Three fixes:
1. architect-run.sh:722 — extract `.result` not `.content` from claude JSON
output. All other callers (dev-agent, formula-session) use `.result`;
this was the direct cause of every pitch being empty.
2. lib/agent-sdk.sh — reset `_AGENT_LAST_OUTPUT=""` at the top of each
`agent_run` call so stale data from a prior invocation can't bleed
into the next caller when claude crashes or returns empty.
3. lib/agent-sdk.sh — scope the diagnostics file by `$LOG_AGENT` instead
of hardcoding `dev/`. Concurrent agents (architect, gardener, planner,
predictor) no longer clobber each other's diag output.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Gate both flock call sites (agent_run main invocation and nudge) behind
CLAUDE_EXTERNAL_LOCK env var. Default off — the native Claude Code
proper-lockfile-based OAuth refresh lock handles concurrency. Set
CLAUDE_EXTERNAL_LOCK=1 to re-enable the external flock for rollback.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Capture exit code from claude invocations instead of suppressing with || true.
Log timeout (rc=124) and non-zero exits distinctly. Skip nudge when output is
empty (claude crashed or failed). Log empty output as a clear diagnostic message.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Some models (especially local) emit end_turn prematurely. After
agent_run completes, check if code was pushed. If not, resume the
session with a nudge: "You stopped but did not push. Complete the
implementation, commit, and push."
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Save agent_run output to agent-run-last.json. On no_push, log the
result text, turn count, and cost. Save full output to
no-push-{issue}-{ts}.json for later analysis.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Move SID_FILE recovery into agent_recover_session() in lib/agent-sdk.sh
to eliminate remaining duplicate block between dev-agent.sh and
review-pr.sh.
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
Extract agent_run() into shared lib/agent-sdk.sh to eliminate code
duplication between dev-agent.sh and review-pr.sh (CI dedup check).
Rewrite review-pr.sh from tmux-based agent-session.sh to synchronous
claude -p invocations via shared agent-sdk.sh, matching the SDK pattern
from dev-agent.sh (#798).
Key changes:
- Create lib/agent-sdk.sh with shared agent_run() function
- Both dev-agent.sh and review-pr.sh now source lib/agent-sdk.sh
instead of defining agent_run() inline
- Replace agent-session.sh (tmux + monitor_phase_loop) with agent_run()
- Add .sid file for session continuity: re-reviews resume the original
session via --resume, so Claude remembers its prior review
- Use worktree.sh for worktree cleanup
- Remove phase file signaling — completion is automatic when claude -p
returns
- Keep all review business logic unchanged
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
