disinto/dev/dev-agent.sh

745 lines
29 KiB
Bash
Raw Normal View History

#!/usr/bin/env bash
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# dev-agent.sh — Autonomous developer agent for a single issue (tmux session manager)
#
# Usage: ./dev-agent.sh <issue-number>
#
# Lifecycle:
# 1. Fetch issue, check dependencies (preflight)
# 2. Claim issue (label: in-progress, remove backlog)
# 3. Create worktree + branch
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# 4. Create tmux session: dev-{project}-{issue} with interactive claude
# 5. Send initial prompt via tmux (issue body, context, phase protocol)
# 6. Monitor phase file — Claude signals when it needs input
# 7. React to phases: create PR, poll CI, inject results, inject review, merge
# 8. Kill session on PHASE:done, PHASE:failed, or 2h idle timeout
#
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Phase file: /tmp/dev-session-{project}-{issue}.phase
# Session: dev-{project}-{issue} (tmux)
# Peek phase: head -1 /tmp/dev-session-{project}-{issue}.phase
# Log: tail -f dev-agent.log
set -euo pipefail
# Load shared environment
source "$(dirname "$0")/../lib/env.sh"
source "$(dirname "$0")/../lib/ci-helpers.sh"
source "$(dirname "$0")/../lib/agent-session.sh"
source "$(dirname "$0")/../lib/formula-session.sh"
source "$(dirname "$0")/../lib/worktree.sh"
# shellcheck source=./phase-handler.sh
source "$(dirname "$0")/phase-handler.sh"
# Auto-pull factory code to pick up merged fixes before any logic runs
git -C "$FACTORY_ROOT" pull --ff-only origin main 2>/dev/null || true
# --- Config ---
ISSUE="${1:?Usage: dev-agent.sh <issue-number>}"
# shellcheck disable=SC2034
REPO="${FORGE_REPO}"
# shellcheck disable=SC2034
REPO_ROOT="${PROJECT_REPO_ROOT}"
API="${FORGE_API}"
LOCKFILE="/tmp/dev-agent-${PROJECT_NAME:-default}.lock"
STATUSFILE="/tmp/dev-agent-status-${PROJECT_NAME:-default}"
# Gitea labels API requires []int64 — look up the "backlog" label ID once
BACKLOG_LABEL_ID=$(forge_api GET "/labels" 2>/dev/null \
| jq -r '.[] | select(.name == "backlog") | .id' 2>/dev/null || true)
BACKLOG_LABEL_ID="${BACKLOG_LABEL_ID:-1300815}"
# Same for "in-progress" label
IN_PROGRESS_LABEL_ID=$(forge_api GET "/labels" 2>/dev/null \
| jq -r '.[] | select(.name == "in-progress") | .id' 2>/dev/null || true)
IN_PROGRESS_LABEL_ID="${IN_PROGRESS_LABEL_ID:-1300818}"
log() {
printf '[%s] #%s %s\n' "$(date -u '+%Y-%m-%d %H:%M:%S UTC')" "$ISSUE" "$*" >> "$LOGFILE"
}
status() {
printf '[%s] dev-agent #%s: %s\n' "$(date -u '+%Y-%m-%d %H:%M:%S UTC')" "$ISSUE" "$*" > "$STATUSFILE"
log "$*"
}
LOGFILE="${DISINTO_LOG_DIR}/dev/dev-agent.log"
PREFLIGHT_RESULT="/tmp/dev-agent-preflight.json"
BRANCH="fix/issue-${ISSUE}"
WORKTREE="/tmp/${PROJECT_NAME}-worktree-${ISSUE}"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Tmux session + phase protocol
PHASE_FILE="/tmp/dev-session-${PROJECT_NAME}-${ISSUE}.phase"
SESSION_NAME="dev-${PROJECT_NAME}-${ISSUE}"
IMPL_SUMMARY_FILE="/tmp/dev-impl-summary-${PROJECT_NAME}-${ISSUE}.txt"
# Scratch file for context compaction survival
SCRATCH_FILE="/tmp/dev-${PROJECT_NAME}-${ISSUE}-scratch.md"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Timing
export PHASE_POLL_INTERVAL=30 # seconds between phase checks (read by agent-session.sh)
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
IDLE_TIMEOUT=7200 # 2h: kill session if phase stale this long
# shellcheck disable=SC2034 # used by phase-handler.sh
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
CI_POLL_TIMEOUT=1800 # 30min max for CI to complete
# shellcheck disable=SC2034 # used by phase-handler.sh
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
REVIEW_POLL_TIMEOUT=10800 # 3h max wait for review
# Limits
# shellcheck disable=SC2034 # used by phase-handler.sh
MAX_CI_FIXES=3
# shellcheck disable=SC2034 # used by phase-handler.sh
MAX_REVIEW_ROUNDS=5
# Counters — global state shared with phase-handler.sh across phase transitions
# shellcheck disable=SC2034
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
CI_RETRY_COUNT=0
# shellcheck disable=SC2034
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
CI_FIX_COUNT=0
# shellcheck disable=SC2034
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
REVIEW_ROUND=0
PR_NUMBER=""
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# --- Cleanup helpers ---
cleanup_worktree() {
worktree_cleanup "$WORKTREE"
}
cleanup_labels() {
curl -sf -X DELETE \
-H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${ISSUE}/labels/${IN_PROGRESS_LABEL_ID}" >/dev/null 2>&1 || true
}
restore_to_backlog() {
cleanup_labels
curl -sf -X POST \
-H "Authorization: token ${FORGE_TOKEN}" \
-H "Content-Type: application/json" \
"${API}/issues/${ISSUE}/labels" \
-d "{\"labels\":[${BACKLOG_LABEL_ID}]}" >/dev/null 2>&1 || true
CLAIMED=false # Don't unclaim again in cleanup()
}
CLAIMED=false
cleanup() {
rm -f "$LOCKFILE" "$STATUSFILE"
# Kill any live session so Claude doesn't run without an orchestrator attached
agent_kill_session "$SESSION_NAME"
# If we claimed the issue but never created a PR, unclaim it
if [ "$CLAIMED" = true ] && [ -z "${PR_NUMBER:-}" ]; then
log "cleanup: unclaiming issue (no PR created)"
curl -sf -X DELETE \
-H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${ISSUE}/labels/${IN_PROGRESS_LABEL_ID}" >/dev/null 2>&1 || true
curl -sf -X POST \
-H "Authorization: token ${FORGE_TOKEN}" \
-H "Content-Type: application/json" \
"${API}/issues/${ISSUE}/labels" \
-d "{\"labels\":[${BACKLOG_LABEL_ID}]}" >/dev/null 2>&1 || true
fi
}
trap cleanup EXIT
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# =============================================================================
# LOG ROTATION
# =============================================================================
if [ -f "$LOGFILE" ] && [ "$(stat -c%s "$LOGFILE" 2>/dev/null || echo 0)" -gt 102400 ]; then
mv "$LOGFILE" "$LOGFILE.old"
log "Log rotated"
fi
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# =============================================================================
# MEMORY GUARD
# =============================================================================
AVAIL_MB=$(awk '/MemAvailable/ {printf "%d", $2/1024}' /proc/meminfo)
if [ "$AVAIL_MB" -lt 2000 ]; then
log "SKIP: only ${AVAIL_MB}MB available (need 2000MB)"
exit 0
fi
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# =============================================================================
# CONCURRENCY LOCK
# =============================================================================
if [ -f "$LOCKFILE" ]; then
LOCK_PID=$(cat "$LOCKFILE" 2>/dev/null || echo "")
if [ -n "$LOCK_PID" ] && kill -0 "$LOCK_PID" 2>/dev/null; then
log "SKIP: another dev-agent running (PID ${LOCK_PID})"
exit 0
fi
log "Removing stale lock (PID ${LOCK_PID:-?})"
rm -f "$LOCKFILE"
fi
echo $$ > "$LOCKFILE"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# =============================================================================
# FETCH ISSUE
# =============================================================================
status "fetching issue"
ISSUE_JSON=$(curl -s -H "Authorization: token ${FORGE_TOKEN}" "${API}/issues/${ISSUE}") || true
if [ -z "$ISSUE_JSON" ] || ! echo "$ISSUE_JSON" | jq -e '.id' >/dev/null 2>&1; then
log "ERROR: failed to fetch issue #${ISSUE} (API down or invalid response)"
exit 1
fi
ISSUE_TITLE=$(echo "$ISSUE_JSON" | jq -r '.title')
ISSUE_BODY=$(echo "$ISSUE_JSON" | jq -r '.body // ""')
ISSUE_BODY_ORIGINAL="$ISSUE_BODY"
# --- Resolve bot username(s) for comment filtering ---
_bot_login=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API%%/repos*}/user" | jq -r '.login // empty' 2>/dev/null || true)
# Build list: token owner + any extra names from FORGE_BOT_USERNAMES (comma-separated)
_bot_logins="${_bot_login}"
if [ -n "${FORGE_BOT_USERNAMES:-}" ]; then
_bot_logins="${_bot_logins:+${_bot_logins},}${FORGE_BOT_USERNAMES}"
fi
# Append human comments to issue body (filter out bot accounts)
ISSUE_COMMENTS=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${ISSUE}/comments" | \
jq -r --arg bots "$_bot_logins" \
'($bots | split(",") | map(select(. != ""))) as $bl |
.[] | select(.user.login as $u | $bl | index($u) | not) |
"### @\(.user.login) (\(.created_at[:10])):\n\(.body)\n"' 2>/dev/null || true)
if [ -n "$ISSUE_COMMENTS" ]; then
ISSUE_BODY="${ISSUE_BODY}
## Issue comments
${ISSUE_COMMENTS}"
fi
ISSUE_STATE=$(echo "$ISSUE_JSON" | jq -r '.state')
if [ "$ISSUE_STATE" != "open" ]; then
log "SKIP: issue #${ISSUE} is ${ISSUE_STATE}"
echo '{"status":"already_done","reason":"issue is closed"}' > "$PREFLIGHT_RESULT"
exit 0
fi
log "Issue: ${ISSUE_TITLE}"
# =============================================================================
# GUARD: Reject formula-labeled issues (feat/formula not yet merged)
# =============================================================================
ISSUE_LABELS=$(echo "$ISSUE_JSON" | jq -r '[.labels[].name] | join(",")') || true
if echo "$ISSUE_LABELS" | grep -qw 'formula'; then
log "SKIP: issue #${ISSUE} has 'formula' label but formula dispatch is not yet implemented (feat/formula branch not merged)"
echo '{"status":"unmet_dependency","blocked_by":"formula dispatch not implemented — feat/formula branch not merged to main","suggestion":null}' > "$PREFLIGHT_RESULT"
exit 0
fi
# =============================================================================
# PREFLIGHT: Check dependencies before doing any work
# =============================================================================
status "preflight check"
# Extract dependency references using shared parser (use original body only — not comments)
DEP_NUMBERS=$(echo "$ISSUE_BODY_ORIGINAL" | bash "${FACTORY_ROOT}/lib/parse-deps.sh")
BLOCKED_BY=()
if [ -n "$DEP_NUMBERS" ]; then
while IFS= read -r dep_num; do
[ -z "$dep_num" ] && continue
# Check if dependency issue is closed (= satisfied)
DEP_STATE=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${dep_num}" | jq -r '.state // "unknown"')
if [ "$DEP_STATE" != "closed" ]; then
BLOCKED_BY+=("$dep_num")
log "dependency #${dep_num} is ${DEP_STATE} (not satisfied)"
else
log "dependency #${dep_num} is closed (satisfied)"
fi
done <<< "$DEP_NUMBERS"
fi
if [ "${#BLOCKED_BY[@]}" -gt 0 ]; then
# Find a suggestion: look for the first blocker that itself has no unmet deps
SUGGESTION=""
for blocker in "${BLOCKED_BY[@]}"; do
BLOCKER_BODY=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${blocker}" | jq -r '.body // ""')
BLOCKER_STATE=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${blocker}" | jq -r '.state')
if [ "$BLOCKER_STATE" != "open" ]; then
continue
fi
# Check if this blocker has its own unmet deps
BLOCKER_DEPS=$(echo "$BLOCKER_BODY" | \
grep -ioP '(?:depends on|blocked by|requires|after)\s+#\K[0-9]+' | sort -un || true)
BLOCKER_SECTION=$(echo "$BLOCKER_BODY" | sed -n '/^## Dependencies/,/^## /p' | sed '1d;$d')
if [ -n "$BLOCKER_SECTION" ]; then
BLOCKER_SECTION_DEPS=$(echo "$BLOCKER_SECTION" | grep -oP '#\K[0-9]+' | sort -un || true)
BLOCKER_DEPS=$(printf '%s\n%s' "$BLOCKER_DEPS" "$BLOCKER_SECTION_DEPS" | sort -un | grep -v '^$' || true)
fi
BLOCKER_BLOCKED=false
if [ -n "$BLOCKER_DEPS" ]; then
while IFS= read -r bd; do
[ -z "$bd" ] && continue
BD_STATE=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${bd}" | jq -r '.state // "unknown"')
if [ "$BD_STATE" != "closed" ]; then
BLOCKER_BLOCKED=true
break
fi
done <<< "$BLOCKER_DEPS"
fi
if [ "$BLOCKER_BLOCKED" = false ]; then
SUGGESTION="$blocker"
break
fi
done
# Write preflight result
BLOCKED_JSON=$(printf '%s\n' "${BLOCKED_BY[@]}" | jq -R 'tonumber' | jq -sc '.')
if [ -n "$SUGGESTION" ]; then
jq -n --argjson blocked "$BLOCKED_JSON" --argjson suggestion "$SUGGESTION" \
'{"status":"unmet_dependency","blocked_by":$blocked,"suggestion":$suggestion}' > "$PREFLIGHT_RESULT"
else
jq -n --argjson blocked "$BLOCKED_JSON" \
'{"status":"unmet_dependency","blocked_by":$blocked,"suggestion":null}' > "$PREFLIGHT_RESULT"
fi
# Post comment ONLY if last comment isn't already an unmet dependency notice
BLOCKED_LIST=$(printf '#%s, ' "${BLOCKED_BY[@]}" | sed 's/, $//')
LAST_COMMENT_IS_BLOCK=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${ISSUE}/comments?limit=1" | \
jq -r '.[0].body // ""' | grep -c 'Dev-agent: Unmet dependency' || true)
if [ "$LAST_COMMENT_IS_BLOCK" -eq 0 ]; then
BLOCK_COMMENT="🚧 **Dev-agent: Unmet dependency**
### Blocked by open issues
This issue depends on ${BLOCKED_LIST}, which $(if [ "${#BLOCKED_BY[@]}" -eq 1 ]; then echo "is"; else echo "are"; fi) not yet closed."
if [ -n "$SUGGESTION" ]; then
BLOCK_COMMENT="${BLOCK_COMMENT}
**Suggestion:** Work on #${SUGGESTION} first."
fi
BLOCK_COMMENT="${BLOCK_COMMENT}
---
*Automated assessment by dev-agent · $(date -u '+%Y-%m-%d %H:%M UTC')*"
printf '%s' "$BLOCK_COMMENT" > /tmp/block-comment.txt
jq -Rs '{body: .}' < /tmp/block-comment.txt > /tmp/block-comment.json
curl -sf -o /dev/null -X POST \
-H "Authorization: token ${FORGE_TOKEN}" \
-H "Content-Type: application/json" \
"${API}/issues/${ISSUE}/comments" \
--data-binary @/tmp/block-comment.json 2>/dev/null || true
rm -f /tmp/block-comment.txt /tmp/block-comment.json
else
log "skipping duplicate dependency comment"
fi
log "BLOCKED: unmet dependencies: ${BLOCKED_BY[*]}$(if [ -n "$SUGGESTION" ]; then echo ", suggest #${SUGGESTION}"; fi)"
exit 0
fi
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Preflight passed (no explicit unmet deps)
log "preflight passed — no explicit unmet dependencies"
# =============================================================================
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# CLAIM ISSUE
# =============================================================================
curl -sf -X POST \
-H "Authorization: token ${FORGE_TOKEN}" \
-H "Content-Type: application/json" \
"${API}/issues/${ISSUE}/labels" \
-d "{\"labels\":[${IN_PROGRESS_LABEL_ID}]}" >/dev/null 2>&1 || true
curl -sf -X DELETE \
-H "Authorization: token ${FORGE_TOKEN}" \
"${API}/issues/${ISSUE}/labels/${BACKLOG_LABEL_ID}" >/dev/null 2>&1 || true
CLAIMED=true
# =============================================================================
# CHECK FOR EXISTING PR (recovery mode)
# =============================================================================
EXISTING_PR=""
EXISTING_BRANCH=""
RECOVERY_MODE=false
BODY_PR=$(echo "$ISSUE_BODY_ORIGINAL" | grep -oP 'Existing PR:\s*#\K[0-9]+' | head -1) || true
if [ -n "$BODY_PR" ]; then
PR_CHECK=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/pulls/${BODY_PR}" | jq -r '{state, head_ref: .head.ref}')
PR_CHECK_STATE=$(echo "$PR_CHECK" | jq -r '.state')
if [ "$PR_CHECK_STATE" = "open" ]; then
EXISTING_PR="$BODY_PR"
EXISTING_BRANCH=$(echo "$PR_CHECK" | jq -r '.head_ref')
log "found existing PR #${EXISTING_PR} on branch ${EXISTING_BRANCH} (from issue body)"
fi
fi
if [ -z "$EXISTING_PR" ]; then
# Priority 1: match by branch name (most reliable)
FOUND_PR=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/pulls?state=open&limit=20" | \
jq -r --arg branch "$BRANCH" \
'.[] | select(.head.ref == $branch) | "\(.number) \(.head.ref)"' | head -1) || true
if [ -n "$FOUND_PR" ]; then
EXISTING_PR=$(echo "$FOUND_PR" | awk '{print $1}')
EXISTING_BRANCH=$(echo "$FOUND_PR" | awk '{print $2}')
log "found existing PR #${EXISTING_PR} on branch ${EXISTING_BRANCH} (from branch match)"
fi
fi
if [ -z "$EXISTING_PR" ]; then
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Priority 2: match "Fixes #NNN" in PR body
FOUND_PR=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/pulls?state=open&limit=20" | \
jq -r --arg issue "ixes #${ISSUE}\\b" \
'.[] | select(.body | test($issue; "i")) | "\(.number) \(.head.ref)"' | head -1) || true
if [ -n "$FOUND_PR" ]; then
EXISTING_PR=$(echo "$FOUND_PR" | awk '{print $1}')
EXISTING_BRANCH=$(echo "$FOUND_PR" | awk '{print $2}')
log "found existing PR #${EXISTING_PR} on branch ${EXISTING_BRANCH} (from body match)"
fi
fi
# Priority 3: check CLOSED PRs for prior art (don't redo work from scratch)
PRIOR_ART_DIFF=""
if [ -z "$EXISTING_PR" ]; then
CLOSED_PR=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/pulls?state=closed&limit=30" | \
jq -r --arg issue "#${ISSUE}" \
'.[] | select(.merged != true) | select((.title | contains($issue)) or (.body // "" | test("ixes " + $issue + "\\b"; "i"))) | "\(.number) \(.head.ref)"' | head -1) || true
if [ -n "$CLOSED_PR" ]; then
CLOSED_PR_NUM=$(echo "$CLOSED_PR" | awk '{print $1}')
log "found closed (unmerged) PR #${CLOSED_PR_NUM} as prior art"
PRIOR_ART_DIFF=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
"${API}/pulls/${CLOSED_PR_NUM}.diff" | head -500) || true
if [ -n "$PRIOR_ART_DIFF" ]; then
log "captured prior art diff from PR #${CLOSED_PR_NUM} ($(echo "$PRIOR_ART_DIFF" | wc -l) lines)"
fi
fi
fi
if [ -n "$EXISTING_PR" ]; then
RECOVERY_MODE=true
PR_NUMBER="$EXISTING_PR"
BRANCH="$EXISTING_BRANCH"
log "RECOVERY MODE: adopting PR #${PR_NUMBER} on branch ${BRANCH}"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
fi
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# =============================================================================
# WORKTREE SETUP
# =============================================================================
status "setting up worktree"
cd "$REPO_ROOT"
# Determine which git remote corresponds to FORGE_URL.
# When the forge is local Forgejo (not Codeberg), the remote is typically named
# "forgejo" rather than "origin". Matching by host ensures pushes target the
# correct forge regardless of remote naming conventions.
_forge_host=$(echo "$FORGE_URL" | sed 's|https\?://||; s|/.*||')
FORGE_REMOTE=$(git remote -v | awk -v host="$_forge_host" '$2 ~ host && /\(push\)/ {print $1; exit}')
FORGE_REMOTE="${FORGE_REMOTE:-origin}"
export FORGE_REMOTE # used by phase-handler.sh
log "forge remote: ${FORGE_REMOTE} (FORGE_URL=${FORGE_URL})"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
if [ "$RECOVERY_MODE" = true ]; then
if ! worktree_recover "$WORKTREE" "$BRANCH" "$FORGE_REMOTE"; then
log "ERROR: worktree recovery failed"
cleanup_labels
exit 1
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
fi
if [ "$_WORKTREE_REUSED" = true ]; then
log "reusing existing worktree (preserves session)"
fi
else
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Normal mode: create fresh worktree from primary branch
# Ensure repo is in clean state (abort stale rebases, checkout primary branch)
if [ -d "$REPO_ROOT/.git/rebase-merge" ] || [ -d "$REPO_ROOT/.git/rebase-apply" ]; then
log "WARNING: stale rebase detected in main repo — aborting"
git rebase --abort 2>/dev/null || true
fi
CURRENT_BRANCH=$(git rev-parse --abbrev-ref HEAD 2>/dev/null || echo "unknown")
if [ "$CURRENT_BRANCH" != "${PRIMARY_BRANCH}" ]; then
log "WARNING: main repo on '$CURRENT_BRANCH' instead of ${PRIMARY_BRANCH} — switching"
git checkout "${PRIMARY_BRANCH}" 2>/dev/null || true
fi
git fetch "${FORGE_REMOTE}" "${PRIMARY_BRANCH}" 2>/dev/null
git pull --ff-only "${FORGE_REMOTE}" "${PRIMARY_BRANCH}" 2>/dev/null || true
if ! worktree_create "$WORKTREE" "$BRANCH" "${FORGE_REMOTE}/${PRIMARY_BRANCH}"; then
log "ERROR: worktree creation failed"
cleanup_labels
exit 1
fi
# Symlink lib node_modules from main repo (submodule init doesn't run npm install)
for lib_dir in "$REPO_ROOT"/onchain/lib/*/; do
lib_name=$(basename "$lib_dir")
if [ -d "$lib_dir/node_modules" ] && [ ! -d "$WORKTREE/onchain/lib/$lib_name/node_modules" ]; then
ln -s "$lib_dir/node_modules" "$WORKTREE/onchain/lib/$lib_name/node_modules" 2>/dev/null || true
fi
done
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
fi
# =============================================================================
# READ SCRATCH FILE (compaction survival)
# =============================================================================
SCRATCH_CONTEXT=$(read_scratch_context "$SCRATCH_FILE")
SCRATCH_INSTRUCTION=$(build_scratch_instruction "$SCRATCH_FILE")
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# =============================================================================
# BUILD PROMPT
# =============================================================================
OPEN_ISSUES_SUMMARY=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
"${API}/issues?state=open&labels=backlog&limit=20&type=issues" | \
jq -r '.[] | "#\(.number) \(.title)"' 2>/dev/null || echo "(could not fetch)")
PHASE_PROTOCOL_INSTRUCTIONS="## Phase-Signaling Protocol (REQUIRED)
You are running in a persistent tmux session managed by an orchestrator.
Communicate progress by writing to the phase file. The orchestrator watches
this file and injects events (CI results, review feedback) back into this session.
### Key files
\`\`\`
PHASE_FILE=\"${PHASE_FILE}\"
SUMMARY_FILE=\"${IMPL_SUMMARY_FILE}\"
\`\`\`
### Phase transitions — write these exactly:
**After committing and pushing your branch:**
\`\`\`bash
# Rebase on target branch before push to avoid merge conflicts
git fetch ${FORGE_REMOTE} ${PRIMARY_BRANCH} && git rebase ${FORGE_REMOTE}/${PRIMARY_BRANCH}
git push ${FORGE_REMOTE} ${BRANCH}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Write a short summary of what you implemented:
printf '%s' \"<your summary>\" > \"\${SUMMARY_FILE}\"
# Signal the orchestrator to create the PR and watch for CI:
echo \"PHASE:awaiting_ci\" > \"${PHASE_FILE}\"
\`\`\`
Then STOP and wait. The orchestrator will inject CI results.
**When you receive a \"CI passed\" injection:**
\`\`\`bash
echo \"PHASE:awaiting_review\" > \"${PHASE_FILE}\"
\`\`\`
Then STOP and wait. The orchestrator will inject review feedback.
**When you receive a \"CI failed:\" injection:**
Fix the CI issue, then rebase on target branch and push:
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
\`\`\`bash
git fetch ${FORGE_REMOTE} ${PRIMARY_BRANCH} && git rebase ${FORGE_REMOTE}/${PRIMARY_BRANCH}
git push --force-with-lease ${FORGE_REMOTE} ${BRANCH}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
echo \"PHASE:awaiting_ci\" > \"${PHASE_FILE}\"
\`\`\`
Then STOP and wait.
**When you receive a \"Review: REQUEST_CHANGES\" injection:**
Address ALL review feedback, then rebase on target branch and push:
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
\`\`\`bash
git fetch ${FORGE_REMOTE} ${PRIMARY_BRANCH} && git rebase ${FORGE_REMOTE}/${PRIMARY_BRANCH}
git push --force-with-lease ${FORGE_REMOTE} ${BRANCH}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
echo \"PHASE:awaiting_ci\" > \"${PHASE_FILE}\"
\`\`\`
(CI runs again after each push — always write awaiting_ci, not awaiting_review)
**When you receive an \"Approved\" injection:**
The orchestrator handles merging and issue closure automatically via the bash
phase handler. You do not need to merge or close anything — stop and wait.
**When you need human help (CI exhausted, merge blocked, stuck on a decision):**
\`\`\`bash
printf 'PHASE:escalate\nReason: %s\n' \"describe what you need\" > \"${PHASE_FILE}\"
\`\`\`
Then STOP and wait. A human will review and respond via the forge.
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
**If refusing (too large, unmet dep, already done):**
\`\`\`bash
printf '%s' '{\"status\":\"too_large\",\"reason\":\"...\"}' > \"\${SUMMARY_FILE}\"
printf 'PHASE:failed\nReason: refused\n' > \"${PHASE_FILE}\"
\`\`\`
**On unrecoverable failure:**
\`\`\`bash
printf 'PHASE:failed\nReason: %s\n' \"describe what failed\" > \"${PHASE_FILE}\"
\`\`\`"
# Write phase protocol to context file for compaction survival
write_compact_context "$PHASE_FILE" "$PHASE_PROTOCOL_INSTRUCTIONS"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
if [ "$RECOVERY_MODE" = true ]; then
# Build recovery context
GIT_DIFF_STAT=$(git -C "$WORKTREE" diff "${FORGE_REMOTE}/${PRIMARY_BRANCH}..HEAD" --stat 2>/dev/null | head -20 || echo "(no diff)")
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
LAST_PHASE=$(read_phase)
rm -f "$PHASE_FILE" # Clear stale phase — new session starts clean
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
CI_RESULT=$(cat "/tmp/ci-result-${PROJECT_NAME}-${ISSUE}.txt" 2>/dev/null || echo "")
REVIEW_COMMENTS=$(curl -sf -H "Authorization: token ${FORGE_TOKEN}" \
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
"${API}/issues/${PR_NUMBER}/comments?limit=10" | \
jq -r '.[-3:] | .[] | "[\(.user.login)] \(.body[:500])"' 2>/dev/null || echo "(none)")
INITIAL_PROMPT="You are working in a git worktree at ${WORKTREE} on branch ${BRANCH}.
This is issue #${ISSUE} for the ${FORGE_REPO} project.
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
## Issue: ${ISSUE_TITLE}
${ISSUE_BODY}
${SCRATCH_CONTEXT}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
## CRASH RECOVERY
Your previous session for this issue was interrupted. Resume from where you left off.
Git is the checkpoint — your code changes survived.
### Work completed before crash:
\`\`\`
${GIT_DIFF_STAT}
\`\`\`
$(if [ "$LAST_PHASE" = "PHASE:escalate" ]; then
printf '### Previous session escalated — starting fresh\nThe previous session hit an issue and escalated. Do NOT re-escalate for the same reason.\nRead the issue and review comments carefully, then address the problem.'
else
printf '### Last known phase: %s' "${LAST_PHASE:-unknown}"
fi)
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
### PR: #${PR_NUMBER} (${BRANCH})
**IMPORTANT: PR #${PR_NUMBER} already exists — do NOT create a new PR.** Do NOT call the Codeberg/Gitea/Forgejo API to create PRs. The orchestrator manages PR creation.
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
### Recent PR comments:
${REVIEW_COMMENTS}
$(if [ -n "$CI_RESULT" ]; then printf '\n### Last CI result:\n%s\n' "$CI_RESULT"; fi)
### Next steps
1. Run \`git log --oneline -5\` and \`git status\` to understand current state.
2. **PR #${PR_NUMBER} already exists.** Address any review comments, commit, push to \`${BRANCH}\`, then write \`PHASE:awaiting_ci\`.
3. Do NOT attempt to create PRs via API calls — the orchestrator handles that.
4. Follow the phase protocol below.
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
${SCRATCH_INSTRUCTION}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
${PHASE_PROTOCOL_INSTRUCTIONS}"
else
# Normal mode: initial implementation prompt
INITIAL_PROMPT="You are working in a git worktree at ${WORKTREE} on branch ${BRANCH}.
You have been assigned issue #${ISSUE} for the ${FORGE_REPO} project.
## Issue: ${ISSUE_TITLE}
${ISSUE_BODY}
${SCRATCH_CONTEXT}
## Other open issues labeled 'backlog' (for context if you need to suggest alternatives):
${OPEN_ISSUES_SUMMARY}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
$(if [ -n "$PRIOR_ART_DIFF" ]; then
printf '## Prior Art (closed PR — DO NOT start from scratch)\n\nA previous PR attempted this issue but was closed without merging. Review the diff below and reuse as much as possible. Fix whatever caused it to fail (merge conflicts, CI errors, review findings).\n\n```diff\n%s\n```\n' "$PRIOR_ART_DIFF"
fi)
## Instructions
**Before implementing, assess whether you should proceed.** You have two options:
### Option A: Implement
If the issue is clear, dependencies are met, and scope is reasonable:
1. Read AGENTS.md in this repo for project context and coding conventions.
2. Implement the changes described in the issue.
3. Run lint and tests before you're done (see AGENTS.md for commands).
4. Commit your changes with message: fix: ${ISSUE_TITLE} (#${ISSUE})
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
5. Follow the phase protocol below to signal progress.
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
### Option B: Refuse (write JSON to SUMMARY_FILE, then write PHASE:failed)
If you cannot or should not implement this issue, write ONLY a JSON object to \$SUMMARY_FILE:
**Unmet dependency** — required code/infrastructure doesn't exist in the repo yet:
\`\`\`
{\"status\": \"unmet_dependency\", \"blocked_by\": \"short explanation of what's missing\", \"suggestion\": <issue-number-to-work-on-first or null>}
\`\`\`
**Too large** — issue needs to be split, spec is too vague, or scope exceeds a single session:
\`\`\`
{\"status\": \"too_large\", \"reason\": \"what makes it too large and how to split it\"}
\`\`\`
**Already done** — the work described is already implemented in the codebase:
\`\`\`
{\"status\": \"already_done\", \"reason\": \"where the existing implementation is\"}
\`\`\`
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
Then write:
\`\`\`bash
printf 'PHASE:failed\nReason: refused\n' > \"${PHASE_FILE}\"
\`\`\`
### How to decide
- Read the issue carefully. Check if files/functions it references actually exist in the repo.
- If it depends on other issues, check if those issues' deliverables are present in the codebase.
- If the issue spec is vague or requires designing multiple new systems, refuse as too_large.
- If another open issue should be done first, suggest it.
- When in doubt, implement. Only refuse if there's a clear, specific reason.
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
**Do NOT invent dependencies that aren't real.** If the code compiles and tests pass, that's ready.
${SCRATCH_INSTRUCTION}
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
${PHASE_PROTOCOL_INSTRUCTIONS}"
fi
# =============================================================================
# CREATE TMUX SESSION
# =============================================================================
status "creating tmux session: ${SESSION_NAME}"
if ! create_agent_session "${SESSION_NAME}" "${WORKTREE}" "${PHASE_FILE}"; then
log "ERROR: failed to create agent session"
cleanup_labels
cleanup_worktree
exit 1
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
fi
# Send initial prompt into the session
inject_formula "${SESSION_NAME}" "${INITIAL_PROMPT}"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
log "initial prompt sent to tmux session"
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
# Signal to dev-poll.sh that we're running (session is up)
echo '{"status":"ready"}' > "$PREFLIGHT_RESULT"
status "monitoring phase: ${PHASE_FILE}"
monitor_phase_loop "$PHASE_FILE" "$IDLE_TIMEOUT" _on_phase_change
# Handle exit reason from monitor_phase_loop
case "${_MONITOR_LOOP_EXIT:-}" in
idle_timeout|idle_prompt)
# Post diagnostic comment + label issue blocked
fix: refactor: replace escalation JSONL with blocked label + diagnostic comment (#352) Replace the unreliable escalation JSONL system (supervisor/escalations-*.jsonl consumed by gardener) with direct blocked label + diagnostic comment on the original issue. When a dev-agent or action-agent session fails (PHASE:failed, idle timeout, crash, CI exhausted): - Capture last 50 lines from tmux pane via tmux capture-pane - Post a structured diagnostic comment on the issue (exit reason, timestamp, PR number, tmux output) - Label the issue "blocked" (instead of restoring "backlog") - Remove in-progress label Removed: - Escalation JSONL write paths in dev-agent.sh, phase-handler.sh, dev-poll.sh, action-agent.sh - is_escalated() helper in dev-poll.sh - Escalation triage (P2f section) in supervisor-poll.sh - Escalation processing + recipe engine in gardener-poll.sh - ci-escalation-recipes step from run-gardener.toml formula - escalations*.jsonl from .gitignore Added: - post_blocked_diagnostic() shared helper in phase-handler.sh - ensure_blocked_label_id() helper (creates label via API if not exists) - is_blocked() helper in dev-poll.sh (replaces is_escalated) - Blocked issues listing in supervisor/preflight.sh Kept: - Matrix notifications on failure (unchanged) - CI fix counter logic (still tracks attempts) - needs_human injection in supervisor/gardener (not escalation-related) - Gardener grooming (gardener-agent.sh still invoked) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-21 04:18:43 +00:00
post_blocked_diagnostic "${_MONITOR_LOOP_EXIT:-idle_timeout}"
if [ -n "${PR_NUMBER:-}" ]; then
log "keeping worktree (PR #${PR_NUMBER} still open)"
else
cleanup_worktree
fi
rm -f "$PHASE_FILE" "${PHASE_FILE%.phase}.context" \
"$IMPL_SUMMARY_FILE" "$SCRATCH_FILE" \
"/tmp/ci-result-${PROJECT_NAME}-${ISSUE}.txt"
[ -n "${PR_NUMBER:-}" ] && rm -f "/tmp/review-injected-${PROJECT_NAME}-${PR_NUMBER}"
;;
crashed)
# Belt-and-suspenders: _on_phase_change(PHASE:crashed) handles primary
fix: refactor: replace escalation JSONL with blocked label + diagnostic comment (#352) Replace the unreliable escalation JSONL system (supervisor/escalations-*.jsonl consumed by gardener) with direct blocked label + diagnostic comment on the original issue. When a dev-agent or action-agent session fails (PHASE:failed, idle timeout, crash, CI exhausted): - Capture last 50 lines from tmux pane via tmux capture-pane - Post a structured diagnostic comment on the issue (exit reason, timestamp, PR number, tmux output) - Label the issue "blocked" (instead of restoring "backlog") - Remove in-progress label Removed: - Escalation JSONL write paths in dev-agent.sh, phase-handler.sh, dev-poll.sh, action-agent.sh - is_escalated() helper in dev-poll.sh - Escalation triage (P2f section) in supervisor-poll.sh - Escalation processing + recipe engine in gardener-poll.sh - ci-escalation-recipes step from run-gardener.toml formula - escalations*.jsonl from .gitignore Added: - post_blocked_diagnostic() shared helper in phase-handler.sh - ensure_blocked_label_id() helper (creates label via API if not exists) - is_blocked() helper in dev-poll.sh (replaces is_escalated) - Blocked issues listing in supervisor/preflight.sh Kept: - Matrix notifications on failure (unchanged) - CI fix counter logic (still tracks attempts) - needs_human injection in supervisor/gardener (not escalation-related) - Gardener grooming (gardener-agent.sh still invoked) Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
2026-03-21 04:18:43 +00:00
# cleanup (diagnostic comment, blocked label, worktree, files).
# Only post if the callback didn't already (guard prevents double comment).
if [ "${_BLOCKED_POSTED:-}" != "true" ]; then
post_blocked_diagnostic "crashed"
fi
;;
done)
# Belt-and-suspenders: callback in phase-handler.sh handles primary cleanup,
# but ensure sentinel files are removed if callback was interrupted
rm -f "$PHASE_FILE" "${PHASE_FILE%.phase}.context" \
"$IMPL_SUMMARY_FILE" "$SCRATCH_FILE" \
"/tmp/ci-result-${PROJECT_NAME}-${ISSUE}.txt"
[ -n "${PR_NUMBER:-}" ] && rm -f "/tmp/review-injected-${PROJECT_NAME}-${PR_NUMBER}"
CLAIMED=false
;;
esac
fix: feat: tmux session manager in dev-agent.sh (#80) Replace fire-and-forget `claude -p` calls with a persistent tmux session that Claude Code runs in interactively. The orchestrator (dev-agent.sh) monitors a phase file and reacts to Claude's signals: - Session lifecycle: create `dev-{project}-{issue}` tmux session, send the full initial prompt (issue body + phase protocol instructions) via `tmux load-buffer` / `tmux paste-buffer`, then enter a phase monitor loop. - Phase monitor loop: polls `/tmp/dev-session-{project}-{issue}.phase` every 30s for mtime changes. Handles all five phase sentinels: - PHASE:awaiting_ci → create PR if needed, poll CI, inject result - PHASE:awaiting_review → poll for review comment, inject verdict - PHASE:needs_human → send Matrix notification, wait for injection - PHASE:done → call do_merge(), exit on success - PHASE:failed → detect refusal JSON vs genuine failure, post comment / escalate, kill session, restore backlog - Crash recovery: if the tmux session dies unexpectedly, dev-agent.sh restarts it in the same worktree and injects a recovery prompt with the last known phase and git diff. - Idle timeout: 2h with no phase update kills the session gracefully. - PR creation moved into the PHASE:awaiting_ci handler; Claude pushes the branch and writes the phase, orchestrator creates the PR and starts CI. - Summary file `/tmp/dev-impl-summary-{project}-{issue}.txt` carries the implementation summary (for PR body) and refusal JSON between Claude and the orchestrator. - All existing logic preserved: dep preflight, label management, do_merge() with rebase retry, CI escalation, prior art detection, log rotation. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
2026-03-17 20:20:38 +00:00
log "dev-agent finished for issue #${ISSUE}"