### Middleware Chain

Compaction keeps state `SystemMessage`s; transient instructions use request
wrappers, and fully rescued partitions skip compaction. If latest-user rescue
empties an AI/Tool-only window, use `_build_summary_input_text(strategy="last")`;
mixed windows keep normal anchoring and final-message fallback. Budget raw text
before escaping/wrapping; pass `trim_tokens_to_summarize=None` to avoid the
LangChain default.

Delegation verdicts are untrusted: revalidate persisted values, ignore malformed
ones, and treat completed work as reusable evidence rather than acceptance.

On new user turns, DurableContext cancels earlier-run unanswered delegations.
It preserves resumes, same-run continuations, and entries without `run_id`.
Any reply prevents cancellation; legacy replies without status metadata may
stay `in_progress`. Never infer status from reply text.

Assembly order: `tool_error_handling_middleware.py::_build_runtime_middlewares` (exposed as `build_lead_runtime_middlewares`), then `../lead_agent/agent.py::build_middlewares` appends lead-only entries. Optional entries require their config/runtime condition.

**Message provenance.** At injection/rewrite, always stamp `additional_kwargs`
via `deerflow_extension_api.provenance.provenance_kwargs()`:
`deerflow_content_kind`, `deerflow_producer_kind`, optional
`deerflow_producer_entity_id`. All are server-owned inbound metadata; stamp even
without observers, since downstream cannot recover producers. Producers:
DynamicContext (reminder/memory), DurableContext (contract/data),
SystemMessageCoalescing, ViewImage, SkillActivation. Summarization/Title use
`SystemOperationKind.SUMMARIZATION`/`.TITLE` model-call attribution; summaries
enter via DurableContext's stamped `durable_context_data`, not separate
messages. Memory only queues extraction; recall uses DynamicContext's
`dynamic_context_memory` stamp.

**Middleware self-description.** Behaviour-configurable middleware implements
`release_policy_parameters() -> dict[str, object]` (duck-typed
`deerflow_extension_api.release.ReleasePolicyProvider`, no base class).
Use JSON-serialisable values and `canonical_hash` for long text, not prompt
copies. `collect_release_policies()` gathers stack declarations; update them
alongside every behaviour-affecting field.
Summarization declares enabled `task_continuity` retention settings (otherwise
`None`). DurableContext declares its normalized skills root, sorted read-tool
names and continuity switch, so each capture/injection policy affects assembly
identity without depending on private-field probing.
Continuity history readers share shape validation, including the capture failure
path and DurableContext rendering, so malformed persisted metadata cannot abort
ordinary compaction or a model call.

**Removing tool calls.** Use `clone_ai_message_with_tool_calls`, not a bare
`tool_calls` update: adapters resend stale `content` tool-call blocks, which
strict providers reject.

**Shared runtime base** (`build_lead_runtime_middlewares`; subagents reuse most of this via `build_subagent_runtime_middlewares`):

1. **InputSanitizationMiddleware** - First, so it is the outermost `wrap_model_call` wrapper; every inner middleware (including LLM retries) sees sanitized messages. `additional_kwargs.original_user_content` is server-owned provenance: Gateway strips caller-supplied values for non-internal run requests, trusted IM calls may carry the string they captured before adding transport/file context, and the middleware replaces any non-string value before wrapping. Uploads and sanitization retain first-writer-wins only for validated strings. Caller markers are marked `untrusted_input`, never stripped; scope is every turn.

   **KnowledgeScopeMiddleware** follows input sanitization: it exposes only Gateway-admitted execution scope, removes scope/display data from model messages, and blocks `knowledge_search` when disabled without reading storage or RAGFlow.
2. **ToolOutputBudgetMiddleware** - Caps model-bound tool output per app config. Externalizes oversized results to `tool_output.storage_subdir` (default `.tool-results`, constant `TOOL_RESULTS_DIRNAME`) under thread outputs, leaving a typed synopsis + `read_file` reference. These process-feedback files are excluded from workspace-change scans and delivery verification. `wrap_model_call` elides successful `write_file` content only in model-bound requests (#5328) after a later successful same-path `read_file`/`write_file`/`str_replace`; the on-disk file becomes the reference. Preserves the newest `keep_recent_writes` writes. Pairs call occurrences via `tool_call_args.pair_tool_call_results` and rewrites through shared `tool_call_args` helpers; controls: `elide_superseded_writes`, `superseded_write_min_chars`.
3. **ToolResultSanitizationMiddleware** - Neutralizes framework/injection tags (e.g. `<system-reminder>`) and boundary markers in *remote-content* tool results (`web_fetch`/`web_search`/`image_search`/`web_capture`) so attacker-controlled fetched pages cannot forge trusted framework context. Mirrors `InputSanitizationMiddleware`'s user-input guardrail for the other untrusted-content entry point; sits inner of `ToolOutputBudgetMiddleware` (neutralizes the raw output, then the budget truncates). Local tool output (bash/read_file) is left untouched. Scope is a name-based allowlist for the first-party web tools, plus every MCP-sourced tool via its `deerflow_mcp` metadata tag, so an MCP server naming its fetcher `fetch_url` is still covered

   Result-rewriting middlewares between the raw callable boundary and the
   model-visible result append a declared entry to
   `additional_kwargs["deerflow_tool_transforms"]` via
   `agents/middlewares/tool_transform_meta.py::append_tool_transform`. The trail is
   ordered by application — the last entry produced the final visible bytes — so an
   observer classifies raw→visible transforms from facts rather than by sniffing
   output wording.
4. **PiiRedactionMiddleware** - *(optional, `pii_redaction.enabled`, default off, #3190)* Rewrites PII in genuine user messages (`wrap_model_call`, request-scoped, raw text stays in thread state) and remote-content tool results (`wrap_tool_call`, ToolResultSanitizationMiddleware's allowlist incl. `Command.update.messages`) to irreversible placeholders (`[EMAIL_1]`, …). Deterministic regex detectors only, no new dependencies; national IDs run before the card detector so a checksum-valid resident ID whose digits also pass Luhn is never mislabeled `[CREDIT_CARD_n]`. One redactor per result keeps numbering continuous; existing summary/message tokens reserve indices before new values; no persistent mapping links compacted identities. Innermost Layer-1 wrapper, so budget-externalized copies hold redacted text; compaction, reinjected summaries, and title-model inputs are redacted; title fields are redacted before truncation. Memory extraction is a follow-up slice.

5. **ThreadDataMiddleware** - Creates per-thread directories under the user's isolation scope (`backend/.deer-flow/users/{user_id}/threads/{thread_id}/user-data/{workspace,uploads,outputs}`); resolves identity via `resolve_runtime_user_id(runtime)`, including Gateway runtime context and standalone LangGraph Server auth, then falls back to the request ContextVar / `"default"`
6. **UploadsMiddleware** - Tracks and injects newly uploaded files into conversation (lead agent only); upload existence checks use the same runtime-resolved user bucket as thread-data creation
7. **SandboxMiddleware** - Acquires sandbox, stores `sandbox_id` in state. The
   lead runtime normally owns the thread's physical Agent-skill projection;
   delegated subagents and the prompt-only bootstrap agent are non-owners, so
   their narrower discovery allowlists never rebuild the shared thread view or
   force eager sandbox acquisition.
8. **DanglingToolCallMiddleware** - Injects placeholder ToolMessages for AIMessage tool_calls that lack responses (e.g., user interruption), preserving raw provider tool-call payloads in `additional_kwargs["tool_calls"]`; malformed tool-call names and arguments are sanitized in the model-bound request so strict OpenAI-compatible providers do not reject the next request
9. **LLMErrorHandlingMiddleware** - Converts provider/model failures to recoverable assistant errors. Sync and async calls carry circuit-generation ownership; only the current owner may settle or release a half-open probe, so stale completions cannot affect a newer recovery attempt. Cancellation propagates unchanged without retry or failure accounting.
10. **Authorization / GuardrailMiddleware** - Up to two independent pre-tool-call gates run here. When `authorization.enabled`, the `AuthorizationProvider` instance already used for Layer 1 capability filtering is wrapped by `GuardrailAuthorizationAdapter` and reused for Layer 2 execution checks. A generated `tool_search` bypasses the adapter's second provider call only when the current build has a concrete deferred setup; its catalog was already filtered by Layer 1, and an ordinary same-named tool without that deferred setup receives no exemption. When `guardrails.enabled`, the explicitly configured `GuardrailProvider` is appended after authorization and still evaluates every call, including `tool_search`. Authorization therefore runs outermost and can deny before an external guardrail call; both use the existing middleware's fail-closed, audit, sync/async, and error-`ToolMessage` behavior. See the authorization RFC and [docs/GUARDRAILS.md](../../../../../docs/GUARDRAILS.md).

   Every guardrail decision path publishes a neutral
   `deerflow.authz.outcome.AuthorizationOutcome` into the per-run runtime context,
   keyed by `tool_call_id` under the `__`-prefixed
   `__authorization_outcome` key (so `build_run_config` strips caller-supplied
   forgeries). Consumers pop it; the publisher and the consumer share only that
   contract module.
11. **SandboxAuditMiddleware** - Audits sandboxed shell/file operations before tool execution; command classification is **defense-in-depth and audit, not a security boundary** (the sandbox is the isolation boundary). Command substitution is judged by *position*, not the presence of `$(`: **command position** (`$(curl url)`, `` `curl url` ``, the word after `|`/`&&`/`;`, an `eval`/`source` argument) executes fetched content and is blocked; **value position** (`x=$(curl url)`, `echo $(curl url)`, an argument, a `for` word list) only captures output and passes (#4611). So `_HIGH_RISK_COMMAND_POSITION_PATTERNS` is matched anchored against each sub-command from `_split_compound_command(split_pipes=True)`, never the whole string; pipe-spanning rules (`| sh`, `base64 -d | ...`) still use `_classify_command`'s whole-command Pass 1. `_COMMAND_POSITION_PREFIX` extends the anchor over leading assignments and exec wrappers (`FOO=1 $(curl url)`, `env`/`command`/`builtin`/`exec`/`nohup`/`time`/`sudo`/`doas`); its assignment branch requires whitespace before the substitution, which keeps `x=$(curl url)` in value position. Two contexts are deliberately **position-blind** (matched whole-command in Pass 1, since they execute their input anywhere, e.g. `xargs sh -c "$(curl url)"`): an `eval`/`source` argument, and an interpreter **code-string flag** — `-c` (shells, `python`), `-e` (`perl`/`ruby`/`node`), `-p` (`perl`/`node`), `-r` (`php`) — plus the here-string (`<<<`) reaching the same place via stdin. All three substitution spellings (`$(`, `<(`, `` ` ``) share one `_RISKY_SUBSTITUTION` opener. An unquoted newline splits like `;` (else `echo hi\n$(curl url)` evades the anchored rules). Heredoc bodies are data: `_split_compound_command` records headers (`<<EOF`, `<<-EOF`, `<<'EOF'`) and consumes their bodies verbatim, so a body line starting `$(curl url)` isn't promoted to a command position; `<<<` (here-string, needs look-ahead + look-behind) and a `<<` inside `$(( ))`/`(( ))` (bit shift, arithmetic depth tracked with the quote flags) must not open one. This is a heuristic, not shell parsing — an unterminated body consumes the rest of the string, an unclosed `((` only disables heredoc detection, and the failure direction is always toward *more* command positions, not fewer. Known gaps: process substitution outside `eval`/`source` (`. <(curl u)`) is undetected, and two-step forms (`x=$(curl u); eval "$x"`) need dataflow analysis. No config gate — appended unconditionally in `_build_runtime_middlewares`, for both lead and subagents.
12. **ReadBeforeWriteMiddleware** - *(optional, `read_before_write.enabled`, default on)* Outermost write gate (#3857): `read_file` stamps a content hash on its ToolMessage; `write_file` (existing file, incl. append) and `str_replace` are blocked unless the newest mark for the path matches its current hash. Sits outside ToolProgress/ToolErrorHandling (a block consumes no ToolProgress slot); blocked results self-stamp `deerflow_tool_meta` and carry `deerflow_write_block` (`{path, tool}`). Marks live on messages, so summarization dropping the read invalidates the gate; writes never refresh marks. Gate check + execution are serialized per (thread, path); `"Error: ..."`-string sandboxes (AIO/E2B) fail open. It owns the composed call's sandbox authorization scope; `SandboxAuthorizationError` becomes an error ToolMessage. Its `wrap_model_call` swaps blocked calls' dead payload (`content`, `old_str`/`new_str`) for a deterministic placeholder in the model-bound request only (`elide_blocked_payloads`, `elide_min_chars`); state, receipts, and the journal keep the originals. Policy stays in the gate; the shared `tool_call_args` helper rewrites every arg surface together and every model-bound arg rewrite must use it.
13. **ToolProgressMiddleware** - *(optional, if `tool_progress.enabled`)* State-machine-based stagnation guard (RFC #3177). Outer wrapper around ToolErrorHandlingMiddleware so its tool wrapper receives results already stamped with `deerflow_tool_meta`. Recoverable problems stay WARNED with hints; retryable non-recoverable problems escalate to BLOCKED; stop-category failures block immediately. It is independent of LoopDetection's turn-level call-pattern guard. Effective phase changes emit `middleware:tool_progress`, including `reset` when a later agent invocation (for example, goal continuation) clears WARNED/BLOCKED state. The state machine supplies the action and count threshold (`null` for category rules, recovery, and reset); recorder calls happen after releasing the state lock. Events include the actual sync/async hook plus bounded status, error, and recovery fields. Server-owned recorder keys determine subagent attribution. Arguments, content, prompts, and hashes are never persisted; recorder failures are fail-open. See [event semantics](../../../../../docs/RUN_EVENT_STREAM.md).
14. **ToolReceiptMiddleware + ToolErrorHandlingMiddleware** - `ToolReceiptMiddleware` is *(optional, if `verification.receipts_enabled`, default on)*. It is the **outermost `wrap_tool_call` layer** — registered ahead of entries 10-13 — because Guardrail/SandboxAudit/ReadBeforeWrite/ToolProgress can short-circuit a call with their own ToolMessage (and SandboxAudit rebuilds medium-risk results); an inner receipt layer would silently gap the ledger on those results (ordering constraints in `deerflow.extensions.ordering`). Normal results still carry the `deerflow_tool_meta` status ToolErrorHandlingMiddleware stamps on the inner return path; short-circuit messages self-stamp meta or fall back to `message.status`. It stamps deterministic provenance (tool name, status, args/output hashes, byte count, timestamp) onto direct `ToolMessage` results and every matching `ToolMessage` carried in `Command.update.messages`, including delegated `task`, `present_file`, `view_image`, and `tool_search` results; before model calls it derives a hidden receipt ledger (display ids r1..rN) from message state, and when the 2,000-character budget is exceeded the newest receipts are retained in chronological order with their original ids plus an older-receipts omission marker. Rendering returns both the text and its retained receipt subset; every response that received a ledger carries only that exact server-owned subset, never omitted receipts. Snapshot validation accepts a strictly consecutive positive original-id range (for example `r24`–`r30`) rather than requiring `r1`, so subagent terminal citation verification resolves ids against evidence present in the citing turn even when later summarization drops and renumbers tool messages. Model-generated citation IDs are digit-bounded before integer conversion; oversized IDs are ignored as malformed input rather than raising through task write-back. Citation parsing deduplicates exact `(id, anchor)` pairs, not IDs alone, so repeated identical references stay compact while every distinct anchor claim is verified. Gateway strips delegated receipts/verdicts from external messages. `ToolErrorHandlingMiddleware` receives `AppConfig`, converts tool exceptions into error `ToolMessage`s so the run can continue instead of aborting, stamps every result with `deerflow_tool_meta` (status / error_type / recoverable_by_model / recommended_next_action / source) via `tool_result_meta.normalize_tool_result`, stamps structured metadata for task exception wrappers, and stamps skill-read metadata for downstream durable-context capture. Task tool result text is generated from the same status/result/error inputs as the structured metadata so callers do not hand-write a second protocol string.

Authorization identity is independent of enforcement. Gateway strips client identity overrides: only the server auth source sets `is_internal`, and only authenticated IM `body.context` supplies `channel_user_id` (never `body.config`). `build_principal_from_context` applies role defaults, strict provenance, and copied attributes; RBAC rejects unknown defaults. Delegation and `GuardrailMiddleware` share this identity. Layer 1 precedes deferred assembly across agent paths and its provider is reused for Layer 2; framework skill/memory ordering stays stable. Trusted `DeerFlowClient.stream()` accepts identity overrides. Its graph key always includes effective storage `user_id` and, when enforced, the full Principal; nested attributes are copied so mutation cannot hide stale cache state.

Gateway route permissions and per-caller skill visibility are documented in the [Gateway guide](../../../../../app/gateway/AGENTS.md#route-and-skill-listing-authorization). Runtime skill activation remains a separate layer.

Model authorization uses `authz.py::resolve_model_authorization()` (same cached-provider, internal-role, and principal-building path as route authorization) as the Gateway integration point for the `models` router: `list_models` filters names through `filter_resources(principal, "model", ...)`, and `get_model` enforces `authorize(resource="model", action="use")` with a deny surfacing as 403; provider errors follow `authorization.fail_closed` (fail-open returns the unfiltered list / proceeds). At runtime, `lead_agent/agent.py::_authorize_model_name` — called from `_make_lead_agent` and from `DeerFlowClient._ensure_agent` — applies the same `model:use` check to the resolved model name. On deny it scans the `filter_resources`-visible names (excluding the denied model), re-verifying each candidate with `authorize("model", "use")` before falling back, because a custom provider may allow `list` while denying `use`; no usable fallback raises under `fail_closed` and keeps the original model under fail-open. The built-in RBAC provider maps this to the per-role `models` policy key. Tests: `tests/test_models_authorization.py`.

Sandbox authorization (`sandbox:execute`) gates every sandbox acquisition before `provider.acquire` — including this middleware's eager path (`before_agent` / `abefore_agent` skip acquisition on deny instead of raising, deferring to the lazy per-tool gate). See the [sandbox module guide](../../sandbox/AGENTS.md) authorization-gate paragraph and `tests/test_sandbox_authorization.py`.

Before changing a later authorization phase, read the [authorization RFC](../../../../../../docs/plans/2026-07-10-pluggable-authorization-rfc.md) and its [implementation notes](../../../../../../docs/plans/2026-07-10-pluggable-authorization-implementation-notes.md). The notes are the cumulative handoff record for merged PR behavior, reviewer feedback, trust-boundary decisions, deferred scope, and required regression coverage.

**Lead-only middlewares** (`build_middlewares`, appended after the base):

15. **DynamicContextMiddleware** - Injects date and optional memory outside the static prompt. Opt-out removes its frozen server memory but retains date/user messages. Date follows server-local time unless `DEER_FLOW_DATE_TIMEZONE` names an IANA zone (invalid values fall back locally).
16. **SkillActivationMiddleware** - Detects strict `/skill-name task` syntax on the latest real user message, resolves only enabled and runtime-allowed skills, injects the `SKILL.md` body as hidden current-turn context, and records a `middleware:skill_activation` audit event
17. **SkillToolPolicyMiddleware** - Applies `allowed-tools` only after real activation; passive enabled skills and a custom agent's configured skill allowlist do not clamp the lead toolset. A run-scoped slash activation is authoritative and suppresses `skill_context` as a policy source, so reading another skill cannot widen the explicit skill's tools; without slash activation, skills captured after configured `read_file` loads retain the existing union semantics. The middleware filters model-visible schemas and blocks unauthorized execution, resolving canonical paths against the live enabled/agent-allowed registry on every model call, then stores a versioned, JSON-safe, middleware-token-bound decision signed by policy source plus active paths in run context for the resulting tool calls to reuse. The next model call always refreshes it, and malformed, foreign, stale, or unmatched decisions fall back to live resolution. `tool_search` and `describe_skill` remain framework-safe discovery tools under a restrictive policy; they may reveal or promote metadata, but a deferred business tool must still be declared by the active policy before its schema or execution can survive the policy middleware. The decision's owner token is authorization-sensitive, so its reserved context key is owned by `runtime.secret_context` and included in `REDACTED_CONTEXT_KEYS` for observable and persisted context copies. Registry load failures and a non-empty active set with no authorized skill fail closed to framework-safe tools; an individual stale path is skipped only when at least one valid active skill remains. This is best-effort behavioral scoping rather than a hard security boundary: alternate loads such as `bash cat` are not captured, and bounded autonomous `skill_context` can evict old entries. `task` is not framework-exempt, so a restricted skill cannot delegate around its policy. The middleware must remain immediately after `SkillActivationMiddleware` (which publishes the slash source through `runtime.secret_context`'s public path helpers authenticated by a required token shared only within the assembled middleware chain) and immediately before `DurableContextMiddleware`; assembly and compiled-graph tests pin ordering, token sharing, schema filtering, and execution blocking.
18. **DurableContextMiddleware** - Captures `task` delegations into `ThreadState.delegations` (including in-progress dispatches and terminal result summaries) and loaded skill-file references (name/path/description, parsed in-memory - not the body) into `ThreadState.skill_context` before summarization can compact the paired tool-call/result messages, then projects durable context into each model request. Static authority rules are injected as a `SystemMessage`; untrusted field values (`summary_text`, delegation results, skill descriptions) are injected separately as a hidden `HumanMessage` data block so compressed history, delegated work, and which skills are active stay visible without being stored as `messages` or promoted to system-role instructions. `build_subagent_runtime_middlewares` also attaches this middleware immediately before subagent summarization so a compacted `summary_text` is projected ahead of a preserved assistant/tool tail instead of leaving strict providers with an assistant-first request.
19. **SummarizationMiddleware** - *(optional, if enabled)* Compacts near token limits; memory flush follows runtime policy, while manual compaction trusts the checkpoint-bound agent, so opt-out never captures removed turns. It preserves the latest real user request by ID and tagged DynamicContext reminders while allowing stale ID-swap peers into the summary. Moving the cutoff backward can retain old AI/tool turns and make first-turn compaction a no-op.
20. **TodoListMiddleware** - *(optional, if `is_plan_mode`)* Task tracking with `write_todos`. Skips reminders on `model_length_termination` so capped turns end cleanly.
21. **TokenUsageMiddleware** - *(optional, if `token_usage.enabled`)* Records token usage metrics; subagent usage is read from terminal `ToolMessage.additional_kwargs` in the current run and merged back into the dispatching AIMessage by message position. The same state update marks the ToolMessage with `subagent_token_usage_attributed=true`, so checkpoint replay or middleware re-entry cannot add the cumulative snapshot twice; missing/malformed usage or a result with no matching dispatch remains unmarked and retryable.
22. **TitleMiddleware** - Auto-generates the thread title after the first complete exchange and normalizes structured message content before prompting the title model. If a first-turn run is interrupted before this middleware can write a title, `runtime/runs/worker.py` keeps the run in a finalizing state, persists a local fallback title from the latest checkpoint or original run input, and then syncs it to `threads_meta.display_name`. Replacement runs admitted by `multitask_strategy="interrupt"` / `"rollback"` wait for older same-thread finalization before entering the graph; the interrupted run only skips the fallback title write once a later run has started and may have advanced the checkpoint.
23. **MemoryMiddleware** - Queues conversations for async memory update (filters to user + final AI responses); captures the runtime-resolved user so standalone LangGraph Server reads and writes stay in the same bucket
24. **ViewImageMiddleware** - *(optional, if the model supports vision)* Appends a hidden HumanMessage with base64 image data, identified by a reserved ID prefix plus a server-owned metadata marker, to `ModelRequest.messages` in `wrap_model_call` / `awrap_model_call`. The payload lives only in that request and is never returned as a state update, so no checkpoint carries it and an interrupted run cannot strand it in history; state keeps only the lightweight `viewed_images` metadata. It owns that context and rebuilds it per call: its own message is swept out of the request first — a thread checkpointed by the earlier `before_model`/`after_model` pair (which wrote the payload into state and took it back out with `RemoveMessage`) can carry one that reached state but was never removed, and leaving it in would resend that base64 in every later request for the life of the thread — then a freshly built one is appended when warranted. The sweep requires both the reserved ID prefix and the server-owned marker, so a client cannot get its own message dropped; unmarked leftovers predating the marker are left in place and merely not duplicated
25. **McpRoutingMiddleware** - Auto-promotes deferred schemas matching the latest real user message before `DeferredToolFilterMiddleware`, without executing tools. New names emit `middleware:tool_promotion` with `source=routing_hint`; repeat passes emit nothing
26. **DeferredToolPromotionAuditMiddleware** - Observes final `tool_search` `Command`s; keep it outer of `SkillToolPolicyMiddleware` so denied names are excluded. It atomically claims new names per lead run or subagent execution to dedupe parallel searches, derives subagent attribution from the server-installed recorder, returns the original `Command`, and omits private payloads
27. **DeferredToolFilterMiddleware** - *(optional, if `tool_search.enabled`)* Hides deferred (MCP) tool schemas from the bound model until `tool_search` or `McpRoutingMiddleware` promotes them (reads per-thread promotions from `ThreadState.promoted`, hash-scoped)
28. **SystemMessageCoalescingMiddleware** - Merges every SystemMessage into a single leading SystemMessage per request; provider-agnostic fix for strict backends (vLLM/SGLang/Qwen/Anthropic) that reject non-leading system messages. Touches the per-request payload only (checkpoint state unchanged); on midnight crossings only the latest `dynamic_context_reminder` SystemMessage survives. The subagent builder places its date-only context middleware immediately before this coalescer, so the built-in subagent prompt and hidden date reminder still reach providers as one leading system block
29. **SubagentLimitMiddleware** - *(optional, if `subagent_enabled`)* Truncates excess ordinary `task` tool calls to enforce both the per-response concurrency limit (`max_concurrent_subagents`, resolved against startup `subagent_runtime.max_running` and the 1-64 safety range before construction) and the per-run total delegation cap (`max_total_subagents` runtime override or `subagents.max_total_per_run`, default 6, clamped to 1-50). The total cap counts current-run entries in the durable delegation ledger (entries are tagged with `run_id` when captured), so repeated planning checkpoints in one run cannot keep launching legal-sized batches indefinitely, while later user turns in the same thread get a fresh run budget. Explicit durable `batch_task` calls are a separate mode with persisted total/live/running limits and are not rewritten into ordinary ledger entries. If the ordinary cap is exhausted, the middleware strips remaining `task` calls, forces `finish_reason="stop"`, and appends a visible limit note so the run can synthesize existing results instead of ending with an empty tool-call response.
30. **LoopDetectionMiddleware** - *(optional, if `loop_detection.enabled`)* Detects repeated tool-call loops; hard-stop clears structured, raw, and content-block tool calls before forcing a final text answer; stamps `loop_capped` via `consume_stop_reason` (#3875 Phase 2), symmetric to `TokenBudgetMiddleware`; persists warned-state transitions (first per call hash or per tool-frequency burst) and hard stops as `middleware:loop_detection`, attributed with `is_subagent` and the optional `agent_id`, without tool arguments, message content, tool results, or argument-derived hashes. Ordinary task subagents get dedicated recorder keys through a parent-loop proxy; never pass `RunJournal` into their isolated loop. Durable batch subagents have no parent run journal and do not persist these transitions
   State is run-scoped: new user runs get fresh budgets; same-run goal
   continuations share history. Keep sibling warnings isolated and lifecycle
   hooks topology-stable. Before changing this guard, read
   [Loop detection lifecycle](../../../../../docs/LOOP_DETECTION.md) for
   fallback identity, cleanup/LRU/reset, severity ordering, and test invariants.
31. **TokenBudgetMiddleware** - `token_budget.enabled`: shares run-ID budgets across continuations; missing/invalid IDs clear invocation state.
32. **Custom middlewares** - *(optional)* Any `custom_middlewares` passed to `build_middlewares` are injected here, before config-declared extensions and the terminal-response/safety/clarification tail
33. **Configured extension middlewares** - `extensions.middlewares` in `config.yaml` or `extensions_config.json` optionally accepts `module.path:ClassName` strings or `{class, kwargs}` objects. `deerflow.reflection.resolve_class` loads `AgentMiddleware` classes; import, class, and constructor errors fail agent creation. `kwargs` must be JSON-compatible; YAML dates/timestamps become ISO strings. Order: built-ins/custom and loop/token guards → extensions → terminal-response/safety/clarification tail. Subagents share the list before their safety tail; separate lead/subagent lists are unsupported. Trusted operator config only: paths instantiate arbitrary code. Gateway skill/MCP toggles preserve it in raw JSON; adding an API write path requires explicit trust-boundary review.
34. **TerminalResponseMiddleware** - When a provider returns an empty terminal `AIMessage` after tool execution, injects a hidden recovery prompt and retries the model once; a second empty response is replaced in checkpoint state by a visible error fallback marked for the run worker, so the run finishes as an error instead of a silent success
35. **ModelLengthFinishReasonMiddleware** - Match stamps `stop_reason=model_length_capped` and ends tool loop. Suppresses calls, appends notice even with partial text, and stamps `model_length_termination` so downstream guards stand down. Preserves content blocks.
36. **SafetyFinishReasonMiddleware** - *(optional, if `safety_finish_reason.enabled`)* Suppresses tool execution when the provider safety-terminated the response (e.g. `finish_reason=content_filter`); registered after terminal-response/custom/configured middlewares so LangChain's reverse-order `after_model` dispatch runs it first
37. **ClarificationMiddleware** - Intercepts `ask_clarification`, writes a readable `ToolMessage.content` fallback plus a structured `ToolMessage.artifact.human_input` payload, and interrupts via `Command(goto=END)` (must be last). `after_model` drops same-turn sibling tool calls so they cannot run before the user answers; a malformed `ask_clarification` parked on `invalid_tool_calls` is the same stop signal. `disable_clarification` runs keep the siblings. Payloads are versioned — legacy `free_text`/`choice_with_other` stay `version: 1`; the v2 `form` mode (from `fields`) is `version: 2` so older frontends reject it and fall back to plain text. Field normalization is deterministic and lives in the middleware (it short-circuits before tool execution, so tool-arg typing gives no runtime validation), and it is atomic: any structurally broken entry — non-dict, bad/duplicate name, a name colliding with a JS `Object.prototype` member (`__proto__`/`constructor`), or exceeding the caps (16 fields / 24 options per field / 200 chars per text / `MAX_FORM_SERIALIZED_BYTES` = 16KB UTF-8, the per-item caps alone admitting forms whose IM text fallback overruns channel limits) — degrades the whole form to the legacy option/free-text modes, so a card never renders "complete" while missing a field. Benign issues degrade locally (unknown types — incl. unhashable JSON like `type: []`, which must not raise from the membership probe — and option-less selects become `text`); options are trimmed/deduped with blanks dropped (form- and top-level) since the frontend rejects blank labels. XML-to-dict option payloads are recursively flattened from dict/list containers in source order, scalar leaves kept, residual XML tags stripped before that trimming. Checkboxes are booleans defaulting to "no"; `required` on one means consent semantics. The response protocol is unchanged (v1 `text`/`option`): form cards submit a text summary as `response_kind: "text"`, so journal persistence needs no new allowlist entries. Because this middleware can short-circuit before `on_tool_end`, `RunJournal` does a root-run reconciliation for `ToolMessage`s whose `tool_call_id` came from the current run, so cards survive checkpoint compaction. That reconciliation is **not** `ask_clarification`-only — any middleware that answers a tool call has the same gap, and a result the user saw must not vanish on reload (#4666 — `ReadBeforeWriteMiddleware` blocked-write errors reached the UI but not the event store). It is bounded by three conditions, not a name allowlist: the message is user-visible, the call belongs to this run's **lead agent** (`_remember_current_run_tool_calls` records lead-agent calls only; subagent results stay in `subagent.step`), and it is not already persisted. Human Input Card replies are `hide_from_ui` `HumanMessage`s with `additional_kwargs.human_input_response`; `RunJournal` persists only allowlisted hidden sources (currently `ask_clarification`) as `llm.human.input`.