forgexa-cli 1.5.2__tar.gz → 1.6.1__tar.gz

{forgexa_cli-1.5.2 → forgexa_cli-1.6.1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: forgexa-cli
-Version: 1.5.2
+Version: 1.6.1
 Summary: Forgexa CLI — command-line client and AI agent runtime for the Forgexa platform
 Author-email: Jason Sun <dev.winds@gmail.com>
 License: MIT

{forgexa_cli-1.5.2 → forgexa_cli-1.6.1}/forgexa_cli/__init__.py

@@ -1,2 +1,2 @@
 """forgexa-cli — Forgexa command-line client."""
-__version__ = "1.5.2"
+__version__ = "1.6.1"

{forgexa_cli-1.5.2 → forgexa_cli-1.6.1}/forgexa_cli/daemon.py

@@ -307,7 +307,7 @@ except (ImportError, ModuleNotFoundError):
 # DAEMON_VERSION is the protocol/logic version of the daemon code.
 # Kept in sync with pyproject.toml version via bump-version.sh.
 # CLIENT_TYPE identifies which packaging/distribution this daemon runs in.
-DAEMON_VERSION = "1.5.2"
+DAEMON_VERSION = "1.6.1"
 
 
 def _detect_client_type() -> str:
@@ -473,9 +473,12 @@ def get_os_info() -> str:
             return f"{distro} {machine}"
         return f"Linux {platform.release().split('-')[0]} {machine}"
     elif system == "Windows":
-        # e.g. "Windows 10.0 AMD64"
-        win_ver = platform.version().split('.')[0:2]
-        return f"Windows {'.'.join(win_ver)} {machine}"
+        # platform.version() returns the NT kernel version (10.0.x) for BOTH
+        # Windows 10 and Windows 11, so it cannot distinguish them.
+        # platform.win32_ver()[0] returns the marketing release: "10" or "11".
+        win_release = platform.win32_ver()[0] or platform.version().split('.')[0]
+        build = platform.version()  # e.g. "10.0.26200"
+        return f"Windows {win_release} ({build}) {machine}"
     else:
         return f"{system} {platform.release()} {machine}"
 
@@ -869,12 +872,19 @@ class WorkspaceManager:
             branch_name = f"feature/{workspace_key}"
 
         # Determine whether this node is the "first" in a new graph that needs
-        # a fresh base from the default branch.  Analysis nodes always start
-        # fresh unless using "refine" mode which preserves the existing branch.
-        analysis_mode = task.input_data.get("analysis_mode", "fresh")
+        # a fresh base from the default branch.  Only analysis nodes with
+        # explicit analysis_mode="fresh" should start fresh.  All other modes
+        # (including the default when no mode is specified) preserve the
+        # existing branch to avoid destroying prior implementation commits.
+        #
+        # SAFETY: Defaulting to "refine" ensures that any code path which
+        # forgets to set analysis_mode will safely preserve the branch.
+        # Only the initial requirement analysis endpoint (requirements.py)
+        # explicitly passes "fresh" when creating a brand-new requirement branch.
+        analysis_mode = task.input_data.get("analysis_mode", "refine")
         is_fresh_start = (
             task.node_type == "analysis"
-            and analysis_mode != "refine"
+            and analysis_mode == "fresh"
         )
 
         if repo_url:
@@ -977,11 +987,52 @@ class WorkspaceManager:
                 logger.warning("Broken worktree detected at %s — removing and recreating", ws_path)
                 await self._remove_broken_worktree(main_repo, ws_path, workspace_key)
             else:
-                # Healthy worktree — fetch and optionally reset
+                # Healthy worktree — fetch and optionally reset.
+                # The _main repo uses --single-branch, so `git fetch origin`
+                # only fetches the default branch.  Explicitly fetch the
+                # feature branch with a full refspec so that
+                # origin/{branch_name} is available for checkout/reset.
                 try:
                     await self._git("fetch", "origin", cwd=ws_path, project_key=project_key)
                 except RuntimeError:
                     pass
+                try:
+                    await self._git(
+                        "fetch", "origin",
+                        f"{branch_name}:refs/remotes/origin/{branch_name}",
+                        cwd=ws_path, project_key=project_key,
+                    )
+                except RuntimeError:
+                    pass  # Branch may not exist on remote yet
+
+                if fresh_start:
+                    # Safety check: if the branch already exists on remote with
+                    # commits beyond the default branch, do NOT reset to
+                    # origin/default_branch — that would destroy prior work.
+                    # This guards against accidental branch destruction when
+                    # analysis_mode is incorrectly set to "fresh" for
+                    # verification or iterative nodes.
+                    branch_has_remote_commits = False
+                    try:
+                        result = await self._git(
+                            "rev-list", "--count",
+                            f"origin/{default_branch}..origin/{branch_name}",
+                            cwd=ws_path,
+                        )
+                        ahead_count = int(result.strip()) if result.strip() else 0
+                        branch_has_remote_commits = ahead_count > 0
+                    except (RuntimeError, ValueError):
+                        pass  # Branch may not exist on remote yet
+
+                    if branch_has_remote_commits:
+                        logger.warning(
+                            "Fresh start requested for %s but remote branch has %d commit(s) "
+                            "ahead of %s — switching to safe sync instead of resetting to "
+                            "avoid destroying prior work",
+                            branch_name, ahead_count, default_branch,
+                        )
+                        # Override: fall through to the non-fresh sync path
+                        fresh_start = False
 
                 if fresh_start:
                     logger.info("Fresh start: resetting %s to origin/%s", ws_path, default_branch)
@@ -1036,7 +1087,9 @@ class WorkspaceManager:
                                     await asyncio.sleep(2 * (_sync_attempt + 1))
                                     try:
                                         await self._git(
-                                            "fetch", "origin", cwd=ws_path,
+                                            "fetch", "origin",
+                                            f"{branch_name}:refs/remotes/origin/{branch_name}",
+                                            cwd=ws_path,
                                             project_key=project_key,
                                         )
                                     except RuntimeError:
@@ -1061,7 +1114,9 @@ class WorkspaceManager:
                                 await asyncio.sleep(2 * (_sync_attempt + 1))
                                 try:
                                     await self._git(
-                                        "fetch", "origin", cwd=ws_path,
+                                        "fetch", "origin",
+                                        f"{branch_name}:refs/remotes/origin/{branch_name}",
+                                        cwd=ws_path,
                                         project_key=project_key,
                                     )
                                 except RuntimeError:
@@ -1096,6 +1151,19 @@ class WorkspaceManager:
         else:
             await self._git("fetch", "--all", cwd=main_repo, timeout=300, project_key=project_key)
 
+        # --single-branch clone only fetches the default branch.
+        # Explicitly fetch the feature branch so origin/{branch_name}
+        # is available for worktree creation and checkout.
+        if not fresh_start:
+            try:
+                await self._git(
+                    "fetch", "origin",
+                    f"{branch_name}:refs/remotes/origin/{branch_name}",
+                    cwd=main_repo, timeout=60, project_key=project_key,
+                )
+            except RuntimeError:
+                pass  # Branch may not exist on remote yet (first analysis)
+
         # Prune stale worktree references (e.g. directories deleted externally
         # when simulating cross-runtime or after disk cleanup).  Without this,
         # `git worktree add` refuses to create a branch that is "already checked out"
@@ -1144,7 +1212,11 @@ class WorkspaceManager:
                 )
                 await asyncio.sleep(2 * (_check_attempt + 1))
                 try:
-                    await self._git("fetch", "--all", cwd=main_repo, timeout=60, project_key=project_key)
+                    await self._git(
+                        "fetch", "origin",
+                        f"{branch_name}:refs/remotes/origin/{branch_name}",
+                        cwd=main_repo, timeout=60, project_key=project_key,
+                    )
                 except RuntimeError:
                     pass
 
@@ -1227,6 +1299,18 @@ class WorkspaceManager:
 
         If a project SSH key is registered, remote-touching commands
         (clone, fetch, pull, push) will use GIT_SSH_COMMAND.
+
+        Process-group isolation: git is started in its own session so that
+        SIGKILL on timeout propagates to all of git's children (especially
+        the ssh subprocess that git forks for remote operations).  Without
+        this, killing git leaves orphaned ssh processes that hold the stderr
+        pipe open — preventing asyncio from detecting EOF and causing the
+        parent to hang until the GIT_CLONE_TIMEOUT fires.
+
+        SSH keepalives: GIT_SSH_COMMAND now includes ServerAliveInterval=30
+        and ServerAliveCountMax=3 so that a stalled TCP connection (server
+        accepts but never sends the git protocol banner) is detected within
+        ~90 s rather than waiting forever.
         """
         env = None
         git_prefix_args: list[str] = []
@@ -1255,6 +1339,14 @@ class WorkspaceManager:
                             f" -o StrictHostKeyChecking=accept-new"
                             f" -o UserKnownHostsFile=/dev/null"
                             f" -o IdentitiesOnly=yes"
+                            # Detect a stalled TCP connection (server accepts but
+                            # never sends the git protocol banner).  After 30 s of
+                            # silence the client sends a keepalive; after 3 missed
+                            # responses (≤ 90 s total) SSH exits, closes the pipes,
+                            # and lets asyncio's communicate() return.
+                            f" -o ConnectTimeout=30"
+                            f" -o ServerAliveInterval=30"
+                            f" -o ServerAliveCountMax=3"
                         ),
                     }
                 except Exception:
@@ -1274,18 +1366,53 @@ class WorkspaceManager:
                 if git_prefix_args:
                     env = {**(env or os.environ), "GIT_TERMINAL_PROMPT": "0"}
 
+        # start_new_session=True puts git in its own process group.
+        # On timeout we send SIGKILL to the entire group, which includes
+        # any ssh/gpg/credential-helper children that git forked — preventing
+        # orphaned processes from keeping pipes alive.
+        # Windows note: start_new_session creates a new console process group;
+        # we use taskkill /T there instead of killpg.
         proc = await asyncio.create_subprocess_exec(
             "git", *git_prefix_args, *args,
             stdout=asyncio.subprocess.PIPE,
             stderr=asyncio.subprocess.PIPE,
             cwd=str(cwd) if cwd else None,
             env=env,
+            start_new_session=True,
         )
         try:
             stdout, stderr = await asyncio.wait_for(proc.communicate(), timeout=timeout)
         except (TimeoutError, asyncio.TimeoutError):
-            proc.kill()
-            await proc.wait()
+            # Kill the entire process group so that SSH (and any other
+            # child processes git may have spawned) are also terminated.
+            # A plain proc.kill() only kills the direct child (git); the
+            # ssh grandchild becomes orphaned, keeps the stderr pipe open,
+            # and proc.communicate() can never return EOF.
+            try:
+                if sys.platform != "win32":
+                    import signal as _signal
+                    try:
+                        os.killpg(os.getpgid(proc.pid), _signal.SIGKILL)
+                    except (ProcessLookupError, PermissionError, OSError):
+                        pass
+                else:
+                    # Windows: taskkill /F /T kills the process tree
+                    import subprocess as _subprocess
+                    _subprocess.run(
+                        ["taskkill", "/F", "/T", "/PID", str(proc.pid)],
+                        capture_output=True,
+                    )
+            except Exception:
+                pass
+            finally:
+                try:
+                    proc.kill()
+                except Exception:
+                    pass
+            try:
+                await asyncio.wait_for(proc.wait(), timeout=5.0)
+            except Exception:
+                pass
             raise RuntimeError(f"git {' '.join(args)} timed out after {timeout}s")
         finally:
             # Clean up temp SSH key file if created
@@ -1353,11 +1480,27 @@ class ProcessManager:
 
     @staticmethod
     def _extract_output_signals(text: str) -> dict[str, Any]:
-        """Parse stdout/stderr-like streams for success and failure signals."""
+        """Parse stdout/stderr-like streams for success and failure signals.
+
+        Uses structural signal detection — NOT keyword matching on result text.
+
+        The key invariant for Claude stream-json:
+        - A genuine execution always emits at least one 'assistant' event AND
+          has total_input_tokens > 0 in the result event, because the CLI made
+          a real API call.
+        - A pre-call failure (API Error, Connection error, auth failure) exits
+          immediately: no 'assistant' events, and total_input_tokens == 0 in
+          the result event. The result text contains the CLI error message.
+
+        This avoids false positives from agent work output that legitimately
+        contains words like 'rate limit', 'connection error', 'API error', etc.
+        (e.g. implementing a rate-limiter feature, or documenting error codes).
+        """
         has_turn_completed = False
         has_turn_failed = False
         has_result = False
         has_meaningful_content = False
+        has_assistant_events = False  # True once any 'assistant' event is seen
         error_messages: list[str] = []
         json_line_count = 0
 
@@ -1387,18 +1530,31 @@ class ProcessManager:
                 elif isinstance(err, str):
                     error_messages.append(err)
             elif ev_type == "result":
+                result_text = str(data.get("result", "") or "")
                 if data.get("is_error"):
-                    err_text = str(data.get("result", "") or data.get("error", "") or "result marked as error")
+                    err_text = result_text or str(data.get("error", "") or "result marked as error")
                     error_messages.append(err_text)
                 else:
-                    has_result = True
-                    has_meaningful_content = True
+                    # Structural check: if no tokens were consumed AND no assistant
+                    # events appeared, the CLI never made an API call. The result
+                    # text is a CLI-level error (e.g. "API Error: Connection error.")
+                    # rather than the agent's actual work output.
+                    tok_in = int(data.get("total_input_tokens", 0) or 0)
+                    tok_out = int(data.get("total_output_tokens", 0) or 0)
+                    no_api_call = (tok_in + tok_out == 0) and not has_assistant_events
+                    if no_api_call and result_text:
+                        error_messages.append(result_text)
+                    else:
+                        has_result = True
+                        has_meaningful_content = True
             elif ev_type == "error":
                 msg = data.get("message", "")
                 if msg:
                     error_messages.append(msg)
+            elif ev_type == "assistant":
+                has_meaningful_content = True
+                has_assistant_events = True
             elif ev_type in (
-                "assistant",
                 "content_block_delta",
                 "message_delta",
                 "step_finish",
@@ -1557,7 +1713,10 @@ class ProcessManager:
         start_time = time.monotonic()
 
         if agent.agent_id == "claude-code":
-            result = await self._run_claude(agent, prompt, workspace_path, timeout, task.task_id, on_chunk)
+            result = await self._run_claude(
+                agent, prompt, workspace_path, timeout, task.task_id, on_chunk,
+                node_type=task.node_type,
+            )
         elif agent.agent_id == "codex":
             result = await self._run_codex(agent, prompt, workspace_path, timeout, task.task_id, on_chunk)
         elif agent.agent_id == "opencode":
@@ -1778,18 +1937,37 @@ class ProcessManager:
 
     async def _run_claude(
         self, agent: DiscoveredAgent, prompt: str, cwd: Path, timeout: int, task_id: str,
-        on_chunk: Any = None,
+        on_chunk: Any = None, node_type: str = "",
     ) -> TaskResult:
         """Run Claude Code CLI in print mode with auto-approve permissions.
 
         Uses stdin pipe for prompt delivery to avoid ARG_MAX limits.
         Uses --dangerously-skip-permissions for autonomous file operations.
+        Uses --max-turns to prevent context window overflow during long sessions.
         """
+        # Determine max turns based on node type.
+        # Analysis/coding nodes need more turns (read many files, write multiple outputs).
+        # Review/testing nodes are typically shorter.
+        # This prevents runaway context accumulation that leads to "Prompt is too long".
+        max_turns_map = {
+            "analysis": 80,
+            "coding": 80,
+            "design": 50,
+            "testing": 60,
+            "review": 40,
+            "fix": 60,
+        }
+        max_turns = int(os.environ.get(
+            "FACTORY_CLAUDE_MAX_TURNS",
+            str(max_turns_map.get(node_type, 60)),
+        ))
+
         cmd = [
             agent.command,
             "-p",
             "--output-format", "stream-json",
             "--verbose",
+            "--max-turns", str(max_turns),
             "--dangerously-skip-permissions",
         ]
 
@@ -1825,6 +2003,24 @@ class ProcessManager:
                     metrics=metrics,
                 )
             else:
+                # Detect context overflow: Claude exits 1 with "Prompt is too long" in output
+                _combined_output = stdout[-20000:] + stderr[-500:]
+                if "Prompt is too long" in _combined_output or "prompt is too long" in _combined_output:
+                    logger.error(
+                        "Context overflow for task %s: Claude context window exceeded "
+                        "(%d stdout chars). Consider reducing prompt size.",
+                        task_id, len(stdout),
+                    )
+                    return TaskResult(
+                        status="failed",
+                        exit_code=returncode,
+                        stdout=stdout[-settings.AGENT_MAX_OUTPUT_SIZE:],
+                        stderr=stderr[-10000:],
+                        error="context_overflow: Claude's context window was exceeded during execution. "
+                              "The accumulated conversation history (initial prompt + tool call results) "
+                              "exceeded the model's limit. Reduce prompt size or shorten file reads.",
+                        metrics=metrics,
+                    )
                 return TaskResult(
                     status="failed",
                     exit_code=returncode,
@@ -3632,9 +3828,13 @@ class RuntimeDaemon:
             required_files = _get_analysis_outputs_for_type(req_type)
 
             # Analysis deliverables live in analysis_output_dir (docs/requirements/...)
+            _input = task.input_data or {}
             doc_dir = (
-                (task.input_data or {}).get("analysis_output_dir", "")
-                or (task.input_data or {}).get("output_dir", "")
+                _input.get("analysis_output_dir")
+                or _input.get("context", {}).get("analysis_output_dir")
+                or _input.get("output_dir")
+                or _input.get("context", {}).get("output_dir")
+                or ""
             )
             if doc_dir:
                 base = workspace_path / doc_dir
@@ -3718,7 +3918,12 @@ class RuntimeDaemon:
                     pass
 
                 if not _skip_test_artifacts:
-                    doc_dir = (task.input_data or {}).get("output_dir", "")
+                    _input = task.input_data or {}
+                    doc_dir = (
+                        _input.get("output_dir")
+                        or _input.get("context", {}).get("output_dir")
+                        or ""
+                    )
                     if doc_dir:
                         base = workspace_path / doc_dir
                     else:
@@ -3809,11 +4014,25 @@ class RuntimeDaemon:
                 ],
             )
 
-            # Build a targeted fix prompt
+            # Build a targeted fix prompt with output directory context
+            _input = task.input_data or {}
+            _fix_doc_dir = (
+                _input.get("output_dir")
+                or _input.get("context", {}).get("output_dir")
+                or ""
+            )
             fix_prompt = (
                 "The previous execution produced output with validation errors.\n"
                 "Please fix ALL of the following issues:\n\n"
                 f"{issues_text}\n\n"
+            )
+            if _fix_doc_dir:
+                fix_prompt += (
+                    f"IMPORTANT: All deliverable files (test-cases.json, coverage-matrix.json, "
+                    f"test-report.md, design.md, etc.) MUST be written to the `{_fix_doc_dir}/` "
+                    f"directory, NOT the workspace root.\n\n"
+                )
+            fix_prompt += (
                 "Fix the issues in-place. Do NOT recreate files that are already correct.\n"
                 "Only fix the specific problems listed above."
             )
@@ -4024,7 +4243,13 @@ class RuntimeDaemon:
             # ── Pre-push rebase onto latest default branch ──
             # This ensures the AI's branch incorporates the latest remote
             # changes, dramatically reducing PR merge conflicts.
-            await self._rebase_onto_latest(workspace_path, default_branch, task, project_key)
+            # SKIP for analysis nodes: analysis produces documentation files
+            # (PRD.md, SDD.md, etc.) that are independent of codebase changes.
+            # Rebasing analysis commits onto default_branch is unnecessary
+            # overhead and rewrites commit SHAs, which complicates pushes
+            # on subsequent iterations of the same requirement.
+            if task.node_type != "analysis":
+                await self._rebase_onto_latest(workspace_path, default_branch, task, project_key)
 
             # ── Verify we're on the correct branch before pushing ──
             current_branch = ""
@@ -4550,10 +4775,79 @@ class RuntimeDaemon:
                     unpushed = "first-push"
 
                 if unpushed:
+                    # Check if remote has commits not in local history (diverged).
+                    try:
+                        remote_ahead = (await git(
+                            "log", f"HEAD..origin/{branch}", "--oneline", cwd=workspace_path,
+                        )).strip()
+                    except RuntimeError:
+                        remote_ahead = ""  # Remote branch doesn't exist yet
+
+                    if remote_ahead:
+                        remote_count = len(remote_ahead.splitlines())
+                        # Divergence detected — but this is expected after rebase.
+                        # Check if the remote commits are already incorporated
+                        # (i.e., their patches are empty against our branch, meaning
+                        # the rebase already applied them with new SHAs).
+                        # `git cherry HEAD origin/branch` lists commits from
+                        # origin/branch not in HEAD; lines starting with "-" are
+                        # already incorporated (equivalent patch exists).
+                        rebase_divergence = False
+                        try:
+                            cherry_out = (await git(
+                                "cherry", "HEAD", f"origin/{branch}",
+                                cwd=workspace_path,
+                            )).strip()
+                            if cherry_out:
+                                # All lines starting with "-" means all remote
+                                # commits are already in our branch (rebased).
+                                # Lines starting with "+" are truly missing.
+                                truly_missing = [
+                                    line for line in cherry_out.splitlines()
+                                    if line.startswith("+ ")
+                                ]
+                                rebase_divergence = len(truly_missing) == 0
+                            else:
+                                # No cherry output — remote commits are empty or
+                                # fully equivalent
+                                rebase_divergence = True
+                        except RuntimeError:
+                            pass
+
+                        if rebase_divergence:
+                            logger.info(
+                                "Branch %s diverged from origin (%d remote commit(s)) "
+                                "but all are already incorporated (rebase). "
+                                "Using --force-with-lease to push safely.",
+                                branch, remote_count,
+                            )
+                            try:
+                                await git(
+                                    "push", "--force-with-lease", "-u", "origin", branch,
+                                    cwd=workspace_path, project_key=project_key,
+                                )
+                                logger.info("Force-pushed (with lease) branch %s to origin", branch)
+                                return None
+                            except RuntimeError as exc:
+                                logger.error("Force-push (with lease) failed for %s: %s", branch, exc)
+                                return f"Push failed: {exc}"
+                        else:
+                            logger.error(
+                                "SAFETY: Refusing to push %s — remote has %d commit(s) "
+                                "not in local branch. This would destroy prior work. "
+                                "Remote-only commits:\n%s",
+                                branch, remote_count, remote_ahead,
+                            )
+                            return (
+                                f"Push refused: remote branch '{branch}' has {remote_count} "
+                                f"commit(s) not in local history. Force-pushing would "
+                                f"destroy prior implementation work."
+                            )
+
                     logger.info("Found unpushed commits on %s, pushing...", branch)
                     try:
                         await git(
-                            "push", "--force-with-lease", "-u", "origin", branch,
+                            "push", "-u", "origin", branch,
                             cwd=workspace_path, project_key=project_key,
                         )
                         logger.info("Pushed branch %s to origin", branch)

{forgexa_cli-1.5.2 → forgexa_cli-1.6.1}/forgexa_cli.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: forgexa-cli
-Version: 1.5.2
+Version: 1.6.1
 Summary: Forgexa CLI — command-line client and AI agent runtime for the Forgexa platform
 Author-email: Jason Sun <dev.winds@gmail.com>
 License: MIT

{forgexa_cli-1.5.2 → forgexa_cli-1.6.1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "forgexa-cli"
-version = "1.5.2"
+version = "1.6.1"
 description = "Forgexa CLI — command-line client and AI agent runtime for the Forgexa platform"
 requires-python = ">=3.9"
 license = { text = "MIT" }