mirror of
https://github.com/bytedance/deer-flow.git
synced 2026-07-26 16:07:53 +00:00
4 Commits
| Author | SHA1 | Message | Date | |
|---|---|---|---|---|
Xinmin Zeng
|
09988caf95
|
feat(skills): request-scoped secrets for skills (closes #3861) (#3871)
* feat(sandbox): per-call env injection + platform-secret scrubbing for skills Add an env parameter to Sandbox.execute_command (abstract + local + AIO) so request-scoped secrets can be injected into skill subprocesses, and scrub platform credentials (*KEY*/*SECRET*/*TOKEN*/*PASSWORD*/*CREDENTIAL*) from the inherited environment by default so scoped injection is not security theatre. LocalSandbox always passes an explicit scrubbed env; AioSandbox routes env-bearing commands through bash.exec(env=) on a fresh session and leaves the legacy persistent-shell path unchanged. Part of #3861. BEHAVIOR CHANGE: execute_command no longer inherits the full os.environ; Windows encoding tests updated to assert the scrubbed dict. * feat(skills): parse required-secrets frontmatter declaration Add SecretRequirement and Skill.required_secrets, and parse the required-secrets SKILL.md frontmatter field (a string list or {name, optional} mappings), dropping malformed entries with a warning so one bad declaration does not invalidate the skill. The declared name is both the context.secrets key and the env var injected at activation. Part of #3861. * feat(runtime): request-scoped secret carrier (context.secrets) Add SECRETS_CONTEXT_KEY + extract_request_secrets, centralising the context.secrets carrier contract. The existing context passthrough (build_run_config -> _build_runtime_context) already carries the sub-key to runtime.context without mirroring it into configurable; characterization tests lock that behaviour. Part of #3861. * feat(skills): inject declared secrets at slash-activation into bash env Binding point A: when a skill is slash-activated, SkillActivationMiddleware resolves its declared required-secrets against the request's context.secrets and writes the per-run injection set to runtime.context. The bash tool forwards that set to execute_command(env=). A skill cannot harvest a host platform credential (is_host_platform_secret guard, cf. GHSA-rhgp-j443-p4rf), and injected values are redacted from bash output (mask_secret_values) so an echoed secret never re-enters the prompt/trace. Part of #3861. * test(skills): lock the five secret leak surfaces + add trace redaction helper Regression tests assert the secret value is absent from all five surfaces: prompt (activation message), checkpoint (graph state vs context separation), audit (journal records names only), trace (metadata builder never copies context; never mirrored to configurable), and stdout (mask_secret_values). Add redact_secret_context_keys as a defensive helper for any context serialization. Part of #3861. * docs(backend): document request-scoped secrets for skills Add Request-Scoped Secrets subsection (Skills) + env policy note (Sandbox) and the execute_command(env=) signature change, per the doc-sync policy. Part of #3861. * fix(skills): close gaps found by end-to-end verification of request-scoped secrets Real-gateway e2e + independent review of #3861 surfaced three defects, now fixed: 1. Slash activation never fired in the live chain. InputSanitizationMiddleware wraps user input in BEGIN/END markers before SkillActivationMiddleware sees it, and the original text was only preserved when an upload or IM channel set it. For a plain text message the slash command became undetectable, so no secret was ever resolved. Fix: the sanitizer now setdefaults the pre-wrap text into ORIGINAL_USER_CONTENT_KEY (additive; sanitization behaviour unchanged), so slash activation works for all messages. Pre-existing latent bug surfaced here. 2. The raw request config (with context.secrets) was persisted to runs.kwargs_json and echoed by the run API (RunResponse.kwargs). Fix: redact_config_secrets() strips secret-bearing context keys from the persisted/echoed copy in start_run; the live config that drives the run keeps them. build_run_config now also sets configurable.thread_id on the context path (the checkpointer requires it). 3. Connection-string credentials (DATABASE_URL, REDIS_URL, SENTRY_DSN, GH_PAT, ...) were not scrubbed from the inherited sandbox env. Fix: env_policy adds a *DSN* pattern plus an explicit connection-string denylist (no blanket *URL* — benign service URLs stay readable). Verified end-to-end via a real gateway run (real LLM + skill activation + bash): the secret reaches the sandbox subprocess and appears in NONE of prompt, trace, checkpoint, audit, stdout, runs.kwargs_json, or the run API. Part of #3861. * docs(backend): document the env scrub, persistence redaction, and sanitizer interaction Sync the Request-Scoped Secrets section with the verification-driven fixes: inherited-env scrub (incl. connection-string denylist), run-record/run-API redaction as the 6th sealed leak surface, and the sanitizer preserving original content so slash activation fires. Part of #3861. * fix(skills): inject caller secret over scrubbed host value; drop redundant host-name guard A real-world demo (a skill calling a third-party cloud API with a request-scoped key) exposed that the is_host_platform_secret guard was both wrong and harmful: it refused to inject a caller-supplied secret whenever a same-named variable existed in the Gateway env — which is exactly the #3861 use case (a per-user key overriding a shared platform key). The guard was also redundant: build_sandbox_env already scrubs secret-looking names from the inherited env before injection, so a skill can never read a host credential — it only ever receives the caller's value. Remove the guard; the injected (caller) value simply wins over the scrubbed host value. Verified end-to-end: the agent called the real cloud API successfully with the caller's key, the host's same-named key was scrubbed and never used, and the caller's key leaked to none of the surfaces. Part of #3861. * fix(skills): address review on request-scoped secrets (#3861) Review fixes from PR #3871: - E2BSandbox.execute_command now accepts env/timeout and routes them to commands.run(envs=, timeout=). The bash tool passes env= unconditionally, so the prior signature (command only) raised TypeError on every e2b bash call and broke e2b deployments entirely. env=None stays backward-compatible. - SkillActivationMiddleware clears the active-secret set before resolving each activation, so a later skill in the same run never inherits an earlier skill's injection set (the #3861 contract: a skill only receives what the caller supplied AND that skill declared). - AioSandbox env path uses a dedicated _DEFAULT_HARD_TIMEOUT — bash.exec exposes no idle/no-change timeout, so the prior reuse of the legacy idle constant conflated wall-clock vs idle semantics. The env path also retries on the ErrorObservation signature now, sharing the legacy persistent-shell recovery contract. - mask_secret_values skips values below a minimum length floor so a short declared secret (e.g. "42") cannot shred unrelated bytes (exit codes, timestamps, sizes) of tool output. The secret is still injected into the subprocess; only the output mask skips it. session_id reuse on the env path is intentionally NOT added: a shared session could let request-scoped secrets ride the session env into later commands, which the SDK does not contractually forbid. The fresh-session choice matches the LocalSandbox model (each call is a fresh subprocess); the trade-off (consecutive env-bearing calls do not share cwd/venv/exports) is documented on _execute_with_env. |
||
|
Zhipeng Zheng
|
e7b88a97ed
|
fix(security): add input sanitization middleware for prompt-injection defense (#3630) (#3662)
InputSanitizationMiddleware: escapes blocked XML tags in user messages (<system> -> <system>) — de-identify, don't reject (AWS PII-style). Wraps user input in plain-text boundary markers (--- BEGIN/END USER INPUT ---) per OWASP structured-prompt guidance. System-Context Confidentiality clause prevents LLM from revealing internal instructions. Boundary-marker layer hardened against delimiter injection (self-suppression + break-out). Multimodal content blocks preserved during sanitization. Addresses review feedback from @fancyboi999 and @WillemJiang: 1. Boundary-marker injection: strict startswith+endswith idempotency, neutralize user-supplied BEGIN/END tokens, forged-idempotency bypass fixed 2. Multimodal data loss: _rebuild_content preserves interleaved non-text blocks 3. CI fixes: replay_provider strips boundary markers before hashing, middleware chain assertions updated 4. Lint: ruff format applied on Linux (slice spacing + f-string collapse) Test: 71 input_sanitization + 1 replay_golden + 14 tool_error + 93 tool_output_budget = 179/179 passed |
||
|
Huixin615
|
64d923b0fd
|
fix(middleware): externalize oversized tool output into sandbox for non-mounted sandboxes (#3417)
* fix(middleware): externalize oversized tool output into sandbox for non-mounted sandboxes
ToolOutputBudgetMiddleware persisted oversized tool results to the host
filesystem and returned a /mnt/user-data/outputs virtual path. For sandboxes
that do not use thread-data mounts (e.g. remote AIO sandbox), that virtual
path does not exist inside the sandbox, so the model's read_file tool could
not read it back and reported 'file not found'.
Branch on SandboxProvider.uses_thread_data_mounts:
- Mounted sandboxes (local Docker, AIO + LocalContainerBackend) keep the
original host-disk path; the host outputs dir is bind-mounted to the same
virtual path inside the sandbox, so behavior is unchanged.
- Non-mounted (remote) sandboxes externalize into the sandbox itself via
execute_command('mkdir -p ...') + write_file + 'test -s' validation. The
validation step is required because AIO sandbox execute_command returns
'Error: ...' as a string on failure instead of raising, so a silent mkdir
failure would otherwise leak through.
Any failure (rejected subdir, mkdir/write/validate error) falls back to the
existing inline head+tail truncation, so an unreadable path is never returned
to the model.
The sandbox resolver reads the sandbox_id that SandboxMiddleware already
writes into runtime.state['sandbox']; it never calls provider.acquire(),
keeping the tool-call hot path free of blocking I/O. Tools that do not use a
sandbox (web_search, MCP, ...) resolve to None and fall through to inline
truncation, which is the safe behavior for them.
Fixes #3416
* fix(middleware): address Copilot review feedback on sandbox externalization
- Make get_sandbox_provider() lookup best-effort in _budget_content: only
query when outputs_path or sandbox is available, and fall back to inline
truncation if provider initialization raises rather than propagating
the error. A resolved sandbox instance is sufficient on its own to take
the non-mounted externalization branch.
- Strict-match the sandbox post-write validation echo
(check.strip() == 'OK') to avoid false positives if execute_command
ever surfaces unrelated stdout/stderr containing 'OK' as a substring.
Refs: #3417
* test: fix flaky tests relying on /nonexistent/... path under container root
Two tests in this module (test_returns_none_on_invalid_path and
test_fallback_when_disk_write_fails) used paths like
'/nonexistent/impossible/path' to trigger _externalize's OSError
fallback. These paths are creatable when the test process runs as root
inside the CI container: os.makedirs(..., exist_ok=True) successfully
creates the entire chain under /, so the OSError branch is never hit
and the tests fail. Reproducible on main independently of this PR.
Switch to '/dev/null/cannot-mkdir-here'. /dev/null is a character
device on both Linux and macOS, so os.makedirs always fails with
NotADirectoryError regardless of privileges, reliably exercising the
OSError fallback.
* fix(tool-output-budget): only consult sandbox provider when a sandbox is resolved
The previous revision called get_sandbox_provider() whenever externalization
was triggered, including on the legacy host-disk path. Environments without
a configured sandbox -- in particular CI runners without a config.yaml --
would raise FileNotFoundError there, get caught, and silently fall back to
inline truncation. That defeated the host-disk externalization path that
predates this PR and was the root cause of the regressing legacy tests.
Restructure the branching so the provider is only consulted when a sandbox
has actually been resolved for the current tool call:
- sandbox resolved + provider.uses_thread_data_mounts: host-disk write
(bind-mounted into the sandbox, equivalent to a sandbox-side write).
- sandbox resolved + non-mounted provider: sandbox write (#3416).
- no sandbox + outputs_path: host-disk write
(legacy / non-sandbox tools, no provider call at all).
- otherwise: inline fallback.
No test changes; the legacy externalization tests are provider-agnostic by
construction and now pass without monkeypatching.
Refs: #3416
* test(tool-output-budget): assert legacy path does not call sandbox provider
Lock in the contract introduced by d6e2d25b: when no sandbox is resolved
for a tool call, _budget_content must externalize to the host outputs
directory without consulting get_sandbox_provider(). Regressing this would
re-break legacy / non-sandbox tools in environments without a configured
sandbox (e.g. CI without config.yaml), which is the failure mode #3416's
fix avoids.
The test injects a get_sandbox_provider that raises on call, so any
future refactor that moves the provider lookup out of the sandbox-only
branch will fail loudly.
Refs: #3416
|
||
|
Xinmin Zeng
|
ca487578a4
|
feat(agent): add ToolOutputBudgetMiddleware for oversized tool output protection (#3303)
* feat(agent): add ToolOutputBudgetMiddleware for oversized tool output protection Closes #3289. Adds a unified middleware that enforces per-result budgets on ALL tool outputs (MCP, sandbox, community, custom), preventing oversized external tool results from blowing the model context window. Design informed by claude-code (persistToolResult), hermes-agent (tool_result_storage), and pi (OutputAccumulator) — the three most mature implementations in production coding-agent frameworks. Key features: - Disk externalization: oversized outputs written to thread-local .tool-results/ directory, replaced with compact preview + file reference. Model can read full output via read_file with offset/limit. - Fallback truncation: head+tail truncation when disk is unavailable (no thread_data, write failure), ensuring the context is always protected. - read_file exemption: prevents persist-read-persist infinite loops (independently discovered by claude-code, hermes-agent, and pi). - Per-tool threshold overrides via config. - Line-boundary-aware truncation (no partial lines in previews). - Multimodal content passthrough (images/structured blocks skip budget). - Historical ToolMessage patching in wrap_model_call for checkpoint recovery scenarios. Related: #3222 (design RFC), #1844 (comprehensive context management), #3137 (write_file args compaction), #1677 (sandbox tool truncation). * test: add MCP content_and_artifact format coverage Add 5 tests for MCP tool output format (list of content blocks): - text content blocks are extracted and budgeted - multiple text blocks are joined and budgeted - image content blocks are skipped (multimodal passthrough) - mixed text+image blocks are skipped - small text blocks pass through unchanged Total test count: 59 (was 54). * fix(agent): address Codex review findings for ToolOutputBudgetMiddleware Three issues identified by Codex code review, all fixed: 1. `enabled` config field was unused — middleware now checks `config.enabled` and skips all processing when disabled. 2. `_build_fallback` could exceed `fallback_max_chars` — the marker text itself (~139 chars) was not deducted from the budget. Now pre-computes marker overhead and falls back to hard slice when max_chars is smaller than the marker. 3. Sync file I/O in async path — `awrap_tool_call` now delegates `_patch_result` to `asyncio.to_thread` to avoid blocking the event loop during disk writes. Tests updated to use realistic fallback_max_chars values (500+) that can accommodate the marker overhead, plus two new tests: - `test_result_never_exceeds_max_chars` (parametric across sizes) - `test_very_small_max_chars_does_not_crash` * fix(agent): address Copilot review — path traversal, async perf, shared config 1. Path traversal defense: sanitize tool_name via _sanitize_tool_name() (strips separators, .., absolute paths), validate storage_subdir is relative, and verify resolved filepath stays inside storage_dir. 2. Async hot-path optimization: add _needs_budget() cheap check before asyncio.to_thread offload — small outputs (99% of calls) skip the thread overhead entirely. 3. Replace shared module-level _DEFAULT_CONFIG with _default_config() factory to prevent cross-instance mutation of mutable fields. 12 new tests: TestSanitizeToolName (5), TestExternalizePathTraversal (3), TestNeedsBudget (4). * fix(agent): correct preview hint to match read_file actual API read_file uses start_line/end_line (1-indexed line numbers), not offset/limit. The previous wording was copied from hermes-agent which has a different read_file interface. * perf(agent): hoist hot-path imports, add model-call pre-scan (review #3303) Address maintainer review feedback: 1. Hoist inline imports to module level — `import asyncio` (was in awrap_tool_call hot path) and `from dataclasses import replace` (was in _patch_result) now live at module top. 2. Add a cheap pre-scan to _patch_model_messages so the historical message list is not rebuilt on every model call when nothing is oversized (the common case once results are budgeted at tool-call time). Also adds the same _needs_budget gate to the sync wrap_tool_call for symmetry with awrap_tool_call. The pre-scan is refactored into per-tool-aware helpers (_effective_trigger / _tool_message_over_budget) that mirror the exact trigger conditions in _budget_content — including tool_overrides — so the fast-path can never produce a false negative (silently skipping budgeting for a tool with a low per-tool threshold). 7 new regression tests lock the per-tool-override-through-pre-scan path and the model-call early return. --------- Co-authored-by: Willem Jiang <willem.jiang@gmail.com> |