@ranger1/dx 0.1.33 → 0.1.35

package/@opencode/agents/claude-reviewer.md

@@ -14,25 +14,26 @@ tools:
 
 - `PR #<number>`
 - `round: <number>`
+- `runId: <string>`（必须透传，禁止自行生成）
 - `contextFile: <filename>`
 
 ## 输出（强制）
 
 只输出一行：
 
-`reviewFile: <filename>`
+`reviewFile: ./.cache/<file>.md`
 
 
 ## 规则
 
 - 默认已在 PR head 分支（可直接读工作区代码）
 - 可用 `git`/`gh` 只读命令获取 diff/上下文
-- 写入 reviewFile：`review-CLD-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
+- 写入 reviewFile：`./.cache/review-CLD-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
 - findings id 必须以 `CLD-` 开头
 
 ## Cache 约定（强制）
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
+
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ## reviewFile 格式（强制）
 

package/@opencode/agents/codex-reviewer.md

@@ -15,26 +15,26 @@ tools:
 
 - `PR #<number>`
 - `round: <number>`
+- `runId: <string>`（必须透传，禁止自行生成）
 - `contextFile: <filename>`
 
 ## 输出（强制）
 
 只输出一行：
 
-`reviewFile: <filename>`
+`reviewFile: ./.cache/<file>.md`
 
 
 ## 规则
 
 - 默认已在 PR head 分支（可直接读工作区代码）
 - 可用 `git`/`gh` 只读命令获取 diff/上下文
-- 写入 reviewFile：`review-CDX-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
+- 写入 reviewFile：`./.cache/review-CDX-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
 - findings id 必须以 `CDX-` 开头
 
 ## Cache 约定（强制）
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
 
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ## reviewFile 格式（强制）
 

package/@opencode/agents/gemini-reviewer.md

@@ -14,26 +14,26 @@ tools:
 
 - `PR #<number>`
 - `round: <number>`
+- `runId: <string>`（必须透传，禁止自行生成）
 - `contextFile: <filename>`
 
 ## 输出（强制）
 
 只输出一行：
 
-`reviewFile: <filename>`
+`reviewFile: ./.cache/<file>.md`
 
 
 ## 规则
 
 - 默认已在 PR head 分支（可直接读工作区代码）
 - 可用 `git`/`gh` 只读命令获取 diff/上下文
-- 写入 reviewFile：`review-GMN-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
+- 写入 reviewFile：`./.cache/review-GMN-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
 - findings id 必须以 `GMN-` 开头
 
 ## Cache 约定（强制）
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
 
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ## reviewFile 格式（强制）
 

package/@opencode/agents/pr-context.md

@@ -20,11 +20,11 @@ tools:
 
 ## 输出（强制）
 
-脚本会写入 `~/.opencode/cache/`，stdout 只输出单一 JSON（可 `JSON.parse()`）。
+脚本会写入项目内 `./.cache/`，stdout 只输出单一 JSON（可 `JSON.parse()`）。
 
 ## Cache 约定（强制）
 
-
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ## 调用脚本（强制）
 

package/@opencode/agents/pr-fix.md

@@ -26,13 +26,14 @@ tools:
 ## 前置条件
 
 ### Cache 约定（强制）
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
+
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ### 必需输入
 
 - **PR 编号**：调用者必须在 prompt 中明确提供（如：`请修复 PR #123`）
-- **fixFile**：调用者必须在 prompt 中提供问题清单文件名（basename）（Structured Handoff）
+- **runId**：调用者必须在 prompt 中提供（必须透传，禁止自行生成）
+- **fixFile**：调用者必须在 prompt 中提供问题清单文件路径（repo 相对路径，例：`./.cache/fix-...md`）（Structured Handoff）
 
 ### 失败快速退出
 
@@ -126,11 +127,11 @@ Round: 2
 
 ## 输出（强制）
 
-写入：`fix-report-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
+写入：`./.cache/fix-report-pr<PR_NUMBER>-r<ROUND>-<RUN_ID>.md`
 
 最终只输出一行：
 
-`fixReportFile: <filename>`
+`fixReportFile: ./.cache/<file>.md`
 
 
 ## fixReportFile 内容格式（强制）
@@ -172,4 +173,4 @@ Rejected: <n>
 ## 输出有效性保证
 
 - fixReportFile 必须成功写入
-- stdout 只能输出一行 `fixReportFile: <filename>`
+- stdout 只能输出一行 `fixReportFile: ./.cache/<file>.md`

package/@opencode/agents/pr-precheck.md

@@ -10,9 +10,8 @@ tools:
 # PR Precheck
 
 ## Cache 约定（强制）
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
 
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ## 输入（prompt 必须包含）
 

package/@opencode/agents/pr-review-aggregate.md

@@ -10,8 +10,8 @@ tools:
 # PR Review Aggregator
 
 ## Cache 约定（强制）
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
+
+- 缓存目录固定为 `./.cache/`；交接一律传 `./.cache/<file>`（repo 相对路径），禁止 basename-only（如 `foo.md`）。
 
 ## 输入（两种模式）
 
@@ -20,15 +20,15 @@ tools:
 - `PR #<number>`
 - `round: <number>`
 - `runId: <string>`
-- `contextFile: <filename>`
-- `reviewFile: <filename>`（三行，分别对应 CDX/CLD/GMN）
+- `contextFile: <path>`（例如：`./.cache/pr-context-...md`）
+- `reviewFile: <path>`（三行，分别对应 CDX/CLD/GMN，例如：`./.cache/review-...md`）
 
 ### 模式 B：发布修复评论（基于 fixReportFile）
 
 - `PR #<number>`
 - `round: <number>`
 - `runId: <string>`
-- `fixReportFile: <filename>`
+- `fixReportFile: <path>`（例如：`./.cache/fix-report-...md`）
 
 示例：
 
@@ -36,10 +36,10 @@ tools:
 PR #123
 round: 1
 runId: abcdef123456
-contextFile: pr-context-pr123-r1-abcdef123456.md
-reviewFile: review-CDX-pr123-r1-abcdef123456.md
-reviewFile: review-CLD-pr123-r1-abcdef123456.md
-reviewFile: review-GMN-pr123-r1-abcdef123456.md
+  contextFile: ./.cache/pr-context-pr123-r1-abcdef123456.md
+  reviewFile: ./.cache/review-CDX-pr123-r1-abcdef123456.md
+  reviewFile: ./.cache/review-CLD-pr123-r1-abcdef123456.md
+  reviewFile: ./.cache/review-GMN-pr123-r1-abcdef123456.md
 ```
 
 ## 执行方式（强制）

package/@opencode/agents/pr_context.py

@@ -2,7 +2,7 @@
 # PR context builder (deterministic).
 # - Reads PR metadata + recent comments via gh
 # - Reads changed files via git diff (no patch)
-# - Writes Markdown context file to ~/.opencode/cache/
+# - Writes Markdown context file to project cache: ./.cache/
 # - Prints exactly one JSON object to stdout
 
 import argparse
@@ -16,7 +16,38 @@ from urllib.parse import urlparse
 from pathlib import Path
 
 
-CACHE_DIR = Path.home() / ".opencode" / "cache"
+def _repo_root():
+    # Prefer git top-level so the cache follows the current repo.
+    try:
+        p = subprocess.run(
+            ["git", "rev-parse", "--show-toplevel"],
+            stdout=subprocess.PIPE,
+            stderr=subprocess.DEVNULL,
+            text=True,
+        )
+        out = (p.stdout or "").strip()
+        if p.returncode == 0 and out:
+            return Path(out)
+    except Exception:
+        pass
+    return Path.cwd()
+
+
+def _cache_dir(repo_root):
+    return (repo_root / ".cache").resolve()
+
+
+def _repo_relpath(repo_root, p):
+    try:
+        rel = p.resolve().relative_to(repo_root.resolve())
+        return "./" + rel.as_posix()
+    except Exception:
+        # Fallback to basename-only.
+        return os.path.basename(str(p))
+
+
+REPO_ROOT = _repo_root()
+CACHE_DIR = _cache_dir(REPO_ROOT)
 MARKER_SUBSTR = "<!-- pr-review-loop-marker"
 
 
@@ -252,7 +283,8 @@ def main(argv):
             "repo": {"nameWithOwner": owner_repo},
             "headOid": head_oid,
             "existingMarkerCount": marker_count,
-            "contextFile": context_file,
+            # Handoff should be repo-relative path so downstream agents can read it directly.
+            "contextFile": _repo_relpath(REPO_ROOT, context_path),
         }
     )
     return 0

package/@opencode/agents/pr_precheck.py

@@ -8,12 +8,14 @@
 # - If mergeable == CONFLICTING: return {"error":"PR_MERGE_CONFLICTS_UNRESOLVED"}
 # - Run dx cache clear
 # - Run dx lint and dx build all concurrently
-# - On failure, write fixFile to ~/.opencode/cache/ and return {"ok":false,"fixFile":"..."}
+# - On failure, write fixFile to project cache: ./.cache/
+#   and return {"ok":false,"fixFile":"./.cache/..."}
 # - On success, return {"ok":true}
 #
 # Stdout contract: print exactly one JSON object and nothing else.
 
 import json
+import os
 import re
 import secrets
 import subprocess
@@ -92,6 +94,34 @@ def _detect_git_remote_host():
     return None
 
 
+def repo_root():
+    try:
+        p = subprocess.run(
+            ["git", "rev-parse", "--show-toplevel"],
+            stdout=subprocess.PIPE,
+            stderr=subprocess.DEVNULL,
+            text=True,
+        )
+        out = (p.stdout or "").strip()
+        if p.returncode == 0 and out:
+            return Path(out)
+    except Exception:
+        pass
+    return Path.cwd()
+
+
+def cache_dir(repo_root_path):
+    return (repo_root_path / ".cache").resolve()
+
+
+def repo_relpath(repo_root_path, p):
+    try:
+        rel = p.resolve().relative_to(repo_root_path.resolve())
+        return "./" + rel.as_posix()
+    except Exception:
+        return str(p)
+
+
 def tail_text(path, max_lines=200, max_chars=12000):
     try:
         data = Path(path).read_text(errors="replace")
@@ -214,7 +244,8 @@ def main():
         return 1
 
     run_id = secrets.token_hex(4)
-    cache = Path.home() / ".opencode" / "cache"
+    root = repo_root()
+    cache = cache_dir(root)
     cache.mkdir(parents=True, exist_ok=True)
     
     cache_clear_log = cache / f"precheck-pr{pr}-{run_id}-cache-clear.log"
@@ -227,9 +258,9 @@ def main():
         "headRefName": head,
         "baseRefName": base,
         "mergeable": mergeable,
-        "cacheClearLog": str(cache_clear_log),
-        "lintLog": str(lint_log),
-        "buildLog": str(build_log),
+        "cacheClearLog": repo_relpath(root, cache_clear_log),
+        "lintLog": repo_relpath(root, lint_log),
+        "buildLog": repo_relpath(root, build_log),
     }, indent=2) + "\n")
 
     cache_rc = run(["dx", "cache", "clear"], stdout_path=str(cache_clear_log), stderr_path=str(cache_clear_log))
@@ -245,10 +276,10 @@ def main():
             "line": None,
             "title": "dx cache clear failed",
             "description": log_tail,
-            "suggestion": f"Open log: {cache_clear_log}",
+            "suggestion": f"Open log: {repo_relpath(root, cache_clear_log)}",
         }]
         write_fixfile(str(fix_path), issues)
-        print(json.dumps({"ok": False, "fixFile": fix_file}))
+        print(json.dumps({"ok": False, "fixFile": repo_relpath(root, fix_path)}))
         return 1
 
     import threading
@@ -285,7 +316,7 @@ def main():
             "line": line,
             "title": "dx lint failed",
             "description": log_tail,
-            "suggestion": f"Open log: {lint_log}",
+            "suggestion": f"Open log: {repo_relpath(root, lint_log)}",
         })
         i += 1
     if results.get("build", 1) != 0:
@@ -299,11 +330,11 @@ def main():
             "line": line,
             "title": "dx build all failed",
             "description": log_tail,
-            "suggestion": f"Open log: {build_log}",
+            "suggestion": f"Open log: {repo_relpath(root, build_log)}",
         })
 
     write_fixfile(str(fix_path), issues)
-    print(json.dumps({"ok": False, "fixFile": fix_file}))
+    print(json.dumps({"ok": False, "fixFile": repo_relpath(root, fix_path)}))
     return 1
 
 

package/@opencode/agents/pr_review_aggregate.py

@@ -2,12 +2,12 @@
 # Deterministic PR review aggregation (script owns all rules).
 #
 # Workflow:
-# - Mode A: read contextFile + reviewFile(s) from ~/.opencode/cache/, parse findings, merge duplicates,
+# - Mode A: read contextFile + reviewFile(s) from project cache: ./.cache/, parse findings, merge duplicates,
 #   post a single PR comment, and optionally generate a fixFile for pr-fix.
 # - Mode B: read fixReportFile from cache and post it as a PR comment.
 #
 # Input rules:
-# - Callers pass only basenames (no paths). This script reads/writes under ~/.opencode/cache/.
+# - Callers should pass repo-relative paths (e.g. ./.cache/foo.md). For backward-compat, basenames are also accepted.
 # - Duplicate groups come from LLM but are passed as an argument (NOT written to disk).
 #   - Prefer: --duplicate-groups-b64 <base64(json)>
 #   - Also supported: --duplicate-groups-json '<json>'
@@ -20,7 +20,7 @@
 #
 # PR comment rules:
 # - Every comment must include marker: <!-- pr-review-loop-marker -->
-# - Comment body must NOT contain local filesystem paths (this script scrubs ~/.opencode/cache and $HOME).
+# - Comment body must NOT contain local filesystem paths (this script scrubs cache paths, $HOME, and repo absolute paths).
 #
 # fixFile rules:
 # - fixFile includes ONLY P0/P1/P2 findings.
@@ -38,7 +38,76 @@ from pathlib import Path
 
 
 MARKER = "<!-- pr-review-loop-marker -->"
-CACHE_DIR = Path.home() / ".opencode" / "cache"
+
+
+def _repo_root():
+    try:
+        p = subprocess.run(
+            ["git", "rev-parse", "--show-toplevel"],
+            stdout=subprocess.PIPE,
+            stderr=subprocess.DEVNULL,
+            text=True,
+        )
+        out = (p.stdout or "").strip()
+        if p.returncode == 0 and out:
+            return Path(out)
+    except Exception:
+        pass
+    return Path.cwd()
+
+
+def _cache_dir(repo_root):
+    return (repo_root / ".cache").resolve()
+
+
+def _is_safe_relpath(p):
+    if p.is_absolute():
+        return False
+    if any(part in ("..",) for part in p.parts):
+        return False
+    return True
+
+
+def _resolve_ref(repo_root, cache_dir, ref):
+    if not ref:
+        return None
+    s = str(ref).strip()
+    if not s:
+        return None
+
+    # If caller already passes a repo-relative path like ./.cache/foo.md
+    looks_like_path = ("/" in s) or ("\\" in s) or s.startswith(".")
+    if looks_like_path:
+        p = Path(s)
+        if p.is_absolute():
+            # Only allow absolute paths under cache_dir.
+            try:
+                p2 = p.resolve()
+                p2.relative_to(cache_dir.resolve())
+                return p2
+            except Exception:
+                return None
+        if not _is_safe_relpath(p):
+            return None
+        return (repo_root / p).resolve()
+
+    # Backward-compat: accept basename-only.
+    b = _safe_basename(s)
+    if not b:
+        return None
+    return (cache_dir / b).resolve()
+
+
+def _repo_relpath(repo_root, p):
+    try:
+        rel = p.resolve().relative_to(repo_root.resolve())
+        return "./" + rel.as_posix()
+    except Exception:
+        return os.path.basename(str(p))
+
+
+REPO_ROOT = _repo_root()
+CACHE_DIR = _cache_dir(REPO_ROOT)
 
 
 def _json_out(obj):
@@ -57,14 +126,19 @@ def _safe_basename(name):
     return base
 
 
-def _read_cache_text(basename):
-    p = CACHE_DIR / basename
+def _read_cache_text(ref):
+    p = _resolve_ref(REPO_ROOT, CACHE_DIR, ref)
+    if not p:
+        raise FileNotFoundError("INVALID_CACHE_REF")
     return p.read_text(encoding="utf-8", errors="replace")
 
 
-def _write_cache_text(basename, content):
+def _write_cache_text(ref, content):
+    p = _resolve_ref(REPO_ROOT, CACHE_DIR, ref)
+    if not p:
+        raise ValueError("INVALID_CACHE_REF")
     CACHE_DIR.mkdir(parents=True, exist_ok=True)
-    p = CACHE_DIR / basename
+    p.parent.mkdir(parents=True, exist_ok=True)
     p.write_text(content, encoding="utf-8", newline="\n")
 
 
@@ -88,13 +162,21 @@ def _sanitize_for_comment(text):
         text = str(text)
 
     home = str(Path.home())
-    cache_abs = str(CACHE_DIR)
+    cache_abs = str(CACHE_DIR.resolve())
+    repo_abs = str(REPO_ROOT.resolve())
 
+    # Backward-compat scrub.
     text = text.replace("~/.opencode/cache/", "[cache]/")
-    text = text.replace(cache_abs + "/", "[cache]/")
     if home:
         text = text.replace(home + "/.opencode/cache/", "[cache]/")
 
+    # New cache scrub.
+    text = text.replace(cache_abs + "/", "[cache]/")
+
+    # Avoid leaking absolute local repo paths.
+    if repo_abs:
+        text = text.replace(repo_abs + "/", "")
+
     return text
 
 
@@ -236,8 +318,14 @@ def _counts(findings):
     return c
 
 
-def _post_pr_comment(pr_number, body_basename):
-    body_path = str(CACHE_DIR / body_basename)
+def _post_pr_comment(pr_number, body_ref):
+    if isinstance(body_ref, Path):
+        p = body_ref
+    else:
+        p = _resolve_ref(REPO_ROOT, CACHE_DIR, body_ref)
+    if not p:
+        return False
+    body_path = str(p)
     rc = subprocess.run(
         ["gh", "pr", "comment", str(pr_number), "--body-file", body_path],
         stdout=subprocess.DEVNULL,
@@ -334,20 +422,25 @@ def main(argv):
     round_num = args.round
     run_id = str(args.run_id)
 
-    fix_report_file = _safe_basename(args.fix_report_file) if args.fix_report_file else None
-    context_file = _safe_basename(args.context_file) if args.context_file else None
+    fix_report_file = (args.fix_report_file or "").strip() or None
+    context_file = (args.context_file or "").strip() or None
     review_files = []
     for rf in args.review_file or []:
-        b = _safe_basename(rf)
-        if b:
-            review_files.append(b)
+        s = (rf or "").strip()
+        if s:
+            review_files.append(s)
 
     if fix_report_file:
+        fix_p = _resolve_ref(REPO_ROOT, CACHE_DIR, fix_report_file)
+        if not fix_p or not fix_p.exists():
+            _json_out({"error": "FIX_REPORT_FILE_NOT_FOUND"})
+            return 1
         fix_md = _read_cache_text(fix_report_file)
         body = _render_mode_b_comment(pr_number, round_num, run_id, fix_md)
-        body_file = f"review-aggregate-fix-comment-pr{pr_number}-r{round_num}-{run_id}.md"
-        _write_cache_text(body_file, body)
-        if not _post_pr_comment(pr_number, body_file):
+        body_basename = f"review-aggregate-fix-comment-pr{pr_number}-r{round_num}-{run_id}.md"
+        body_ref = _repo_relpath(REPO_ROOT, CACHE_DIR / body_basename)
+        _write_cache_text(body_ref, body)
+        if not _post_pr_comment(pr_number, body_ref):
             _json_out({"error": "GH_PR_COMMENT_FAILED"})
             return 1
         _json_out({"ok": True})
@@ -360,6 +453,21 @@ def main(argv):
         _json_out({"error": "MISSING_REVIEW_FILES"})
         return 1
 
+    ctx_p = _resolve_ref(REPO_ROOT, CACHE_DIR, context_file)
+    if not ctx_p or not ctx_p.exists():
+        _json_out({"error": "CONTEXT_FILE_NOT_FOUND"})
+        return 1
+
+    valid_review_files = []
+    for rf in review_files:
+        p = _resolve_ref(REPO_ROOT, CACHE_DIR, rf)
+        if p and p.exists():
+            valid_review_files.append(rf)
+    review_files = valid_review_files
+    if not review_files:
+        _json_out({"error": "REVIEW_FILES_NOT_FOUND"})
+        return 1
+
     raw_reviews = []
     all_findings = []
     for rf in review_files:
@@ -377,9 +485,10 @@ def main(argv):
     stop = len(must_fix) == 0
 
     body = _render_mode_a_comment(pr_number, round_num, run_id, counts, must_fix, merged_map, raw_reviews)
-    body_file = f"review-aggregate-comment-pr{pr_number}-r{round_num}-{run_id}.md"
-    _write_cache_text(body_file, body)
-    if not _post_pr_comment(pr_number, body_file):
+    body_basename = f"review-aggregate-comment-pr{pr_number}-r{round_num}-{run_id}.md"
+    body_ref = _repo_relpath(REPO_ROOT, CACHE_DIR / body_basename)
+    _write_cache_text(body_ref, body)
+    if not _post_pr_comment(pr_number, body_ref):
         _json_out({"error": "GH_PR_COMMENT_FAILED"})
         return 1
 
@@ -415,8 +524,9 @@ def main(argv):
         lines.append(f"  description: {desc}")
         lines.append(f"  suggestion: {sugg}")
 
-    _write_cache_text(fix_file, "\n".join(lines) + "\n")
-    _json_out({"stop": False, "fixFile": fix_file})
+    fix_ref = _repo_relpath(REPO_ROOT, CACHE_DIR / fix_file)
+    _write_cache_text(fix_ref, "\n".join(lines) + "\n")
+    _json_out({"stop": False, "fixFile": fix_ref})
     return 0
 
 

package/@opencode/commands/pr-review-loop.md

@@ -12,8 +12,8 @@ agent: sisyphus
 
 ## Cache 约定（强制）
 
-- 本流程所有中间文件都存放在 `~/.opencode/cache/`
-- agent/命令之间仅传递文件名（basename），不传目录
+- 本流程所有中间文件都存放在项目内：`./.cache/`
+- agent/命令之间传递**repo 相对路径**（例如：`./.cache/pr-context-...md`），不要只传 basename
 
 ## 固定 subagent_type（直接用 Task 调用，不要反复确认）
 
@@ -56,17 +56,18 @@ agent: sisyphus
 - 每个 reviewer prompt 必须包含：
   - `PR #{{PR_NUMBER}}`
   - `round: <ROUND>`
-  - `contextFile: <filename>`（来自 Step 1 的输出）
+  - `runId: <RUN_ID>`（来自 Step 1 的输出，必须透传，禁止自行生成）
+  - `contextFile: ./.cache/<file>.md`（来自 Step 1 的输出）
 - reviewer 默认读 `contextFile`；必要时允许用 `git/gh` 只读命令拿 diff
 - 忽略问题：1.格式化代码引起的噪音 2.已经lint检查以外的格式问题
 - 特别关注: 逻辑、安全、性能、可维护性
 - 同时要注意 pr 前面轮次的 修复和讨论，对于已经拒绝、已修复的问题不要反复的提出
 - 同时也要注意fix的过程中有没有引入新的问题。
-- 每个 reviewer 输出：`reviewFile: <filename>`（Markdown）
+- 每个 reviewer 输出：`reviewFile: ./.cache/<file>.md`（Markdown）
 
 3. Task: `pr-review-aggregate`
 
-- prompt 必须包含：`PR #{{PR_NUMBER}}`、`round: <ROUND>`、`runId: <RUN_ID>`、`contextFile: <filename>`、三条 `reviewFile: <filename>`
+- prompt 必须包含：`PR #{{PR_NUMBER}}`、`round: <ROUND>`、`runId: <RUN_ID>`、`contextFile: ./.cache/<file>.md`、三条 `reviewFile: ./.cache/<file>.md`
 - 输出：`{"stop":true}` 或 `{"stop":false,"fixFile":"..."}`
 - 若 `stop=true`：本轮结束并退出循环
 
@@ -75,14 +76,15 @@ agent: sisyphus
 - prompt 必须包含：
   - `PR #{{PR_NUMBER}}`
   - `round: <ROUND>`
-  - `fixFile: <filename>`
+  - `runId: <RUN_ID>`（来自 Step 1 的输出，必须透传，禁止自行生成）
+  - `fixFile: ./.cache/<file>.md`
 - 约定：`pr-fix` 对每个 findingId 单独 commit + push（一个 findingId 一个 commit），结束后再 `git push` 兜底
 
-- pr-fix 输出：`fixReportFile: <filename>`（Markdown）
+- pr-fix 输出：`fixReportFile: ./.cache/<file>.md`（Markdown）
 
 5. Task: `pr-review-aggregate`（发布修复评论）
 
-- prompt 必须包含：`PR #{{PR_NUMBER}}`、`round: <ROUND>`、`runId: <RUN_ID>`、`fixReportFile: <filename>`
+- prompt 必须包含：`PR #{{PR_NUMBER}}`、`round: <ROUND>`、`runId: <RUN_ID>`、`fixReportFile: ./.cache/<file>.md`
 - 输出：`{"ok":true}`
 
 6. 下一轮

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@ranger1/dx",
-  "version": "0.1.33",
+  "version": "0.1.35",
   "type": "module",
   "license": "MIT",
   "repository": {