@seanyao/roll 2026.601.2 → 2026.601.3

package/CHANGELOG.md

@@ -1,5 +1,17 @@
 # Changelog
 
+## v2026.601.3
+
+### 可见性
+
+- **kimi cycle 现在也能看到 token 和成本(FIX-154)** — 以前 dashboard 对 kimi 那一行全是 `—/—`,看不到主力 agent 花了多少钱;现在 cycle 跑完读 kimi-code 的 `wire.jsonl`,把 token 数和成本写进事件流,RECENT 视图和成本总闸都看得见 `[loop]`
+
+### 稳定性
+
+- **loop 把活派给 AI 后现在真会动手,不再空转零产出(FIX-152)** — kimi 等对话式 agent 拿到 SKILL.md 会把它当成"贴过来的文档"反问"What would you like me to do?",8 秒空返没交付;技能正文前置一条 agent 无关的自主执行指令,kimi/claude/pi/codex/agy 现在都会直接动手 `[loop]`
+- **agy 在 loop / cron 自动化里不再卡 tty 等待(FIX-153)** — antigravity(agy)默认要 tty 批准操作,自动化场景拿不到 tty 就一直挂着等;现在 headless 模式自动加 `-p` 和跳过权限标记,跑得到结果 `[loop]`
+- **测试不再在桌面弹空报错终端(FIX-155)** — bats 测试跑完临时沙箱删了,但 peer auto-attach 弹的 Terminal 窗口指向那个已不存在的路径,桌面堆一堆空报错的死窗口;给 peer 弹窗补上和 loop 弹窗一样的测试守卫,测试上下文不再弹 `[loop]`
+
 ## v2026.601.2
 
 ### 新功能

package/bin/roll

@@ -4,7 +4,7 @@ set -euo pipefail
 # Roll — AI Agent Convention Manager
 # Single source of truth for how all AI coding agents behave.
 
-VERSION="2026.601.2"
+VERSION="2026.601.3"
 ROLL_HOME="${ROLL_HOME:-${HOME}/.roll}"
 ROLL_CONFIG="${ROLL_HOME}/config.yaml"
 ROLL_GLOBAL="${ROLL_HOME}/conventions/global"
@@ -3841,6 +3841,8 @@ _peer_route() {
 _peer_auto_attach() {
   local session="$1"
   [ "$(uname)" = "Darwin" ] || return 0
+  [ -n "${BATS_TEST_NUMBER:-}" ] && return 0
+  [ -n "${ROLL_LOOP_NO_POPUP:-}" ] && return 0
   [ -f "$_LOOP_MUTE_FILE" ] && return 0
   local attach_cmd="${_SHARED_ROOT}/loop/attach-${session}.command"
   # Drop `exec` so the wrapping shell survives `tmux attach` exiting; pause
@@ -4411,10 +4413,12 @@ _agent_argv() {
       # late 2025. agy reuses ~/.gemini/ for config and reads GEMINI.md
       # natively, so the convention sync target is unchanged — only the
       # invoked binary changes.
+      # FIX-153: non-interactive modes must use -p (headless) +
+      # --dangerously-skip-permissions so the agent does not hang waiting
+      # for a tty approval that never comes in loop/cron contexts.
       case "$mode" in
         interactive) _AGENT_ARGV=(agy -i "$prompt") ;;
-        text|peer)   _AGENT_ARGV=(agy "$prompt") ;;
-        *)           _AGENT_ARGV=(agy "$prompt") ;;
+        *)           _AGENT_ARGV=(agy -p --dangerously-skip-permissions "$prompt") ;;
       esac ;;
     qwen)
       # qwen has the same argv shape in both modes (positional prompt).
@@ -8162,7 +8166,11 @@ _write_loop_runner_script() {
   # US-LOOP-026: post-cycle single-shot usage writer for non-claude agents.
   # pi -p text mode prints no usage, so we recover it from pi's session jsonl
   # exactly once per cycle (loop-fmt passthrough is display-only).
+  # FIX-154: kimi-code's `-p` mode also writes nothing to stdout but persists
+  # usage to wire.jsonl; kimi_emit covers that path. bin/roll dispatches by
+  # agent (pi/deepseek → pi_emit, kimi → kimi_emit).
   local pi_emit_script="${ROLL_PKG_DIR}/lib/agent_usage/pi_emit.py"
+  local kimi_emit_script="${ROLL_PKG_DIR}/lib/agent_usage/kimi_emit.py"
   local roll_bin="${ROLL_PKG_DIR}/bin/roll"
   # US-EVAL-002: pure-function rubric scorer (US-EVAL-001). Baked in at
   # generation time so the inner runner can compute result_eval at cycle finish.
@@ -8831,23 +8839,37 @@ else
   _phase_end agent_invoke ok
 fi
 
-# US-LOOP-026: non-claude agents (pi/deepseek/kimi) print no usage in -p text
-# mode. Recover token+cost once per cycle from the agent's session jsonl and
-# append a single authoritative usage event. Done here (not in loop-fmt's
-# per-attempt passthrough) so retries can't write N duplicate events that the
-# dashboard's same-label SUM would inflate. Runs before the timeout-abort exit
-# so partial cycles still get whatever usage the session recorded. The events
-# path is resolved exactly like _loop_event (rt_dir first, shared fallback) so
-# pi_emit appends to the same file the reader consumes.
-if [ "\$(_project_agent)" != "claude" ] && [ -f "${pi_emit_script}" ]; then
+# US-LOOP-026 + FIX-154: non-claude agents (pi/deepseek/kimi) print no usage
+# in -p text mode. Recover token+cost once per cycle from the agent's session
+# jsonl and append a single authoritative usage event. Done here (not in
+# loop-fmt's per-attempt passthrough) so retries can't write N duplicate
+# events that the dashboard's same-label SUM would inflate. Runs before the
+# timeout-abort exit so partial cycles still get whatever usage the session
+# recorded. The events path is resolved exactly like _loop_event (rt_dir
+# first, shared fallback) so the emitter appends to the same file the reader
+# consumes. Dispatch by agent so each emitter reads the right session format
+# (pi.usage_from_session vs kimi.usage_from_session).
+if [ "\$(_project_agent)" != "claude" ]; then
   _pi_rt=\$(_loop_runtime_dir "${slug}" 2>/dev/null || echo "")
   if [ -n "\$_pi_rt" ]; then
     _pi_evfile="\${_pi_rt}/events.ndjson"
   else
     _pi_evfile="\${_SHARED_ROOT:-\$HOME/.shared/roll}/loop/events-${slug}.ndjson"
   fi
-  python3 "${pi_emit_script}" --cwd "\$WT" --cycle "\${CYCLE_ID}" \\
-    --slug "${slug}" --events "\$_pi_evfile" 2>/dev/null || true
+  case "\$(_project_agent)" in
+    kimi)
+      if [ -f "${kimi_emit_script}" ]; then
+        python3 "${kimi_emit_script}" --cwd "\$WT" --cycle "\${CYCLE_ID}" \\
+          --slug "${slug}" --events "\$_pi_evfile" 2>/dev/null || true
+      fi
+      ;;
+    *)
+      if [ -f "${pi_emit_script}" ]; then
+        python3 "${pi_emit_script}" --cwd "\$WT" --cycle "\${CYCLE_ID}" \\
+          --slug "${slug}" --events "\$_pi_evfile" 2>/dev/null || true
+      fi
+      ;;
+  esac
 fi
 
 # FIX-057: timed out — skip publish; EXIT trap writes cycle_end blocked + ALERT.
@@ -9423,7 +9445,15 @@ _agent_skill_cmd() {
   for ((i = 1; i < prompt_idx; i++)); do
     out+=" ${_AGENT_ARGV[i]}"
   done
-  echo "${out} \"\$(${strip})\""
+  # FIX-152: prepend an explicit autonomous-execution directive ahead of the bare
+  # SKILL.md body. Without it, conversational `-p` agents (notably kimi-code) read
+  # the skill doc as pasted context and reply "what would you like me to do?",
+  # returning in seconds with zero output → the cycle ends idle, no delivery.
+  # pi/deepseek/claude tolerate the bare doc, but the directive is agent-agnostic
+  # and hardens every autonomous cron skill (loop/dream/brief share this chokepoint).
+  # Must stay free of " $ ` \ so it survives the later `eval` of the cycle command.
+  local _autorun='[roll 自主模式] 你正在无人值守的自动化循环中运行,这不是对话。请立即、完整地执行下面这份技能文档描述的工作流,直到完成交付或写出 ALERT 为止;严禁反问、严禁等待确认、严禁只复述或总结而不动手。技能文档如下: '
+  echo "${out} \"${_autorun}\$(${strip})\""
 }
 
 # FIX-134: build the full per-cycle agent command at RUNTIME, routing-aware.

package/lib/agent_usage/kimi.py

@@ -1,29 +1,33 @@
 """
 kimi (Moonshot Kimi CLI) agent usage extractor.
 
-Like openai and gemini (and unlike pi, which persists usage to session
-files), the Kimi CLI prints a token-usage summary to stdout at the end of a
-session.  So this plugin implements the standard ``extract()`` registry
-contract: scrape the passthrough stdout lines for the usage / model lines.
+Two paths are supported, mirroring pi.py:
 
-Recognised lines (case-insensitive, robust to thousands separators)::
+1. ``extract()`` — the registry stdout-scrape contract, kept for legacy
+   callers (and as a fallback when session files are absent).
+2. ``usage_from_session()`` — authoritative recovery from kimi-code's
+   persisted session files at ``~/.kimi-code/sessions/wd_*/session_*/agents/main/wire.jsonl``.
+   Each wire file is NDJSON with one or more ``{"type":"usage.record","model":...,"usage":{...}}``
+   lines whose token fields are summed per cycle.
 
-    Model: kimi-k2
-    Tokens: input=15300 output=3120
+FIX-154 added the session path so loop cycles run by kimi-code (the
+default agent today) no longer show ``—/—`` for tokens and cost in the
+RECENT dashboard.
 
-The Kimi CLI's "usage" / session-summary block is also accepted::
+The stdout-scrape contract still recognises (case-insensitive)::
 
+    Model: kimi-k2
+    Tokens: input=15300 output=3120
     Input tokens:  15,300
     Output tokens:  3,120
     Total tokens:  18,420
-    model: kimi-k2
 
 When an explicit USD cost line isn't present, cost is computed from
-``lib/model_prices.py`` (list price) so the dashboard never shows ``—``
-for a recognised kimi cycle.  Returns None if no usage line is found,
-so the caller falls back to the null payload (US-LOOP-010 compatible).
+``lib/model_prices.py`` (list price).
 """
 
+import glob
+import json
 import os
 import re
 import sys
@@ -125,3 +129,150 @@ def extract(stdin_lines: list[str]) -> Optional[dict]:
         "cost_list_usd": cost,
         "duration_ms": None,
     }
+
+
+# ── Session-file extraction (authoritative, FIX-154) ───────────────────────
+
+# kimi-code persists every CLI session under
+# ``~/.kimi-code/sessions/wd_<cwd-basename>_<8-hex>/session_<uuid>/agents/main/wire.jsonl``
+# where ``<cwd-basename>`` is the basename of the cycle's worktree
+# (e.g. ``roll-ecf079-cycle-20260601-170905-54957``).
+# Each wire file is NDJSON; one or more lines have::
+#
+#     {"type": "usage.record", "model": "kimi-code/kimi-for-coding",
+#      "usage": {"inputOther": <int>, "output": <int>,
+#                "inputCacheRead": <int>, "inputCacheCreation": <int>},
+#      "usageScope": "turn", "time": <ms>}
+#
+# We sum across all matching wire files (retries reuse the same worktree).
+
+
+def _kimi_sessions_base_dir(base_dir: Optional[str]) -> str:
+    """Resolve kimi-code's sessions root: arg → env → default."""
+    return (
+        base_dir
+        or os.environ.get("ROLL_KIMI_SESSIONS_DIR")
+        or os.path.expanduser("~/.kimi-code/sessions")
+    )
+
+
+def _sum_wire_file(path: str) -> Optional[dict]:
+    """Sum ``usage.record`` lines in a single kimi wire.jsonl.
+
+    Returns a usage dict or None when no usage records are found.
+    Field mapping kimi → roll::
+
+        inputOther         → input_tokens
+        output             → output_tokens
+        inputCacheRead     → cache_read_tokens
+        inputCacheCreation → cache_creation_tokens
+    """
+    tin = tout = tcr = tcw = 0
+    model = None
+    seen = False
+    try:
+        with open(path) as f:
+            for line in f:
+                line = line.strip()
+                if not line:
+                    continue
+                try:
+                    o = json.loads(line)
+                except json.JSONDecodeError:
+                    continue
+                if o.get("type") != "usage.record":
+                    continue
+                u = o.get("usage") or {}
+                seen = True
+                if o.get("model"):
+                    model = o["model"]
+                tin += int(u.get("inputOther") or 0)
+                tout += int(u.get("output") or 0)
+                tcr += int(u.get("inputCacheRead") or 0)
+                tcw += int(u.get("inputCacheCreation") or 0)
+    except OSError:
+        return None
+    if not seen:
+        return None
+    return {
+        "model": model or _DEFAULT_MODEL,
+        "input_tokens": tin,
+        "output_tokens": tout,
+        "cache_creation_tokens": tcw,
+        "cache_read_tokens": tcr,
+        "duration_ms": None,
+    }
+
+
+def usage_from_session(
+    cwd: Optional[str] = None,
+    cycle_id: Optional[str] = None,
+    slug: Optional[str] = None,
+    base_dir: Optional[str] = None,
+) -> Optional[dict]:
+    """Recover a kimi cycle's usage by reading its persisted wire file(s).
+
+    Matching: scan ``<base>/wd_*/session_*/agents/main/wire.jsonl`` and
+    select files whose ``wd_*`` directory name contains the worktree
+    basename (authoritative when ``cwd`` is given) or the ``cycle_id``
+    substring (fallback).
+
+    Retries can produce multiple wire files for the same cycle; their
+    usage is SUMMED so token totals reflect retry work too.
+
+    Returns the merged usage dict (tokens + model), or None when nothing
+    matches / zero tokens — caller writes nothing in that case, preserving
+    "n/a, not fake zero".
+    """
+    base = _kimi_sessions_base_dir(base_dir)
+    files = sorted(glob.glob(
+        os.path.join(base, "wd_*", "session_*", "agents", "main", "wire.jsonl")
+    ))
+    if not files:
+        return None
+
+    cwd_basename = os.path.basename(cwd.rstrip("/")) if cwd else None
+    matched = []
+    for path in files:
+        # Session dir name: wd_<cwd-basename>_<8-hex>
+        # Path: <base>/wd_<cwd-basename>_<hash>/session_<uuid>/agents/main/wire.jsonl
+        wd_seg = path[len(base):].lstrip(os.sep).split(os.sep, 1)[0]
+        if cwd_basename and ("wd_%s_" % cwd_basename) in (wd_seg + "_"):
+            matched.append(path)
+            continue
+        if cycle_id and ("cycle-%s" % cycle_id) in wd_seg:
+            matched.append(path)
+
+    if not matched:
+        return None
+
+    agg = {
+        "model": None,
+        "input_tokens": 0,
+        "output_tokens": 0,
+        "cache_creation_tokens": 0,
+        "cache_read_tokens": 0,
+        "duration_ms": None,
+    }
+    got = False
+    for path in matched:
+        s = _sum_wire_file(path)
+        if s is None:
+            continue
+        got = True
+        agg["model"] = agg["model"] or s["model"]
+        agg["input_tokens"] += s["input_tokens"]
+        agg["output_tokens"] += s["output_tokens"]
+        agg["cache_creation_tokens"] += s["cache_creation_tokens"]
+        agg["cache_read_tokens"] += s["cache_read_tokens"]
+
+    if not got:
+        return None
+    has_tokens = (
+        agg["input_tokens"] or agg["output_tokens"]
+        or agg["cache_creation_tokens"] or agg["cache_read_tokens"]
+    )
+    if not has_tokens:
+        return None
+    agg["model"] = agg["model"] or _DEFAULT_MODEL
+    return agg

package/lib/agent_usage/kimi_emit.py

@@ -0,0 +1,123 @@
+#!/usr/bin/env python3
+"""
+kimi_emit — write ONE authoritative usage event for a finished kimi cycle.
+
+Mirror of ``pi_emit.py``: invoked once by bin/roll after the agent phase
+when ROLL_LOOP_AGENT == "kimi". Recovers the cycle's real usage from
+kimi-code's persisted ``wire.jsonl`` files via ``kimi.usage_from_session``
+and appends a single ``stage=="usage"`` event to the loop events file.
+
+Exactly one event per cycle — the dashboard SUMS token fields across
+same-label usage events, so a per-retry write path would inflate ×N.
+
+Cost is frozen at the active price snapshot via
+``model_prices.compute_list_cost`` in the model's native currency.
+
+When ``usage_from_session`` finds nothing (no matching session, zero
+tokens) we write nothing — preserving "show n/a, not a fake zero".
+"""
+
+import argparse
+import importlib.util
+import json
+import os
+import sys
+from datetime import datetime, timezone
+
+_THIS_DIR = os.path.dirname(os.path.abspath(__file__))
+_LIB_DIR = os.path.dirname(_THIS_DIR)
+
+
+def _load_model_prices():
+    spec = importlib.util.spec_from_file_location(
+        "model_prices", os.path.join(_LIB_DIR, "model_prices.py")
+    )
+    mp = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(mp)
+    return mp
+
+
+def _load_kimi():
+    spec = importlib.util.spec_from_file_location(
+        "agent_usage_kimi", os.path.join(_THIS_DIR, "kimi.py")
+    )
+    kimi = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(kimi)
+    return kimi
+
+
+def build_event(cwd=None, cycle_id=None, slug=None, base_dir=None):
+    """Return the (line dict) usage event for a kimi cycle, or None to skip."""
+    kimi = _load_kimi()
+    u = kimi.usage_from_session(
+        cwd=cwd, cycle_id=cycle_id, slug=slug, base_dir=base_dir
+    )
+    if u is None:
+        return None
+
+    mp = _load_model_prices()
+    model = u.get("model") or "kimi-k2.5"
+    totals = {
+        "input_tokens": int(u.get("input_tokens") or 0),
+        "output_tokens": int(u.get("output_tokens") or 0),
+        "cache_creation_tokens": int(u.get("cache_creation_tokens") or 0),
+        "cache_read_tokens": int(u.get("cache_read_tokens") or 0),
+    }
+    cost_list = mp.compute_list_cost(model, **totals)
+    currency = mp.currency_for(model)
+
+    payload = {
+        "model": model,
+        "input_tokens": totals["input_tokens"],
+        "output_tokens": totals["output_tokens"],
+        "cache_creation_tokens": totals["cache_creation_tokens"],
+        "cache_read_tokens": totals["cache_read_tokens"],
+        "duration_ms": u.get("duration_ms"),
+        "cost_list_usd": cost_list,
+        "cost_currency": currency,
+        "prices_version": getattr(mp, "VERSION", None),
+    }
+    return {
+        "ts": datetime.now(timezone.utc).strftime("%Y-%m-%dT%H:%M:%SZ"),
+        "stage": "usage",
+        "label": cycle_id,
+        "detail": payload,
+        "outcome": "ok",
+    }
+
+
+def _default_events_path(slug, shared):
+    base = shared or os.environ.get("LOOP_SHARED_ROOT") \
+        or os.path.expanduser("~/.shared/roll")
+    return os.path.join(base, "loop", "events-%s.ndjson" % slug)
+
+
+def main(argv=None):
+    ap = argparse.ArgumentParser(description="emit one kimi usage event")
+    ap.add_argument("--cwd", help="cycle worktree path (authoritative match)")
+    ap.add_argument("--cycle", help="cycle id (label + dir-name fallback)")
+    ap.add_argument("--slug", help="project slug (events filename)")
+    ap.add_argument("--shared", help="shared root (for default events path)")
+    ap.add_argument("--events", help="explicit events file path (preferred)")
+    ap.add_argument("--base-dir", help="kimi sessions root override (tests)")
+    args = ap.parse_args(argv)
+
+    event = build_event(
+        cwd=args.cwd, cycle_id=args.cycle, slug=args.slug, base_dir=args.base_dir
+    )
+    if event is None:
+        return 0  # nothing recoverable — write nothing (n/a, not fake zero)
+
+    evfile = args.events or _default_events_path(args.slug, args.shared)
+    try:
+        os.makedirs(os.path.dirname(evfile), exist_ok=True)
+        with open(evfile, "a") as f:
+            f.write(json.dumps(event) + "\n")
+    except OSError as e:
+        print("[kimi_emit] failed to write %s: %s" % (evfile, e), file=sys.stderr)
+        return 1
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@seanyao/roll",
-  "version": "2026.601.2",
+  "version": "2026.601.3",
   "description": "Roll — Roll out features with AI agents",
   "scripts": {
     "test": "bash tests/run.sh"