Kanban goal_mode workers never enter the goal loop — dispatcher spawns `chat -q` but the hook is only wired in the `-Q/--quiet` path

[MarkWin91]

Summary

Cards created with --goal never exercise the Ralph-style goal loop. The dispatcher sets HERMES_KANBAN_GOAL_MODE=1 in the worker env, but the spawned worker command (hermes -p <profile> chat -q "work kanban task <id>") takes the single-query branch of main(), which never reads that env var. The goal hook is only wired in the fully-quiet -Q/--quiet machine-readable branch. Result: goal_mode is stored on the card and silently ignored; any apparent progress comes from the dispatcher's ordinary retry loop, and a worker that leaves the task open (expecting the judge to continue it) is reaped as protocol_violation.

Where the wires cross (0.16.0 line numbers; same structure in 0.18.2)

• hermes_cli/kanban_db.py:6673-6676 (_default_spawn) — sets HERMES_KANBAN_GOAL_MODE=1 and spawns [hermes, -p, <profile>, --accept-hooks, chat, -q, <prompt>]. The comment says "the worker reads these and wraps its run in the Ralph-style /goal judge loop (see cli.py quiet-mode path)" — but the spawned form never reaches that path.
• cli.py:16010-16014 — the only reader of HERMES_KANBAN_GOAL_MODE, calling _run_kanban_goal_loop_q(...). It lives in the -Q/--quiet machine-readable branch (the same block that owns the kanban EX_TEMPFAIL exit-code contract).
• cli.py:16048-16070 — the else branch, explicitly documented "Single-query mode (hermes chat -q "…")": cli.chat(query) + _print_exit_summary(). No goal hook.

Evidence from a live board

Worker logs of a goal_mode card end with the _print_exit_summary() block ("Resume this session with… / Duration / Messages") and never contain the session_id: stderr line of the quiet branch — although _default_spawn merges stderr into the log (stderr=subprocess.STDOUT, kanban_db.py:6768). Three consecutive runs of one goal card each produced exactly one turn in a fresh session, no judge verdict, no continuation; goals.run_kanban_goal_loop itself works when invoked directly (unit-style, mocked cli).

Still present in 0.18.2 (latest)

• spawn unchanged: kanban_db.py:7794-7796 still ["chat", "-q", prompt]
• hook unchanged: cli.py:16113 (quiet branch) vs single-query branch 16152-16171

Suggested fix (either side works)

1. Wire the HERMES_KANBAN_GOAL_MODE check into the single-query branch too (after cli.chat(...) returns the first response), or
2. Spawn goal-mode workers through the -Q/--quiet path the hook already lives in.

[MarkWin91]

Proposed patch — against the single-query branch in cli.py (the else branch documented "Single-query mode (hermes chat -q "…")"). It mirrors the existing goal-loop wiring in the -Q/--quiet branch, reusing cli.chat()'s documented Optional[str] return (the agent's final response) as the loop's first_response:

             cli._show_security_advisories()
-            cli.chat(query, images=single_query_images or None)
+            _kanban_first = cli.chat(query, images=single_query_images or None)
+            # Kanban goal-loop mode also applies to the human-facing single-query
+            # path: the dispatcher spawns goal_mode workers via `chat -q`, which
+            # lands here (not the -Q branch). Without this the env var is set but
+            # never read, so goal_mode is silently ignored.
+            if os.environ.get("HERMES_KANBAN_GOAL_MODE") == "1":
+                try:
+                    _run_kanban_goal_loop_q(cli, _kanban_first or "")
+                except Exception as _goal_exc:
+                    logger.debug("kanban goal loop failed: %s", _goal_exc)
             cli._print_exit_summary()

cli.chat() is documented as returning Optional[str] ("The agent's response, or None on error"), so it supplies the same first_response the -Q branch already passes to _run_kanban_goal_loop_q. This is the smaller of the two fixes suggested above (wire the check into the single-query branch).

Transparency: I have not run this against your CI — flagging for maintainer review + testing rather than opening an untested PR.

[teknium1]

Thanks for the precise trace and the proposed single-query-path patch. This is already fixed on current main through the alternative spawn-side route described in the issue.

Automated hermes-sweeper review:

• Commit 80b58ec71dd37d181358787912d40f1a8137fdf8 (fix(kanban): spawn goal_mode workers with -Q so the goal loop actually runs) appends -Q for task.goal_mode workers.
• Current hermes_cli/kanban_db.py:8203-8209 therefore launches goal-mode workers through the fully quiet path.
• Current cli.py:16391-16402 consumes HERMES_KANBAN_GOAL_MODE in that path and calls _run_kanban_goal_loop_q.
• Cross-referenced PR fix(kanban): run goal loop for worker queries #63443 proposed the other viable implementation (wiring the human-facing chat -q branch); main already satisfies the requested behavior via the dispatcher spawn fix.