk.sh"
 
 LOG_FILE="$SCRIPT_DIR/planner.log"
-# shellcheck disable=SC2034  # consumed by run_formula_and_monitor
+# shellcheck disable=SC2034  # consumed by agent-sdk.sh
-SESSION_NAME="planner-${PROJECT_NAME}"
+LOGFILE="$LOG_FILE"
-PHASE_FILE="/tmp/planner-session-${PROJECT_NAME}.phase"
+# shellcheck disable=SC2034  # consumed by agent-sdk.sh
-
+SID_FILE="/tmp/planner-session-${PROJECT_NAME}.sid"
-# shellcheck disable=SC2034  # read by monitor_phase_loop in lib/agent-session.sh
-PHASE_POLL_INTERVAL=15
-
 SCRATCH_FILE="/tmp/planner-${PROJECT_NAME}-scratch.md"
+WORKTREE="/tmp/${PROJECT_NAME}-planner-run"
 
 log() { echo "[$(date -u +%Y-%m-%dT%H:%M:%S)Z] $*" >> "$LOG_FILE"; }
 
@@ -92,14 +96,13 @@ SCRATCH_CONTEXT=$(read_scratch_context "$SCRATCH_FILE")
 SCRATCH_INSTRUCTION=$(build_scratch_instruction "$SCRATCH_FILE")
 
 # ── Build prompt ─────────────────────────────────────────────────────────
-build_prompt_footer "
+build_sdk_prompt_footer "
   Relabel:     curl -sf -H \"Authorization: token \${FORGE_TOKEN}\" -X PUT -H 'Content-Type: application/json' '${FORGE_API}/issues/{number}/labels' -d '{\"labels\":[LABEL_ID]}'
   Comment:     curl -sf -H \"Authorization: token \${FORGE_TOKEN}\" -X POST -H 'Content-Type: application/json' '${FORGE_API}/issues/{number}/comments' -d '{\"body\":\"...\"}'
   Close:       curl -sf -H \"Authorization: token \${FORGE_TOKEN}\" -X PATCH -H 'Content-Type: application/json' '${FORGE_API}/issues/{number}' -d '{\"state\":\"closed\"}'
 "
 
-# shellcheck disable=SC2034  # consumed by run_formula_and_monitor
+PROMPT="You are the strategic planner for ${FORGE_REPO}. Work through the formula below.
-PROMPT="You are the strategic planner for ${FORGE_REPO}. Work through the formula below. You MUST write PHASE:done to '${PHASE_FILE}' when finished — the orchestrator will time you out if you return to the prompt without signalling.
 
 ## Project context
 ${CONTEXT_BLOCK}${MEMORY_BLOCK}${JOURNAL_BLOCK}
@@ -113,12 +116,14 @@ ${SCRATCH_INSTRUCTION}
 
 ${PROMPT_FOOTER}"
 
-# ── Run session ──────────────────────────────────────────────────────────
+# ── Create worktree ──────────────────────────────────────────────────────
-export CLAUDE_MODEL="opus"
+formula_worktree_setup "$WORKTREE"
-run_formula_and_monitor "planner"
 
-# ── Cleanup scratch file on normal exit ──────────────────────────────────
+# ── Run agent ─────────────────────────────────────────────────────────────
-# FINAL_PHASE already set by run_formula_and_monitor
+export CLAUDE_MODEL="opus"
-if [ "${FINAL_PHASE:-}" = "PHASE:done" ]; then
+
-  rm -f "$SCRATCH_FILE"
+agent_run --worktree "$WORKTREE" "$PROMPT"
-fi
+log "agent_run complete"
+
+rm -f "$SCRATCH_FILE"
+log "--- Planner run done ---"

predictor/predictor-run.sh

@@ -1,10 +1,15 @@
 #!/usr/bin/env bash
 # =============================================================================
-# predictor-run.sh — Cron wrapper: predictor execution via Claude + formula
+# predictor-run.sh — Cron wrapper: predictor execution via SDK + formula
 #
-# Runs daily (or on-demand). Guards against concurrent runs and low memory.
+# Synchronous bash loop using claude -p (one-shot invocation).
-# Creates a tmux session with Claude (sonnet) reading formulas/run-predictor.toml.
+# No tmux sessions, no phase files — the bash script IS the state machine.
-# Files prediction/unreviewed issues for the planner to triage.
+#
+# Flow:
+#   1. Guards: cron lock, memory check
+#   2. Load formula (formulas/run-predictor.toml)
+#   3. Context: AGENTS.md, ops:RESOURCES.md, VISION.md, structural graph
+#   4. agent_run(worktree, prompt) → Claude analyzes, writes to ops repo
 #
 # Usage:
 #   predictor-run.sh [projects/disinto.toml]   # project config (default: disinto)
@@ -22,24 +27,22 @@ export PROJECT_TOML="${1:-$FACTORY_ROOT/projects/disinto.toml}"
 source "$FACTORY_ROOT/lib/env.sh"
 # Use predictor-bot's own Forgejo identity (#747)
 FORGE_TOKEN="${FORGE_PREDICTOR_TOKEN:-${FORGE_TOKEN}}"
-# shellcheck source=../lib/agent-session.sh
-source "$FACTORY_ROOT/lib/agent-session.sh"
 # shellcheck source=../lib/formula-session.sh
 source "$FACTORY_ROOT/lib/formula-session.sh"
 # shellcheck source=../lib/worktree.sh
 source "$FACTORY_ROOT/lib/worktree.sh"
 # shellcheck source=../lib/guard.sh
 source "$FACTORY_ROOT/lib/guard.sh"
+# shellcheck source=../lib/agent-sdk.sh
+source "$FACTORY_ROOT/lib/agent-sdk.sh"
 
 LOG_FILE="$SCRIPT_DIR/predictor.log"
-# shellcheck disable=SC2034  # consumed by run_formula_and_monitor
+# shellcheck disable=SC2034  # consumed by agent-sdk.sh
-SESSION_NAME="predictor-${PROJECT_NAME}"
+LOGFILE="$LOG_FILE"
-PHASE_FILE="/tmp/predictor-session-${PROJECT_NAME}.phase"
+# shellcheck disable=SC2034  # consumed by agent-sdk.sh
-
+SID_FILE="/tmp/predictor-session-${PROJECT_NAME}.sid"
-# shellcheck disable=SC2034  # read by monitor_phase_loop in lib/agent-session.sh
-PHASE_POLL_INTERVAL=15
-
 SCRATCH_FILE="/tmp/predictor-${PROJECT_NAME}-scratch.md"
+WORKTREE="/tmp/${PROJECT_NAME}-predictor-run"
 
 log() { echo "[$(date -u +%Y-%m-%dT%H:%M:%S)Z] $*" >> "$LOG_FILE"; }
 
@@ -62,10 +65,10 @@ SCRATCH_CONTEXT=$(read_scratch_context "$SCRATCH_FILE")
 SCRATCH_INSTRUCTION=$(build_scratch_instruction "$SCRATCH_FILE")
 
 # ── Build prompt ─────────────────────────────────────────────────────────
-build_prompt_footer
+build_sdk_prompt_footer
+export CLAUDE_MODEL="sonnet"
 
-# shellcheck disable=SC2034  # consumed by run_formula_and_monitor
+PROMPT="You are the prediction agent (goblin) for ${FORGE_REPO}. Work through the formula below.
-PROMPT="You are the prediction agent (goblin) for ${FORGE_REPO}. Work through the formula below. You MUST write PHASE:done to '${PHASE_FILE}' when finished — the orchestrator will time you out if you return to the prompt without signalling.
 
 Your role: abstract adversary. Find the project's biggest weakness, challenge
 planner claims, and generate evidence. Explore when uncertain (file a prediction),
@@ -88,12 +91,12 @@ ${FORMULA_CONTENT}
 ${SCRATCH_INSTRUCTION}
 ${PROMPT_FOOTER}"
 
-# ── Run session ──────────────────────────────────────────────────────────
+# ── Create worktree ──────────────────────────────────────────────────────
-export CLAUDE_MODEL="sonnet"
+formula_worktree_setup "$WORKTREE"
-run_formula_and_monitor "predictor"
 
-# ── Cleanup scratch file on normal exit ──────────────────────────────────
+# ── Run agent ─────────────────────────────────────────────────────────────
-# FINAL_PHASE already set by run_formula_and_monitor
+agent_run --worktree "$WORKTREE" "$PROMPT"
-if [ "${FINAL_PHASE:-}" = "PHASE:done" ]; then
+log "agent_run complete"
-  rm -f "$SCRATCH_FILE"
+
-fi
+rm -f "$SCRATCH_FILE"
+log "--- Predictor run done ---"

supervisor/supervisor-run.sh

@@ -1,14 +1,17 @@
 #!/usr/bin/env bash
 # =============================================================================
-# supervisor-run.sh — Cron wrapper: supervisor execution via Claude + formula
+# supervisor-run.sh — Cron wrapper: supervisor execution via SDK + formula
 #
-# Runs every 20 minutes (or on-demand). Guards against concurrent runs and
+# Synchronous bash loop using claude -p (one-shot invocation).
-# low memory. Collects metrics via preflight.sh, then creates a tmux session
+# No tmux sessions, no phase files — the bash script IS the state machine.
-# with Claude (sonnet) reading formulas/run-supervisor.toml.
 #
-# Replaces supervisor-poll.sh (bash orchestrator + claude -p one-shot) with
+# Flow:
-# formula-driven interactive Claude session matching the planner/predictor
+#   1. Guards: cron lock, memory check
-# pattern.
+#   2. Housekeeping: clean up stale crashed worktrees
+#   3. Collect pre-flight metrics (supervisor/preflight.sh)
+#   4. Load formula (formulas/run-supervisor.toml)
+#   5. Context: AGENTS.md, preflight metrics, structural graph
+#   6. agent_run(worktree, prompt) → Claude monitors, may clean up
 #
 # Usage:
 #   supervisor-run.sh [projects/disinto.toml]   # project config (default: disinto)
@@ -26,24 +29,22 @@ export PROJECT_TOML="${1:-$FACTORY_ROOT/projects/disinto.toml}"
 source "$FACTORY_ROOT/lib/env.sh"
 # Use supervisor-bot's own Forgejo identity (#747)
 FORGE_TOKEN="${FORGE_SUPERVISOR_TOKEN:-${FORGE_TOKEN}}"
-# shellcheck source=../lib/agent-session.sh
-source "$FACTORY_ROOT/lib/agent-session.sh"
 # shellcheck source=../lib/formula-session.sh
 source "$FACTORY_ROOT/lib/formula-session.sh"
 # shellcheck source=../lib/worktree.sh
 source "$FACTORY_ROOT/lib/worktree.sh"
 # shellcheck source=../lib/guard.sh
 source "$FACTORY_ROOT/lib/guard.sh"
+# shellcheck source=../lib/agent-sdk.sh
+source "$FACTORY_ROOT/lib/agent-sdk.sh"
 
 LOG_FILE="$SCRIPT_DIR/supervisor.log"
-# shellcheck disable=SC2034  # consumed by run_formula_and_monitor
+# shellcheck disable=SC2034  # consumed by agent-sdk.sh
-SESSION_NAME="supervisor-${PROJECT_NAME}"
+LOGFILE="$LOG_FILE"
-PHASE_FILE="/tmp/supervisor-session-${PROJECT_NAME}.phase"
+# shellcheck disable=SC2034  # consumed by agent-sdk.sh
-
+SID_FILE="/tmp/supervisor-session-${PROJECT_NAME}.sid"
-# shellcheck disable=SC2034  # read by monitor_phase_loop in lib/agent-session.sh
-PHASE_POLL_INTERVAL=15
-
 SCRATCH_FILE="/tmp/supervisor-${PROJECT_NAME}-scratch.md"
+WORKTREE="/tmp/${PROJECT_NAME}-supervisor-run"
 
 log() { echo "[$(date -u +%Y-%m-%dT%H:%M:%S)Z] $*" >> "$LOG_FILE"; }
 
@@ -75,10 +76,13 @@ SCRATCH_CONTEXT=$(read_scratch_context "$SCRATCH_FILE")
 SCRATCH_INSTRUCTION=$(build_scratch_instruction "$SCRATCH_FILE")
 
 # ── Build prompt ─────────────────────────────────────────────────────────
-build_prompt_footer
+build_sdk_prompt_footer
+export CLAUDE_MODEL="sonnet"
 
-# shellcheck disable=SC2034  # consumed by run_formula_and_monitor
+# ── Create worktree (before prompt assembly so trap is set early) ────────
-PROMPT="You are the supervisor agent for ${FORGE_REPO}. Work through the formula below. You MUST write PHASE:done to '${PHASE_FILE}' when finished — the orchestrator will time you out if you return to the prompt without signalling.
+formula_worktree_setup "$WORKTREE"
+
+PROMPT="You are the supervisor agent for ${FORGE_REPO}. Work through the formula below.
 
 You have full shell access and --dangerously-skip-permissions.
 Fix what you can. File vault items for what you cannot. Do NOT ask permission — act first, report after.
@@ -97,12 +101,9 @@ ${FORMULA_CONTENT}
 ${SCRATCH_INSTRUCTION}
 ${PROMPT_FOOTER}"
 
-# ── Run session ──────────────────────────────────────────────────────────
+# ── Run agent ─────────────────────────────────────────────────────────────
-export CLAUDE_MODEL="sonnet"
+agent_run --worktree "$WORKTREE" "$PROMPT"
-run_formula_and_monitor "supervisor" 1200
+log "agent_run complete"
 
-# ── Cleanup scratch file on normal exit ──────────────────────────────────
+rm -f "$SCRATCH_FILE"
-# FINAL_PHASE already set by run_formula_and_monitor
+log "--- Supervisor run done ---"
-if [ "${FINAL_PHASE:-}" = "PHASE:done" ]; then
-  rm -f "$SCRATCH_FILE"
-fi