agentx-dev 3.3.0__tar.gz → 3.3.1__tar.gz

{agentx_dev-3.3.0 → agentx_dev-3.3.1}/CHANGELOG.md

@@ -1,566 +1,644 @@
-# Changelog
-
-All notable changes to `agentx-dev` are documented here. Format loosely
-follows [Keep a Changelog](https://keepachangelog.com/); versioning is
-[Semver](https://semver.org/).
-
-## [3.3.0] - 2026-08-19
-
-Dependency DAGs for the Supervisor. Plans declare which steps consume
-which, and the executor derives ordering, parallelism, AND context
-routing from those edges — unifying the old split where sequential
-mode had threading but no parallelism and concurrent mode had
-parallelism but no threading. Design: docs/design/3.3-depends-on-dag.md.
-
-### Added
-
-- **`depends_on` plan steps.** Every plan step now carries an `id`;
-  a step that consumes an earlier step's output lists that id in
-  `depends_on`. Sync `Supervisor` executes in stable topological
-  order; `AsyncSupervisor` runs a completion-driven scheduler where a
-  step starts the MOMENT its dependencies finish (not on wave
-  barriers) and independent steps run concurrently. In DAG mode each
-  step is threaded ONLY its direct dependencies' results — explicit
-  routing instead of "everything prior", which also stops the
-  per-entry context budget shrinking as plans grow.
-
-- **Plan sanitization that never fails a run.** Missing/duplicate ids
-  auto-assigned, unknown dependencies dropped (a planner typo degrades
-  to a root step, not a dead run), self-deps dropped, cycles broken
-  deterministically (back-edge in plan order), spawn steps cannot be
-  depended on. If repairs were needed, the plan is re-requested once
-  (`max_plan_retries`, default 1) with the repair warnings appended;
-  the sanitized original is kept when the retry is no better.
-
-- **Failure cascade + `skipped` flag.** A step whose dependency failed
-  (after retries / success-check) is skipped, transitively, with
-  `SubtaskResult.skipped=True` and an error naming the failed dep.
-  Independent branches keep running; synthesis runs over what
-  succeeded. Skipped steps never dispatch — no tokens burned
-  downstream of garbage.
-
-- **`skip_when` conditional execution.** A step may declare
-  `{"step": <direct dep id>, "field": <typed output field>, "is": <value>}`;
-  evaluated in Python (no LLM call) against the dependency's 3.2
-  structured output, dotted paths supported, strictly FAIL-OPEN (any
-  doubt → the step runs). Condition-skips do NOT cascade — dependents
-  treat them as empty successes ("retrieval unnecessary" is not
-  "answering impossible").
-
-- **`Specialist` registry entries.** `agents={}` now also accepts
-  `Specialist(description, runner, depends_on=[...names...],
-  output_schema=..., when_to_use=...)`. The extras render into the
-  planning catalog (`typically after:` / `returns: Schema(fields)` /
-  `use when:`) so the planner can write real dependency graphs and
-  `skip_when` conditions against actual field names. `depends_on`
-  here is a planner HINT, never an execution constraint (the same
-  specialist can appear twice in one plan; step-ids disambiguate).
-  Classic `(description, runner)` tuples keep working — they are
-  wrapped internally, and `Specialist` tuple-unpacks for older code.
-
-- **`AsyncSupervisor(max_parallel=N)`.** Caps concurrent sub-tasks for
-  rate-limited deployments. `sequential=True` is now sugar for
-  `max_parallel=1` with all-prior threading.
-
-- **`SubtaskResult.step_id` / `.depends_on` / `.skipped`** and a
-  `step_id` field on `dispatch` / `subtask_result` stream events.
-
-### Backward compatibility
-
-A plan where NO step declares `depends_on` runs with byte-identical
-legacy semantics: sync + async-sequential thread all prior results in
-plan order; async-concurrent runs everything at once with no
-threading. Verified by regression tests against the 3.2 behaviour.
-
-## [3.2.0] - 2026-08-13
-
-Typed multi-agent pipelines. Specialists can now declare a Pydantic
-output schema once and pass validated instances to each other through
-the Supervisor, instead of downstream agents re-parsing prose.
-
-### Added
-
-- **`output_schema` on the `AgentRunner` / `AsyncAgentRunner`
-  constructor.** Declare the runner's output shape once
-  (`AgentRunner(..., output_schema=QueryIntent)`) instead of passing it
-  on every call or describing JSON in the prompt. A per-call
-  `output_schema=` still wins when both are set. `None` keeps the exact
-  pre-3.2 behaviour: no coercion, `completion.output` stays `None`.
-
-- **Schema coercion via forced native function calling.** When a schema
-  is in play, the final answer is converted by forcing a provider-native
-  tool call against the schema (constrained decoding), not by regexing
-  JSON out of prose. The ReAct loop itself is untouched: tool selection
-  and intermediate reasoning run exactly as before, and the coercion
-  happens once, after the loop finishes. Models without a
-  `call_with_tools` implementation fall back to the previous text-JSON
-  parsing, so custom `BaseChatModel` subclasses keep working.
-  `completion.content` keeps the human-readable answer alongside
-  `completion.output` in every case.
-
-- **`SubtaskResult.output`.** The Supervisor now preserves each
-  specialist's validated Pydantic instance next to its `content` text.
-  Consumers that only read `content` are unaffected.
-
-- **Structured specialist-to-specialist handoff.** When an earlier step
-  produced typed output, `_build_augmented_query` serializes it into the
-  next specialist's context as a labelled JSON block
-  (`STRUCTURED OUTPUT (QueryIntent): {...}`) followed by the summary
-  text, so downstream steps parse fields rather than interpreting
-  sentences like `INTENT: ... SEARCH_QUERY: ...`.
-
-- **`vector_search_tool` pipeline options.** New kwargs:
-  `max_text_chars` (default 500; pass `0` for full untruncated passages,
-  which a reranker judging evidence actually needs) and
-  `structured_output` (default False; when True the tool returns a JSON
-  array of `{id, text, vector_score, metadata}` instead of the
-  human-formatted list). Defaults preserve existing behaviour byte-for-
-  byte.
-
-### Fixed
-
-- **Supervisor planning prompt contradicted the execution engine.** The
-  planner rule said sub-agents "do NOT see previous steps' output" and
-  discouraged dependency chains, but the dispatcher has threaded prior
-  findings into every step since `_build_augmented_query` shipped.
-  The rule now tells the planner that sequential steps receive earlier
-  results (structured when available) and that chains like
-  intent -> retrieval -> reranking are a good plan shape, while still
-  requiring same-specialist steps to merge and banning report-only steps.
-
-## [3.1.7] — 2026-07-27
-
-### Changed
-
-- **`use_function_calling` default flipped to auto-detect** on
-  `AgentRunner` / `AsyncAgentRunner`. The parameter's default type is
-  now `Optional[bool] = None`; `None` resolves to `True` when the
-  model class overrides `BaseChatModel.call_with_tools` (both `GPT`
-  and `Claude` do) and to `False` when it doesn't (or when
-  `bind_tools_natively=True`). Callers passing `True`/`False`
-  explicitly are unaffected. Rationale: text-mode ReAct requires the
-  model to emit strict JSON with any long `action_input` string
-  properly escaped — a 1200-word markdown draft with unescaped
-  newlines or quotes reliably breaks `json.loads` and killed the run.
-  Function-calling mode routes the parser through the SDK's typed
-  channel so escaping is handled automatically. The historical
-  default (`False`) was the fragile option; the new default matches
-  what most users actually want.
-
-### Fixed
-
-- **Malformed parser JSON no longer crashes the run.** When the
-  text-mode assistant response failed `json.loads` (typically because
-  a long `action_input` string had unescaped `"`, `\n`, or backticks),
-  the framework used to raise `JSONDecodeError` and unwind the whole
-  invocation. The runner now (1) tries a regex-based salvage that
-  extracts `{Thought, action, action_input}` from the raw text
-  covering the common "outer envelope valid, inner string broke
-  escaping" failure, and (2) if salvage fails, feeds a targeted fix
-  hint back to the model (`"your last response was not valid JSON;
-  emit …, escape newlines as \n"`) and continues the loop bounded
-  by `max_iterations`. Exhaustion returns a clear framework message
-  rather than an uncaught exception. Applied to both sync and async
-  runners via a shared `_salvage_react_json` helper.
-  The salvager's action-name regex is intentionally strict
-  (`[A-Za-z_][A-Za-z0-9_.\- ]{0,79}`) so it can't hallucinate an
-  "action" out of an unrelated `"key":"value"` pair inside malformed
-  JSON.
-
-- **Verbose trace in `bind_tools_natively` mode now prints tool
-  name + args + response.** Previously native runs showed blank
-  `[tool.call.start]` / `[tool.call.complete]` pairs (the
-  observability layer fires them without the trace context), so you
-  couldn't tell which tool the model actually invoked or what came
-  back. The runner now prints `[tool] Invoking '<name>' with args:
-  <input>` and `[tool] Response: <preview>` (or `[tool] Error: ...`
-  when the dispatch raised) in the post-dispatch loop, matching the
-  format text-mode and function-calling mode use. Mirrored to the
-  async runner.
-
-- **`web_fetch_tool(vector_store=...)` auto-ingests fetched pages into
-  a vector store** instead of dumping raw HTML into the model's
-  context. Fixes the TPM-limit trap: when a research agent fetches
-  four articles in parallel (via ``multi_tool_use.parallel`` or
-  native binding), the combined bodies can easily exceed 40k tokens
-  and blow past a 30k TPM ceiling on the very next model call.
-  New parameters on ``web_fetch_tool``:
-
-  | Kwarg | Default | Effect |
-  |---|---|---|
-  | ``vector_store`` | ``None`` | When set, each fetch is HTML-stripped, chunked with ``TextSplitter``, embedded via the store's embeddings, and added with ``{src: url, chunk_index, total_chunks}`` metadata. The tool response becomes a compact summary (URL, byte count, chunk count, 240-char preview) — NOT the raw body. The model then calls ``vector_search`` / ``Rag`` to pull only the passages it needs. |
-  | ``chunk_size`` | ``1500`` | Characters per chunk when ``vector_store`` is set. Ignored otherwise. |
-  | ``chunk_overlap`` | ``200`` | Overlap between adjacent chunks so a fact spanning a boundary is still retrievable. Ignored otherwise. |
-
-  Backwards-compatible: the positional ``cache_dir`` signature keeps
-  working; `web_fetch_tool()` with no ``vector_store`` returns raw
-  body as before. Ingest and cache_dir compose — enable both and get
-  disk-cached full bodies AND searchable chunks. HTML stripping is
-  minimal and dependency-free (regex-based: script/style blocks
-  dropped whole, then tags stripped, whitespace collapsed) so the
-  ingest path adds no new install dependency. On JSON/plain-text
-  responses the stripper is a near no-op.
-
-  The observation returned to the model shows topical coverage --
-  first, middle, and last chunk previews (up to 3 samples,
-  deduplicated for short pages) -- so the model can tell what
-  topics the page actually covers, not just the intro paragraph.
-  Without this the model would only see the page's opening and
-  wouldn't know to query for topics discussed later in the same
-  page. Explicit instruction in the observation ("query with
-  SPECIFIC keywords from the topics above; do NOT re-fetch; do
-  NOT ask for the full body") steers the model toward the RAG path
-  on follow-up turns.
-
-- **`multi_tool_use.parallel` now reaches its dispatch path.**
-  When GPT wanted to batch several tool calls into one turn (fetch N
-  URLs concurrently, run M searches at once), it emitted OpenAI's
-  synthetic `multi_tool_use.parallel` meta-tool. The registry's
-  `_dispatch_multi_parallel` / `_adispatch_multi_parallel` handlers
-  already knew how to unpack it, but the runner loop's known-tools
-  guardrail rejected the name FIRST as unregistered — dumping the
-  raw `{"tool_uses": [...]}` payload into the user-facing "final
-  answer" and never invoking any of the nested calls. Added
-  `multi_tool_use.parallel` to the recognized action set in both
-  sync and async runners so the meta-tool flows through to dispatch
-  and the existing unpackers run. Nested calls with the `functions.`
-  prefix are normalized before dispatch (same as top-level FC
-  calls), so the model can emit either shape.
-
-- **`Permissions.full_access` / `read_only` auto-wrap a bare string.**
-  Passing `full_access("./workspace")` used to iterate the string
-  into 11 single-character "subtrees" (Python's `list("./workspace")`)
-  — every path check silently rejected because no real path could
-  ever match a `"."` or `"/"` "allowed subtree". The classmethod
-  now detects a bare string and treats it as `[allowed_paths]`, so
-  `full_access("./workspace")` does the intuitive thing (equivalent
-  to `full_access(["./workspace"])` and auto-infers the workspace).
-  Same fix on `read_only`. List inputs are unchanged.
-
-- **`Permissions.full_access` now accepts (and auto-infers)
-  `workspace`.** The classmethod set `allowed_paths` but not
-  `workspace`, so short paths like `write_file(path="report.md")`
-  resolved to CWD (outside the sandbox) and raised
-  `PermissionError: access denied` — a landmine that every caller of
-  `Permissions.full_access(["./workspace"])` hit sooner or later.
-  New signature: `full_access(allowed_paths, *, workspace=None)`.
-  When `workspace` isn't passed AND `allowed_paths` has exactly one
-  entry, that path is auto-set as the workspace (the "project-scoped
-  agent whose one allowed subtree IS its workspace" case, which is
-  99% of use). Two or more paths stay ambiguous and require an
-  explicit `workspace=` if short-path resolution is wanted. Pass an
-  explicit `workspace=` string to override the auto-choice.
-  Backwards-compatible on the positional signature; adds a keyword
-  argument that existing callers didn't use.
-
-## [3.1.5] — 2026-07-26
-
-### Fixed
-
-- **Text-mode tool results no longer use `role: "function"`.** In text
-  mode (the default — no `use_function_calling`) the runner fed each tool
-  observation back to the model as a `role: "function"` message. Newer
-  OpenAI models reject that role outright (`400 … 'messages[N].role' does
-  not support 'function' with this model`, e.g. gpt-5.x), and Anthropic
-  never accepted it — text-mode multi-tool runs on Claude were latently
-  broken too; older GPT models simply still tolerated the legacy role.
-  Tool observations now go back as a plain `role: "user"` turn framed as
-  `Observation: …`, which every provider and model generation accepts and
-  which matches the ReAct template's own few-shot convention.
-  Function-calling mode is unchanged (native `role: "tool"` +
-  `tool_call_id`). The async runner was additionally emitting `function`
-  unconditionally (even in FC mode); it now uses the same shared helper.
-
-### Added
-
-- **`subtask_success_check` on `Supervisor` / `AsyncSupervisor`.** Opt-in
-  predicate `(SubtaskResult) -> bool | str` that decides whether a
-  *returned* (non-raised) sub-task result is actually acceptable — the
-  "ran fine but produced nothing useful" case a plain retry can't catch
-  (a scraper that saved 0 links, an extractor that found nothing). Return
-  `True` to accept, or `False`/a `str` reason to reject; a rejected
-  result is retried like a raised error, with the reason fed back into the
-  query, bounded by `max_subtask_retries`. After retries are exhausted the
-  last result is returned with its `error` set (content preserved). A
-  check that itself raises is treated as "accept" so a buggy predicate
-  can't wedge the run. Default `None` keeps the exceptions-only behavior.
-  New example `examples/robust_link_scraper.py` wires it together with a
-  scraping `system_addendum` (parse relative+absolute hrefs, fall back to
-  `sitemap.xml` on JS-rendered sites).
-
-## [3.1.4] — 2026-07-26
-
-### Fixed
-
-- **A malformed-JSON tool argument no longer crashes the whole agent
-  run.** When a model emitted a Python snippet or a Windows path as a
-  tool-call argument — `re.findall(r'\d+')`, `C:\Users` — the `\d` / `\U`
-  are illegal JSON escapes, and the OpenAI adapter's eager
-  `json.loads(call.function.arguments)` raised `JSONDecodeError` and
-  `raise`d it, unwinding the entire ReAct loop before the agent's own
-  retry machinery could act. Under a Supervisor this surfaced as a bare
-  `ERROR: Invalid \escape: line 1 column 598` and the sub-task was
-  abandoned. Now:
-    - `_parse_tool_arguments` repairs the common case (backslashes that
-      don't begin a valid JSON escape are doubled), recovering `\d`,
-      `\w`, `\s`, etc. with zero extra round-trips. A backslash before a
-      valid-escape letter (`\b`, `\n`, …) remains ambiguous and is left
-      as the escape — a documented limit.
-    - When repair fails, `call_with_tools` returns a dedicated
-      `invalid_tool_args` result and the loop feeds the error back as a
-      retryable observation ("your arguments weren't valid JSON — escape
-      backslashes and resend"), bounded by `max_iterations`, in all
-      three modes across both `AgentRunner` and `AsyncAgentRunner`.
-    `Claude` was already immune (its tool inputs arrive pre-parsed).
-
-- **A tool-call preamble is no longer returned as the final answer.**
-  Models routinely end a turn with an announcement instead of an action
-  — "I'll look up your recent scores to get a clear view of your
-  communication skills. Just a second!" — and every "no tool call
-  found" branch in both runners was coded as *this text is the answer,
-  break*. The loop terminated on iteration 1 and the caller got a
-  promise instead of a result. Three sites per runner were affected:
-  the native-binding path (`type != "tool_use"`), the
-  `use_function_calling` path (parser unresolved), and the JSON-text
-  path (response didn't parse). `max_iterations` never helped, because
-  the break happened before any iteration was spent.
-
-  The runner now feeds the model one corrective nudge — "your last turn
-  had no action, so nothing happened; do it, don't announce it" — and
-  continues the loop. Verified against both `AgentRunner` and
-  `AsyncAgentRunner` in all three modes.
-
-### Added
-
-- **Proactive "act, don't announce" system-prompt clause.** The reactive
-  `text_turn_nudges` fix corrects an agent *after* it narrates instead of
-  acting; this clause heads it off. When (and only when) an agent has
-  tools, its system prompt now tells it to call the tool rather than
-  reply "I'll do X / just a second" and stop — and to report what it DID
-  in past tense. Injected in all three modes across both runners; skipped
-  for tool-less chat agents, where prose is the correct answer. Sits
-  before any `system_addendum` so a caller's role instructions still win.
-
-- **`max_subtask_retries` on `Supervisor` / `AsyncSupervisor`** (default
-  `1`). A sub-task that raised used to be recorded as an error and the
-  Supervisor moved straight to synthesis — no second attempt. Now a
-  failed sub-task is re-dispatched up to this many times, with the prior
-  error appended to the query so the specialist knows what to fix
-  ("your previous attempt failed with X — diagnose and try again").
-  Bounded and informed: only raised exceptions trigger a retry (a
-  sub-task that returns content is accepted as-is, since the Supervisor
-  can't tell "terse but correct" from "wrong"), and the error text is
-  fed back rather than blindly re-running. Set to `0` for the old
-  quit-on-first-failure behavior. Applies in both sequential and
-  concurrent async modes.
-
-- **`text_turn_nudges` on `AgentRunner` / `AsyncAgentRunner`** (default
-  `1`). Caps the re-prompts described above at one extra LLM call per
-  run; after the budget is spent the model's text stands as the answer.
-  Set to `0` for the previous behavior. Automatically skipped when no
-  tools are registered, since a runner with no tools is a plain chat
-  call and prose genuinely is the answer there.
-
-## [3.1.3] — 2026-07-22
-
-Docs-only patch. No code changes since 3.1.2. Users on 3.1.2 don't
-need to upgrade for functionality; upgrade to pick up the improved
-onboarding docs bundled in the sdist.
-
-### Documentation
-
-- **Tools doc rewritten to answer "how do I actually use these?"**
-  Added §0 `How each built-in tool is registered` as the entry
-  section. Two registration paths — auto vs manual — laid out in a
-  table on the first screen. Six runnable subsections covering every
-  combination:
-    - §0.1 DefaultTools via `Permissions(...)` (auto)
-    - §0.2 WebTools via `tools=[web_search_tool(), web_fetch_tool()]`
-    - §0.3 RAG via `TextSplitter` -> `VectorStore.add_documents` ->
-      `vector_search_tool(store)`
-    - §0.4 Handoffs via `handoff_tool` + `HandoffCoordinator`
-    - §0.5 Custom `StructuredTool` from scratch
-    - §0.6 Fully-loaded runner combining all of the above
-    - §0.7 Rules on name collisions, invisible-denied-capabilities,
-      async-tool behavior
-  The existing inventory + wrapper / controls / cheat-sheet sections
-  are unchanged; they now sit after the "how to use them" primer
-  instead of before it.
-
-## [3.1.2] — 2026-07-22
-
-Patch release. Two independent fixes.
-
-### Fixed
-
-- **`llm_judge` correctly parses YES/NO across providers.** The judge
-  parser was comparing the reply's first word to the literal string
-  `"YES"`. GPT-4o answers `"YES,"` (comma-suffixed), which failed the
-  equality check and marked every genuine PASS as FAIL. Claude replies
-  `"YES"` without punctuation so the bug hid during local development.
-  Fixed by matching `\b(YES|NO)\b` (word-boundary regex, case-
-  insensitive) at the start of the reply. Handles every real shape:
-  `YES`, `YES.`, `YES!`, `YES, exactly right`, `Yes.`, `yes -- reason`.
-  Ambiguous replies (`Maybe`, empty string) still fail closed.
-- Regression test `test_llm_judge_parses_various_verdict_shapes`
-  covers 8 YES shapes, 5 NO shapes, and 4 ambiguous replies.
-
-### Added
-
-- **`agentx_dev.Tools` is a one-stop tools namespace.** Users no
-  longer need to remember which module each tool lives in:
-
-  ```python
-  from agentx_dev.Tools import (
-      StandardTool, StructuredTool,
-      AsyncStandardTool, AsyncStructuredTool,
-      web_search_tool, web_fetch_tool,
-      vector_search_tool, handoff_tool,
-      DefaultTools, Permissions,
-  )
-  ```
-
-  Both this form and the pre-existing `from agentx_dev import X`
-  form coexist. Implementation uses PEP 562 module `__getattr__`
-  and `__dir__` so re-exports are lazy (no import cost for modules
-  the caller doesn't touch) and show up in IDE autocomplete +
-  `dir(agentx_dev.Tools)`.
-
-## [3.1.1] — 2026-07-21
-
-Second batch of 3.1 features + a full docs + brand pass.
-
-### Added
-
-**Streaming through orchestration**
-- `Supervisor.stream()` / `AsyncSupervisor.astream()` emit
-  `plan_start` / `plan` / `dispatch` / `subtask_result` /
-  `synthesize_start` / `final` / `completion` events.
-- `HandoffCoordinator.stream()` / `.astream()` emit `invoke` /
-  `completion` / `handoff` / `final` / `result` events per hop.
-- Legacy `.run()` / `.arun()` refactored to consume the streams (no
-  code duplication).
-
-**Prompt optimization — `Compiled`**
-- New `agentx_dev.Compiler` module.
-- `Compiled(runner_factory, trainset, ...)` iteratively refines a
-  runner's `system_addendum` against the eval harness. Half of
-  DSPy's power at a tenth of the surface.
-
-**Anthropic Batch API**
-- `Claude.batch(requests)` submits many prompts at Anthropic's 50%-off
-  batch rate, polls to completion, returns results in submission order.
-- Per-request error dicts on failure; token usage funneled into
-  `TokenUsage` so cost tracking stays a single source of truth.
-
-**Vector store adapters — `agentx_dev.VectorStores`**
-- `ChromaVectorStore`, `QdrantVectorStore`, `PgVectorStore` — same
-  public shape as the in-memory `VectorStore` (`add` / `search` /
-  `delete` / `clear` / `__len__` / `embeddings`).
-- `vector_search_tool()` and `SemanticMemory` accept any of them.
-- SDK imports lazy; friendly `ImportError` when the underlying SDK
-  is missing.
-
-**Trace viewer (`viewer/`)**
-- Self-hosted single-page app that reads `FileHook` JSONL and renders
-  a timeline with type/text filters, summary sidebar, JSON drill-down.
-- Works from `file://`, no server required.
-
-**Docs site (`host/`)**
-- Full editorial dark-first design system (JetBrains Mono headings,
-  Inter body, `#B8FF3E` electric-lime accent).
-- Command palette (`Cmd+K`) with keyboard navigation and live search.
-- Hero code snippet with hand-tinted syntax highlighting.
-- Reading progress bar, breadcrumbs, header anchor links.
-- Sidebar sliding active marker, collapsible groups.
-- Code copy buttons, language labels.
-- Right-rail auto-TOC with `IntersectionObserver` scrollspy.
-- Dark/light theme toggle, persisted.
-- Cache-busted assets so edits land on refresh without hard-reload.
-
-**Brand identity (`brand/`)**
-- Full brand kit: 5 SVG assets (`mark`, `mono`, `wordmark`, `logo-full`,
-  `app-icon`), `BRAND.md` strategy doc, rendered brand-kit HTML deck.
-- Copy audit dropped "small" (weak) and "LangChain" references from
-  all marketing surfaces.
-- Favicon wired into docs + trace viewer.
-
-**Test suite (`tests/`)**
-- Restored + expanded pytest suite: 127 tests passing (3 skipped for
-  absent optional SDKs).
-- Coverage: parser + all `AgentType` variants, `ToolRegistry`
-  (dispatch / dup-guard / circuit-breaker / timeout), Permissions
-  (capability gating + sandbox + traversal), budgets (cost / rate /
-  retry / non-retryable HTTP), runner loop (streaming + output_schema
-  + chat history), embeddings + `VectorStore` + `SemanticMemory`,
-  handoffs (bounded hops + history sanitization), evals harness
-  (all assertion helpers + JSON case loaders), vector-store adapter
-  shape conformance.
-
-**Docs (`docs/`)**
-- Full docs tree (34 pages), including new pages for:
-  vector store adapters, prompt optimization, batch API, trace viewer,
-  and a **use-cases** landing (13 concrete scenarios with runnable code).
-- Rewrote **Tools** page to enumerate every built-in tool with args,
-  return shape, capability flag, and use-case guidance.
-- Rewrote **Agents** page to cover all four orchestration
-  architectures (Solo / Supervisor / Handoffs / Compiled) with
-  decision trees, worked examples, and cheat sheet.
-- **Agentic RAG chatbot** as use case §13 — multi-query decomposition,
-  parallel retrieval, self-critique, citations, user memory.
-
-**Examples**
-- `examples/agentic_rag_demo.py` — the runnable version of the
-  agentic RAG use case. Auto-seeds a KB if none exists, `--demo` flag
-  runs a 3-turn scripted session proving user-notes recall works.
-
-**Package**
-- `[chroma]`, `[qdrant]`, `[pgvector]`, `[dev]` extras added.
-- `[anthropic]` bumped to `>=0.36` (Batch API + prompt cache).
-
-### Fixed
-
-- `AgentRunner._iter_run` in `bind_tools_natively=True` mode uses a
-  minimal system prompt instead of the AgentType template so the
-  ReAct `action/action_input` scaffold no longer fights the native
-  tool interface. Previously produced JSON-blob answers under GPT.
-- `HandoffCoordinator._sanitize_history_for_next_agent` strips tool
-  and function role messages between hops so tool_call_ids from a
-  previous agent don't leak into the next model's call (OpenAI 400).
-- Docs site marker positioning uses double-`requestAnimationFrame` +
-  `document.fonts.ready` so the sidebar accent bar lands on the
-  correct row even on a cold font cache.
-- Primary hero CTA color uses `#doc .hero-cta a.primary` selector to
-  outrank `#doc a` link styling (previously rendered lime-on-lime
-  and was invisible).
-
-### Notes
-
-- Package version bumped from `3.0.6` to `3.1.1`. The 3.1.0 release
-  did not ship publicly — 3.1.1 is the first 3.1-tagged PyPI release
-  and includes both batches of features.
-
-## [3.1.0] — internal only (commits 52840e7)
-
-First batch of 3.1 features. Committed but not released to PyPI.
-Merged into 3.1.1 for the public release.
-
-### Added
-- Anthropic prompt caching (`Claude(enable_prompt_cache=True)`).
-- Parallel per-turn tool dispatch in `AgentRunner`
-  (`bind_tools_natively=True`, `parallel_tool_workers`).
-- Semantic memory (`SemanticMemory`, embeddings-backed retrieval).
-- RAG core (`Embeddings`, `HashEmbeddings`, `OpenAIEmbeddings`,
-  `VectorStore`, `VectorHit`, `vector_search_tool()`).
-- Agent-to-agent handoffs (`HandoffRequest`, `handoff_tool`,
-  `HandoffCoordinator`, `HandoffResult`).
-- Evals harness (`EvalCase`, `EvalRunner`, `EvalReport`, 7 assertion
-  helpers, JSON case loader, `python -m agentx_dev.Evals run` CLI).
-- `TokenUsage.cache_hit_ratio` property.
-
-## [3.0.6] — 2026-03 (baseline)
-
-Security hardening baseline (SSRF guard on `web_fetch`, HMAC-signed
-persistent state, scrubbed subprocess env, path sanitizer,
-`permissions.json` mode 0o600, ReDoS guard on `grep`,
-`invoke`/`ainvoke` accept bare strings and message lists).
+# Changelog
+
+All notable changes to `agentx-dev` are documented here. Format loosely
+follows [Keep a Changelog](https://keepachangelog.com/); versioning is
+[Semver](https://semver.org/).
+
+## [3.3.1] - 2026-09-09
+
+Correctness fixes in the agent loop, found by driving it with a
+scripted model rather than reading it. No API changes; every fix
+replaces a silent wrong answer or a runaway cost with correct
+behaviour. Nine regression tests ship with them
+(`tests/test_agent_loop_flaws.py`).
+
+### Fixed
+
+- **Native binding had no loop-level spiral guard.** The repeat breaker
+  sat after the native branch's `continue`, so `bind_tools_natively`
+  runs never reached it: a model stuck re-issuing one call burned the
+  full `max_iterations` (20 LLM turns where text mode stopped at 3) and
+  returned the "hit max_iterations" recap quoting the dup-guard's own
+  warning text instead of the data.
+
+- **`respond` batched with real tool calls dropped them.** A model that
+  emitted "do X, and here is my answer" in one turn had X silently
+  discarded -- never dispatched, never in `completion.tool_calls`, no
+  error. The batch now runs and the answer defers one turn.
+
+- **`completion.history` could not be replayed as `chat_history`.** The
+  copy filter kept only truthy `{role, content}`, which dropped every
+  tool-calling assistant turn (`content=""`), stripped `tool_call_id`
+  off `role="tool"` messages, and replayed a stored system prompt on top
+  of the fresh one -- so a follow-up call sent two system messages, zero
+  assistant turns, and an orphaned tool message providers reject. Fixed
+  in both runners.
+
+- **A sync `AgentRunner` silently ignored async tools.** `known_tools`
+  unioned only the two sync registry tables while the registry accepted,
+  listed and prompt-advertised async ones, so calling one fell through
+  to implicit-final and returned `action_input` as the answer with an
+  empty `tool_calls` list. Now unions all four, matching the async
+  runner; the sync dispatcher already returns a clear `ToolError`, so
+  the silent wrong answer became an actionable one.
+
+- **`AsyncAgentRunner` regressed the function-calling message shape.**
+  It appended its FC turn as raw JSON text with no `tool_calls` block
+  and never set `_last_function_call_id`, so observations went back as
+  `role="user"` -- text-mode shape while in FC mode, losing provider
+  correlation and the cached prefix.
+
+- **A `BaseException` from a specialist orphaned its siblings.**
+  `CancelledError` and `KeyboardInterrupt` are not `Exception`, so
+  `_run_subtask`'s handler never saw them; `t.result()` re-raised and
+  left sibling tasks running unowned, with in-flight LLM calls still
+  billing. The scheduler now owns its tasks in a `try`/`finally`, which
+  also covers a consumer that stops iterating `astream` early.
+
+- **The tool cache collided on tool name.** It is a process-wide
+  singleton keyed on `(tool_name, args)` with no record of which
+  implementation ran, so two runners whose tools merely share a name --
+  `search`, `fetch`, `query`, routine across Supervisor specialists --
+  served each other's results and the second function never ran. Keys
+  now fold in the callable's `module.qualname`, so same-name /
+  different-implementation misses while genuinely identical tools still
+  share, including across processes for the disk-backed `FileCache`.
+
+### Docs
+
+- Streaming documentation described parameters the API rejects.
+  `AsyncAgentRunner.astream()` and `HandoffCoordinator.stream()` take no
+  `stream_tokens` (only the sync `AgentRunner.stream()` does), and every
+  `text_delta` example was built on a default model where
+  `use_function_calling` auto-detects to `True` and yields zero deltas.
+  `simple_stream` was documented with the wrong signature entirely.
+  Added a table naming exactly which stream methods accept what.
+
+### Tests
+
+- Repaired 20 stale text-mode tests. `MockModel` defined
+  `call_with_tools` unconditionally, so 3.1.7's auto-detect routed every
+  text script down the function-calling path and the runner returned
+  `""`. The mock now advertises the capability only when scripted for
+  it. Suite: 206 passed, 3 skipped.
+
+## [3.3.0] - 2026-08-19
+
+Dependency DAGs for the Supervisor. Plans declare which steps consume
+which, and the executor derives ordering, parallelism, AND context
+routing from those edges — unifying the old split where sequential
+mode had threading but no parallelism and concurrent mode had
+parallelism but no threading. Design: docs/design/3.3-depends-on-dag.md.
+
+### Added
+
+- **`depends_on` plan steps.** Every plan step now carries an `id`;
+  a step that consumes an earlier step's output lists that id in
+  `depends_on`. Sync `Supervisor` executes in stable topological
+  order; `AsyncSupervisor` runs a completion-driven scheduler where a
+  step starts the MOMENT its dependencies finish (not on wave
+  barriers) and independent steps run concurrently. In DAG mode each
+  step is threaded ONLY its direct dependencies' results — explicit
+  routing instead of "everything prior", which also stops the
+  per-entry context budget shrinking as plans grow.
+
+- **Plan sanitization that never fails a run.** Missing/duplicate ids
+  auto-assigned, unknown dependencies dropped (a planner typo degrades
+  to a root step, not a dead run), self-deps dropped, cycles broken
+  deterministically (back-edge in plan order), spawn steps cannot be
+  depended on. If repairs were needed, the plan is re-requested once
+  (`max_plan_retries`, default 1) with the repair warnings appended;
+  the sanitized original is kept when the retry is no better.
+
+- **Failure cascade + `skipped` flag.** A step whose dependency failed
+  (after retries / success-check) is skipped, transitively, with
+  `SubtaskResult.skipped=True` and an error naming the failed dep.
+  Independent branches keep running; synthesis runs over what
+  succeeded. Skipped steps never dispatch — no tokens burned
+  downstream of garbage.
+
+- **`skip_when` conditional execution.** A step may declare
+  `{"step": <direct dep id>, "field": <typed output field>, "is": <value>}`;
+  evaluated in Python (no LLM call) against the dependency's 3.2
+  structured output, dotted paths supported, strictly FAIL-OPEN (any
+  doubt → the step runs). Condition-skips do NOT cascade — dependents
+  treat them as empty successes ("retrieval unnecessary" is not
+  "answering impossible").
+
+- **`Specialist` registry entries.** `agents={}` now also accepts
+  `Specialist(description, runner, depends_on=[...names...],
+  output_schema=..., when_to_use=...)`. The extras render into the
+  planning catalog (`typically after:` / `returns: Schema(fields)` /
+  `use when:`) so the planner can write real dependency graphs and
+  `skip_when` conditions against actual field names. `depends_on`
+  here is a planner HINT, never an execution constraint (the same
+  specialist can appear twice in one plan; step-ids disambiguate).
+  Classic `(description, runner)` tuples keep working — they are
+  wrapped internally, and `Specialist` tuple-unpacks for older code.
+
+- **`AsyncSupervisor(max_parallel=N)`.** Caps concurrent sub-tasks for
+  rate-limited deployments. `sequential=True` is now sugar for
+  `max_parallel=1` with all-prior threading.
+
+- **`SubtaskResult.step_id` / `.depends_on` / `.skipped`** and a
+  `step_id` field on `dispatch` / `subtask_result` stream events.
+
+### Backward compatibility
+
+A plan where NO step declares `depends_on` runs with byte-identical
+legacy semantics: sync + async-sequential thread all prior results in
+plan order; async-concurrent runs everything at once with no
+threading. Verified by regression tests against the 3.2 behaviour.
+
+## [3.2.0] - 2026-08-13
+
+Typed multi-agent pipelines. Specialists can now declare a Pydantic
+output schema once and pass validated instances to each other through
+the Supervisor, instead of downstream agents re-parsing prose.
+
+### Added
+
+- **`output_schema` on the `AgentRunner` / `AsyncAgentRunner`
+  constructor.** Declare the runner's output shape once
+  (`AgentRunner(..., output_schema=QueryIntent)`) instead of passing it
+  on every call or describing JSON in the prompt. A per-call
+  `output_schema=` still wins when both are set. `None` keeps the exact
+  pre-3.2 behaviour: no coercion, `completion.output` stays `None`.
+
+- **Schema coercion via forced native function calling.** When a schema
+  is in play, the final answer is converted by forcing a provider-native
+  tool call against the schema (constrained decoding), not by regexing
+  JSON out of prose. The ReAct loop itself is untouched: tool selection
+  and intermediate reasoning run exactly as before, and the coercion
+  happens once, after the loop finishes. Models without a
+  `call_with_tools` implementation fall back to the previous text-JSON
+  parsing, so custom `BaseChatModel` subclasses keep working.
+  `completion.content` keeps the human-readable answer alongside
+  `completion.output` in every case.
+
+- **`SubtaskResult.output`.** The Supervisor now preserves each
+  specialist's validated Pydantic instance next to its `content` text.
+  Consumers that only read `content` are unaffected.
+
+- **Structured specialist-to-specialist handoff.** When an earlier step
+  produced typed output, `_build_augmented_query` serializes it into the
+  next specialist's context as a labelled JSON block
+  (`STRUCTURED OUTPUT (QueryIntent): {...}`) followed by the summary
+  text, so downstream steps parse fields rather than interpreting
+  sentences like `INTENT: ... SEARCH_QUERY: ...`.
+
+- **`vector_search_tool` pipeline options.** New kwargs:
+  `max_text_chars` (default 500; pass `0` for full untruncated passages,
+  which a reranker judging evidence actually needs) and
+  `structured_output` (default False; when True the tool returns a JSON
+  array of `{id, text, vector_score, metadata}` instead of the
+  human-formatted list). Defaults preserve existing behaviour byte-for-
+  byte.
+
+### Fixed
+
+- **Supervisor planning prompt contradicted the execution engine.** The
+  planner rule said sub-agents "do NOT see previous steps' output" and
+  discouraged dependency chains, but the dispatcher has threaded prior
+  findings into every step since `_build_augmented_query` shipped.
+  The rule now tells the planner that sequential steps receive earlier
+  results (structured when available) and that chains like
+  intent -> retrieval -> reranking are a good plan shape, while still
+  requiring same-specialist steps to merge and banning report-only steps.
+
+## [3.1.7] — 2026-07-27
+
+### Changed
+
+- **`use_function_calling` default flipped to auto-detect** on
+  `AgentRunner` / `AsyncAgentRunner`. The parameter's default type is
+  now `Optional[bool] = None`; `None` resolves to `True` when the
+  model class overrides `BaseChatModel.call_with_tools` (both `GPT`
+  and `Claude` do) and to `False` when it doesn't (or when
+  `bind_tools_natively=True`). Callers passing `True`/`False`
+  explicitly are unaffected. Rationale: text-mode ReAct requires the
+  model to emit strict JSON with any long `action_input` string
+  properly escaped — a 1200-word markdown draft with unescaped
+  newlines or quotes reliably breaks `json.loads` and killed the run.
+  Function-calling mode routes the parser through the SDK's typed
+  channel so escaping is handled automatically. The historical
+  default (`False`) was the fragile option; the new default matches
+  what most users actually want.
+
+### Fixed
+
+- **Malformed parser JSON no longer crashes the run.** When the
+  text-mode assistant response failed `json.loads` (typically because
+  a long `action_input` string had unescaped `"`, `\n`, or backticks),
+  the framework used to raise `JSONDecodeError` and unwind the whole
+  invocation. The runner now (1) tries a regex-based salvage that
+  extracts `{Thought, action, action_input}` from the raw text
+  covering the common "outer envelope valid, inner string broke
+  escaping" failure, and (2) if salvage fails, feeds a targeted fix
+  hint back to the model (`"your last response was not valid JSON;
+  emit …, escape newlines as \n"`) and continues the loop bounded
+  by `max_iterations`. Exhaustion returns a clear framework message
+  rather than an uncaught exception. Applied to both sync and async
+  runners via a shared `_salvage_react_json` helper.
+  The salvager's action-name regex is intentionally strict
+  (`[A-Za-z_][A-Za-z0-9_.\- ]{0,79}`) so it can't hallucinate an
+  "action" out of an unrelated `"key":"value"` pair inside malformed
+  JSON.
+
+- **Verbose trace in `bind_tools_natively` mode now prints tool
+  name + args + response.** Previously native runs showed blank
+  `[tool.call.start]` / `[tool.call.complete]` pairs (the
+  observability layer fires them without the trace context), so you
+  couldn't tell which tool the model actually invoked or what came
+  back. The runner now prints `[tool] Invoking '<name>' with args:
+  <input>` and `[tool] Response: <preview>` (or `[tool] Error: ...`
+  when the dispatch raised) in the post-dispatch loop, matching the
+  format text-mode and function-calling mode use. Mirrored to the
+  async runner.
+
+- **`web_fetch_tool(vector_store=...)` auto-ingests fetched pages into
+  a vector store** instead of dumping raw HTML into the model's
+  context. Fixes the TPM-limit trap: when a research agent fetches
+  four articles in parallel (via ``multi_tool_use.parallel`` or
+  native binding), the combined bodies can easily exceed 40k tokens
+  and blow past a 30k TPM ceiling on the very next model call.
+  New parameters on ``web_fetch_tool``:
+
+  | Kwarg | Default | Effect |
+  |---|---|---|
+  | ``vector_store`` | ``None`` | When set, each fetch is HTML-stripped, chunked with ``TextSplitter``, embedded via the store's embeddings, and added with ``{src: url, chunk_index, total_chunks}`` metadata. The tool response becomes a compact summary (URL, byte count, chunk count, 240-char preview) — NOT the raw body. The model then calls ``vector_search`` / ``Rag`` to pull only the passages it needs. |
+  | ``chunk_size`` | ``1500`` | Characters per chunk when ``vector_store`` is set. Ignored otherwise. |
+  | ``chunk_overlap`` | ``200`` | Overlap between adjacent chunks so a fact spanning a boundary is still retrievable. Ignored otherwise. |
+
+  Backwards-compatible: the positional ``cache_dir`` signature keeps
+  working; `web_fetch_tool()` with no ``vector_store`` returns raw
+  body as before. Ingest and cache_dir compose — enable both and get
+  disk-cached full bodies AND searchable chunks. HTML stripping is
+  minimal and dependency-free (regex-based: script/style blocks
+  dropped whole, then tags stripped, whitespace collapsed) so the
+  ingest path adds no new install dependency. On JSON/plain-text
+  responses the stripper is a near no-op.
+
+  The observation returned to the model shows topical coverage --
+  first, middle, and last chunk previews (up to 3 samples,
+  deduplicated for short pages) -- so the model can tell what
+  topics the page actually covers, not just the intro paragraph.
+  Without this the model would only see the page's opening and
+  wouldn't know to query for topics discussed later in the same
+  page. Explicit instruction in the observation ("query with
+  SPECIFIC keywords from the topics above; do NOT re-fetch; do
+  NOT ask for the full body") steers the model toward the RAG path
+  on follow-up turns.
+
+- **`multi_tool_use.parallel` now reaches its dispatch path.**
+  When GPT wanted to batch several tool calls into one turn (fetch N
+  URLs concurrently, run M searches at once), it emitted OpenAI's
+  synthetic `multi_tool_use.parallel` meta-tool. The registry's
+  `_dispatch_multi_parallel` / `_adispatch_multi_parallel` handlers
+  already knew how to unpack it, but the runner loop's known-tools
+  guardrail rejected the name FIRST as unregistered — dumping the
+  raw `{"tool_uses": [...]}` payload into the user-facing "final
+  answer" and never invoking any of the nested calls. Added
+  `multi_tool_use.parallel` to the recognized action set in both
+  sync and async runners so the meta-tool flows through to dispatch
+  and the existing unpackers run. Nested calls with the `functions.`
+  prefix are normalized before dispatch (same as top-level FC
+  calls), so the model can emit either shape.
+
+- **`Permissions.full_access` / `read_only` auto-wrap a bare string.**
+  Passing `full_access("./workspace")` used to iterate the string
+  into 11 single-character "subtrees" (Python's `list("./workspace")`)
+  — every path check silently rejected because no real path could
+  ever match a `"."` or `"/"` "allowed subtree". The classmethod
+  now detects a bare string and treats it as `[allowed_paths]`, so
+  `full_access("./workspace")` does the intuitive thing (equivalent
+  to `full_access(["./workspace"])` and auto-infers the workspace).
+  Same fix on `read_only`. List inputs are unchanged.
+
+- **`Permissions.full_access` now accepts (and auto-infers)
+  `workspace`.** The classmethod set `allowed_paths` but not
+  `workspace`, so short paths like `write_file(path="report.md")`
+  resolved to CWD (outside the sandbox) and raised
+  `PermissionError: access denied` — a landmine that every caller of
+  `Permissions.full_access(["./workspace"])` hit sooner or later.
+  New signature: `full_access(allowed_paths, *, workspace=None)`.
+  When `workspace` isn't passed AND `allowed_paths` has exactly one
+  entry, that path is auto-set as the workspace (the "project-scoped
+  agent whose one allowed subtree IS its workspace" case, which is
+  99% of use). Two or more paths stay ambiguous and require an
+  explicit `workspace=` if short-path resolution is wanted. Pass an
+  explicit `workspace=` string to override the auto-choice.
+  Backwards-compatible on the positional signature; adds a keyword
+  argument that existing callers didn't use.
+
+## [3.1.5] — 2026-07-26
+
+### Fixed
+
+- **Text-mode tool results no longer use `role: "function"`.** In text
+  mode (the default — no `use_function_calling`) the runner fed each tool
+  observation back to the model as a `role: "function"` message. Newer
+  OpenAI models reject that role outright (`400 … 'messages[N].role' does
+  not support 'function' with this model`, e.g. gpt-5.x), and Anthropic
+  never accepted it — text-mode multi-tool runs on Claude were latently
+  broken too; older GPT models simply still tolerated the legacy role.
+  Tool observations now go back as a plain `role: "user"` turn framed as
+  `Observation: …`, which every provider and model generation accepts and
+  which matches the ReAct template's own few-shot convention.
+  Function-calling mode is unchanged (native `role: "tool"` +
+  `tool_call_id`). The async runner was additionally emitting `function`
+  unconditionally (even in FC mode); it now uses the same shared helper.
+
+### Added
+
+- **`subtask_success_check` on `Supervisor` / `AsyncSupervisor`.** Opt-in
+  predicate `(SubtaskResult) -> bool | str` that decides whether a
+  *returned* (non-raised) sub-task result is actually acceptable — the
+  "ran fine but produced nothing useful" case a plain retry can't catch
+  (a scraper that saved 0 links, an extractor that found nothing). Return
+  `True` to accept, or `False`/a `str` reason to reject; a rejected
+  result is retried like a raised error, with the reason fed back into the
+  query, bounded by `max_subtask_retries`. After retries are exhausted the
+  last result is returned with its `error` set (content preserved). A
+  check that itself raises is treated as "accept" so a buggy predicate
+  can't wedge the run. Default `None` keeps the exceptions-only behavior.
+  New example `examples/robust_link_scraper.py` wires it together with a
+  scraping `system_addendum` (parse relative+absolute hrefs, fall back to
+  `sitemap.xml` on JS-rendered sites).
+
+## [3.1.4] — 2026-07-26
+
+### Fixed
+
+- **A malformed-JSON tool argument no longer crashes the whole agent
+  run.** When a model emitted a Python snippet or a Windows path as a
+  tool-call argument — `re.findall(r'\d+')`, `C:\Users` — the `\d` / `\U`
+  are illegal JSON escapes, and the OpenAI adapter's eager
+  `json.loads(call.function.arguments)` raised `JSONDecodeError` and
+  `raise`d it, unwinding the entire ReAct loop before the agent's own
+  retry machinery could act. Under a Supervisor this surfaced as a bare
+  `ERROR: Invalid \escape: line 1 column 598` and the sub-task was
+  abandoned. Now:
+    - `_parse_tool_arguments` repairs the common case (backslashes that
+      don't begin a valid JSON escape are doubled), recovering `\d`,
+      `\w`, `\s`, etc. with zero extra round-trips. A backslash before a
+      valid-escape letter (`\b`, `\n`, …) remains ambiguous and is left
+      as the escape — a documented limit.
+    - When repair fails, `call_with_tools` returns a dedicated
+      `invalid_tool_args` result and the loop feeds the error back as a
+      retryable observation ("your arguments weren't valid JSON — escape
+      backslashes and resend"), bounded by `max_iterations`, in all
+      three modes across both `AgentRunner` and `AsyncAgentRunner`.
+    `Claude` was already immune (its tool inputs arrive pre-parsed).
+
+- **A tool-call preamble is no longer returned as the final answer.**
+  Models routinely end a turn with an announcement instead of an action
+  — "I'll look up your recent scores to get a clear view of your
+  communication skills. Just a second!" — and every "no tool call
+  found" branch in both runners was coded as *this text is the answer,
+  break*. The loop terminated on iteration 1 and the caller got a
+  promise instead of a result. Three sites per runner were affected:
+  the native-binding path (`type != "tool_use"`), the
+  `use_function_calling` path (parser unresolved), and the JSON-text
+  path (response didn't parse). `max_iterations` never helped, because
+  the break happened before any iteration was spent.
+
+  The runner now feeds the model one corrective nudge — "your last turn
+  had no action, so nothing happened; do it, don't announce it" — and
+  continues the loop. Verified against both `AgentRunner` and
+  `AsyncAgentRunner` in all three modes.
+
+### Added
+
+- **Proactive "act, don't announce" system-prompt clause.** The reactive
+  `text_turn_nudges` fix corrects an agent *after* it narrates instead of
+  acting; this clause heads it off. When (and only when) an agent has
+  tools, its system prompt now tells it to call the tool rather than
+  reply "I'll do X / just a second" and stop — and to report what it DID
+  in past tense. Injected in all three modes across both runners; skipped
+  for tool-less chat agents, where prose is the correct answer. Sits
+  before any `system_addendum` so a caller's role instructions still win.
+
+- **`max_subtask_retries` on `Supervisor` / `AsyncSupervisor`** (default
+  `1`). A sub-task that raised used to be recorded as an error and the
+  Supervisor moved straight to synthesis — no second attempt. Now a
+  failed sub-task is re-dispatched up to this many times, with the prior
+  error appended to the query so the specialist knows what to fix
+  ("your previous attempt failed with X — diagnose and try again").
+  Bounded and informed: only raised exceptions trigger a retry (a
+  sub-task that returns content is accepted as-is, since the Supervisor
+  can't tell "terse but correct" from "wrong"), and the error text is
+  fed back rather than blindly re-running. Set to `0` for the old
+  quit-on-first-failure behavior. Applies in both sequential and
+  concurrent async modes.
+
+- **`text_turn_nudges` on `AgentRunner` / `AsyncAgentRunner`** (default
+  `1`). Caps the re-prompts described above at one extra LLM call per
+  run; after the budget is spent the model's text stands as the answer.
+  Set to `0` for the previous behavior. Automatically skipped when no
+  tools are registered, since a runner with no tools is a plain chat
+  call and prose genuinely is the answer there.
+
+## [3.1.3] — 2026-07-22
+
+Docs-only patch. No code changes since 3.1.2. Users on 3.1.2 don't
+need to upgrade for functionality; upgrade to pick up the improved
+onboarding docs bundled in the sdist.
+
+### Documentation
+
+- **Tools doc rewritten to answer "how do I actually use these?"**
+  Added §0 `How each built-in tool is registered` as the entry
+  section. Two registration paths — auto vs manual — laid out in a
+  table on the first screen. Six runnable subsections covering every
+  combination:
+    - §0.1 DefaultTools via `Permissions(...)` (auto)
+    - §0.2 WebTools via `tools=[web_search_tool(), web_fetch_tool()]`
+    - §0.3 RAG via `TextSplitter` -> `VectorStore.add_documents` ->
+      `vector_search_tool(store)`
+    - §0.4 Handoffs via `handoff_tool` + `HandoffCoordinator`
+    - §0.5 Custom `StructuredTool` from scratch
+    - §0.6 Fully-loaded runner combining all of the above
+    - §0.7 Rules on name collisions, invisible-denied-capabilities,
+      async-tool behavior
+  The existing inventory + wrapper / controls / cheat-sheet sections
+  are unchanged; they now sit after the "how to use them" primer
+  instead of before it.
+
+## [3.1.2] — 2026-07-22
+
+Patch release. Two independent fixes.
+
+### Fixed
+
+- **`llm_judge` correctly parses YES/NO across providers.** The judge
+  parser was comparing the reply's first word to the literal string
+  `"YES"`. GPT-4o answers `"YES,"` (comma-suffixed), which failed the
+  equality check and marked every genuine PASS as FAIL. Claude replies
+  `"YES"` without punctuation so the bug hid during local development.
+  Fixed by matching `\b(YES|NO)\b` (word-boundary regex, case-
+  insensitive) at the start of the reply. Handles every real shape:
+  `YES`, `YES.`, `YES!`, `YES, exactly right`, `Yes.`, `yes -- reason`.
+  Ambiguous replies (`Maybe`, empty string) still fail closed.
+- Regression test `test_llm_judge_parses_various_verdict_shapes`
+  covers 8 YES shapes, 5 NO shapes, and 4 ambiguous replies.
+
+### Added
+
+- **`agentx_dev.Tools` is a one-stop tools namespace.** Users no
+  longer need to remember which module each tool lives in:
+
+  ```python
+  from agentx_dev.Tools import (
+      StandardTool, StructuredTool,
+      AsyncStandardTool, AsyncStructuredTool,
+      web_search_tool, web_fetch_tool,
+      vector_search_tool, handoff_tool,
+      DefaultTools, Permissions,
+  )
+  ```
+
+  Both this form and the pre-existing `from agentx_dev import X`
+  form coexist. Implementation uses PEP 562 module `__getattr__`
+  and `__dir__` so re-exports are lazy (no import cost for modules
+  the caller doesn't touch) and show up in IDE autocomplete +
+  `dir(agentx_dev.Tools)`.
+
+## [3.1.1] — 2026-07-21
+
+Second batch of 3.1 features + a full docs + brand pass.
+
+### Added
+
+**Streaming through orchestration**
+- `Supervisor.stream()` / `AsyncSupervisor.astream()` emit
+  `plan_start` / `plan` / `dispatch` / `subtask_result` /
+  `synthesize_start` / `final` / `completion` events.
+- `HandoffCoordinator.stream()` / `.astream()` emit `invoke` /
+  `completion` / `handoff` / `final` / `result` events per hop.
+- Legacy `.run()` / `.arun()` refactored to consume the streams (no
+  code duplication).
+
+**Prompt optimization — `Compiled`**
+- New `agentx_dev.Compiler` module.
+- `Compiled(runner_factory, trainset, ...)` iteratively refines a
+  runner's `system_addendum` against the eval harness. Half of
+  DSPy's power at a tenth of the surface.
+
+**Anthropic Batch API**
+- `Claude.batch(requests)` submits many prompts at Anthropic's 50%-off
+  batch rate, polls to completion, returns results in submission order.
+- Per-request error dicts on failure; token usage funneled into
+  `TokenUsage` so cost tracking stays a single source of truth.
+
+**Vector store adapters — `agentx_dev.VectorStores`**
+- `ChromaVectorStore`, `QdrantVectorStore`, `PgVectorStore` — same
+  public shape as the in-memory `VectorStore` (`add` / `search` /
+  `delete` / `clear` / `__len__` / `embeddings`).
+- `vector_search_tool()` and `SemanticMemory` accept any of them.
+- SDK imports lazy; friendly `ImportError` when the underlying SDK
+  is missing.
+
+**Trace viewer (`viewer/`)**
+- Self-hosted single-page app that reads `FileHook` JSONL and renders
+  a timeline with type/text filters, summary sidebar, JSON drill-down.
+- Works from `file://`, no server required.
+
+**Docs site (`host/`)**
+- Full editorial dark-first design system (JetBrains Mono headings,
+  Inter body, `#B8FF3E` electric-lime accent).
+- Command palette (`Cmd+K`) with keyboard navigation and live search.
+- Hero code snippet with hand-tinted syntax highlighting.
+- Reading progress bar, breadcrumbs, header anchor links.
+- Sidebar sliding active marker, collapsible groups.
+- Code copy buttons, language labels.
+- Right-rail auto-TOC with `IntersectionObserver` scrollspy.
+- Dark/light theme toggle, persisted.
+- Cache-busted assets so edits land on refresh without hard-reload.
+
+**Brand identity (`brand/`)**
+- Full brand kit: 5 SVG assets (`mark`, `mono`, `wordmark`, `logo-full`,
+  `app-icon`), `BRAND.md` strategy doc, rendered brand-kit HTML deck.
+- Copy audit dropped "small" (weak) and "LangChain" references from
+  all marketing surfaces.
+- Favicon wired into docs + trace viewer.
+
+**Test suite (`tests/`)**
+- Restored + expanded pytest suite: 127 tests passing (3 skipped for
+  absent optional SDKs).
+- Coverage: parser + all `AgentType` variants, `ToolRegistry`
+  (dispatch / dup-guard / circuit-breaker / timeout), Permissions
+  (capability gating + sandbox + traversal), budgets (cost / rate /
+  retry / non-retryable HTTP), runner loop (streaming + output_schema
+  + chat history), embeddings + `VectorStore` + `SemanticMemory`,
+  handoffs (bounded hops + history sanitization), evals harness
+  (all assertion helpers + JSON case loaders), vector-store adapter
+  shape conformance.
+
+**Docs (`docs/`)**
+- Full docs tree (34 pages), including new pages for:
+  vector store adapters, prompt optimization, batch API, trace viewer,
+  and a **use-cases** landing (13 concrete scenarios with runnable code).
+- Rewrote **Tools** page to enumerate every built-in tool with args,
+  return shape, capability flag, and use-case guidance.
+- Rewrote **Agents** page to cover all four orchestration
+  architectures (Solo / Supervisor / Handoffs / Compiled) with
+  decision trees, worked examples, and cheat sheet.
+- **Agentic RAG chatbot** as use case §13 — multi-query decomposition,
+  parallel retrieval, self-critique, citations, user memory.
+
+**Examples**
+- `examples/agentic_rag_demo.py` — the runnable version of the
+  agentic RAG use case. Auto-seeds a KB if none exists, `--demo` flag
+  runs a 3-turn scripted session proving user-notes recall works.
+
+**Package**
+- `[chroma]`, `[qdrant]`, `[pgvector]`, `[dev]` extras added.
+- `[anthropic]` bumped to `>=0.36` (Batch API + prompt cache).
+
+### Fixed
+
+- `AgentRunner._iter_run` in `bind_tools_natively=True` mode uses a
+  minimal system prompt instead of the AgentType template so the
+  ReAct `action/action_input` scaffold no longer fights the native
+  tool interface. Previously produced JSON-blob answers under GPT.
+- `HandoffCoordinator._sanitize_history_for_next_agent` strips tool
+  and function role messages between hops so tool_call_ids from a
+  previous agent don't leak into the next model's call (OpenAI 400).
+- Docs site marker positioning uses double-`requestAnimationFrame` +
+  `document.fonts.ready` so the sidebar accent bar lands on the
+  correct row even on a cold font cache.
+- Primary hero CTA color uses `#doc .hero-cta a.primary` selector to
+  outrank `#doc a` link styling (previously rendered lime-on-lime
+  and was invisible).
+
+### Notes
+
+- Package version bumped from `3.0.6` to `3.1.1`. The 3.1.0 release
+  did not ship publicly — 3.1.1 is the first 3.1-tagged PyPI release
+  and includes both batches of features.
+
+## [3.1.0] — internal only (commits 52840e7)
+
+First batch of 3.1 features. Committed but not released to PyPI.
+Merged into 3.1.1 for the public release.
+
+### Added
+- Anthropic prompt caching (`Claude(enable_prompt_cache=True)`).
+- Parallel per-turn tool dispatch in `AgentRunner`
+  (`bind_tools_natively=True`, `parallel_tool_workers`).
+- Semantic memory (`SemanticMemory`, embeddings-backed retrieval).
+- RAG core (`Embeddings`, `HashEmbeddings`, `OpenAIEmbeddings`,
+  `VectorStore`, `VectorHit`, `vector_search_tool()`).
+- Agent-to-agent handoffs (`HandoffRequest`, `handoff_tool`,
+  `HandoffCoordinator`, `HandoffResult`).
+- Evals harness (`EvalCase`, `EvalRunner`, `EvalReport`, 7 assertion
+  helpers, JSON case loader, `python -m agentx_dev.Evals run` CLI).
+- `TokenUsage.cache_hit_ratio` property.
+
+## [3.0.6] — 2026-03 (baseline)
+
+Security hardening baseline (SSRF guard on `web_fetch`, HMAC-signed
+persistent state, scrubbed subprocess env, path sanitizer,
+`permissions.json` mode 0o600, ReDoS guard on `grep`,
+`invoke`/`ainvoke` accept bare strings and message lists).