Manager route parsing rejects markdown-formatted Next: line (backtick-wrapped or inline with Task:), causing repeated empty rounds #80

Summary
When parse_role_manager_next_step parses the manager's plan output, a Next: cli / Next: done route that is wrapped in backticks (as shown in the manager prompt template examples) or placed on the same line as the following Task: is treated as MANAGER_NEXT_INVALID. The harness then returns "did not contain a valid route" feedback and the manager retries, causing many empty protocol-correction rounds (observed up to 10 rounds, wasting time and tokens on every run).
Environment

lh-harness 0.1.7 (lh-harness doctor all green)
Agent backend: opencode 1.18.29
Model: custom OpenAI-compatible endpoint (LiteLLM → DeepSeek API)
OS: Ubuntu 24.04

Observed behavior
With a moderately complex task, the manager output a valid route but the harness rejected it, looping up to 10 rounds before any executor/auditor round ran:

Harness feedback (same for every empty round):
    "Audit facts: the manager output did not contain a valid route, so no GUI or CLI executor can be assigned. Gap: emit one dominant GUI/CLI subtask or an explicit ask/done/blocked route. Next step: manage again using exactly Next: gui, Next: cli, Next: ask, Next: done, or Next: blocked."

The actual manager output (from provider_stream.jsonl, raw model stream) contained the route, e.g.:
    `Next: cliTask:` Read-only CLI diagnostic ...``

and later for completion:
    `Next: done`

i.e. the route line is backtick-wrapped (exactly matching the style shown in the manager system prompt: "then exactly one route: Next: gui, Next: cli, ...") and in the first case shares the line with the following Task: section.

manager_plan.txt confirmed the same content: the model output the required sections plus a route, yet next_step came back as MANAGER_NEXT_INVALID.

Root cause
parse_role_manager_next_step in role_prompts.py (line ~481) matches each line as an exact, whole-line string after strip().strip("*").replace(" ","").lower():

for line in str(text or "").splitlines():
normalized = line.strip().strip("*").replace(" ", "").replace("　", "").lower()
normalized = re.split(r"(?:—|–|--|//|#|[（(])", normalized, maxsplit=1)[0]
if normalized in {"下一步:gui任务", "下一步：gui任务", "next:gui"}:
return MANAGER_NEXT_GUI
if normalized in {"...", "next:cli"}:
return MANAGER_NEXT_CLI
...
return MANAGER_NEXT_INVALID

Because the manager system prompt itself teaches the route with backticks ("Next: cli"), models commonly emit:
backtick-wrapped routes: next:cli → normalized to next:cli ≠ next:cli
or merge the route with the next section: Next: cliTask:... `` → normalized to ``next:clitask:... ≠ next:cli
Both fail the exact-match set, so a perfectly valid route is dropped.

Reproduction
1、Create any multi-step file task (e.g. "create a config.toml and a build-note.md in output/, then verify").
2、Run lh-harness run --agent opencode --task @task.md.
3、Observe several rounds where manager is done · xx s · next=invalid before the first next=cli appears, and check runs//harness/orchestration/rounds/round_001/manager_plan.txt (backtick-wrapped route) + harness_feedback.txt ("did not contain a valid route").
4、In the completion round, a backtick-wrapped Next: done is also rejected once.

Expected behavior
parse_role_manager_next_step should be tolerant of markdown decoration and trailing commentary. Concretely: strip backticks/asterisks before comparison, and accept the route when the normalized line contains one of the route tokens (with a word boundary), e.g.:

bare = normalized.replace("`", "").replace("*", "")
if "next:gui" in bare or bare in {"下一步:gui任务", "下一步：gui任务"}:
...
This also keeps the existing Chinese-route set working.

Impact
Wasted rounds (up to 10) and tokens on every multi-round task.
Under a strict round budget the task can exhaust max_rounds purely on protocol-correction rounds.
Only affects manager role; executor/auditor pipelines work fine.

Verification of proposed fix
With the tolerant parse locally patched, the same task went from 10 empty rounds → 0 empty rounds (2 rounds / 2.4 min total), including a clean Next: done acceptance.