hypermind 2.2.1__tar.gz → 2.3.0__tar.gz

{hypermind-2.2.1 → hypermind-2.3.0}/.gitignore

@@ -152,8 +152,8 @@ experiments/
 site/
 
 # Local MCP config (may contain real tokens — setup.sh writes these from user input)
-config/mcp.toml
-config/mcp.json
+**/config/mcp.toml
+**/config/mcp.json
 
 # -----------------------------------------------------------------------------
 # Rust port (rust/) — build artefacts only; Cargo.lock IS committed (workspace

{hypermind-2.2.1 → hypermind-2.3.0}/CHANGELOG.md

@@ -14,14 +14,48 @@ changes only on major-version bumps.
 
 ## [Unreleased]
 
+## [2.3.0] — 2026-07-31
+
+Bundles the four epics from the "swarm can research a topic end-to-end"
+effort: web page/PDF discovery gaps, storage alignment with RelataDB, and
+making PI-decomposed swarms genuinely goal/objective-driven. Additive,
+backward-compatible — all new parameters default to prior behavior.
+
 ### Added
-- **Single-container streamable-HTTP MCP gateway (#1013).** `deploy/Dockerfile.mcp-http`
-  + `deploy/docker-compose.mcp-http.yml` front `hmctl mcp serve` with
-  [`supergateway`](https://github.com/supercorp-ai/supergateway), exposing a
-  plain streamable-HTTP `/mcp` endpoint for platform integrations (gateway-style
-  tool discovery services) that don't speak raw TCP/SSE. Safe-by-default posture
-  matches `deploy/Dockerfile.mcp`: write mode on by default, `MCP_WRITE=0` for
-  read-only. Documented in `deploy/README.md`.
+- **PDF fetch + site crawl for the research swarm (#1018).** `fetch_page`'s
+  Content-Type: application/pdf is no longer silently skipped/garbled — it
+  routes through pypdf extraction, shared with the local-file ingest path
+  (`hypermind.ingest._extract_pdf_bytes`). New `crawl_site` tool
+  (`orgs/web_crawl.py`): SSRF-validated BFS crawl of same-domain links,
+  hard-capped at 15 pages / depth 4 (default 5/2), covers HTML and PDFs.
+- **RelataDB `document_ingest` alignment (#1019).** `hypermind.ingest` gained
+  `ExtractedChunk`/`DocumentManifest`/`chunk_to_manifest()`/`to_jsonl()` —
+  a shared chunk/manifest shape matching RelataDB's datagrep extractor
+  protocol (section_path breadcrumbs, page numbers, chunking metadata).
+  `StorageBackend` gained `ingest_document()`/`search_chunks()`:
+  `RelataBackend` wraps the `relata` SDK's native `RelataClient.ingest_document()`;
+  `SqliteBackend` mirrors it locally (`intel_reports`/`intel_chunks` tables +
+  FTS5 index) so ingested documents are searchable without a Relata server.
+  `OrgSim(ingest_via_manifest=True)` (opt-in, default off) lands
+  `ingest_dir=` files this way in addition to the existing claim-publish path.
+- **Goal-driven charter-decomposed swarms (#1020).** `charter.py` gained
+  `plan_target_confidences()`; `OrgSim(unit_goal_confidence=...)` +
+  `DynamicsConfig(mode="deliberator")` let a PI-decomposed swarm
+  (`build_expert_model()`) stop via real goal satisfaction (per-sub-area
+  confidence + synthesis published) instead of only a fixed/converging round
+  count — bridges `charter.py`'s decomposition to the existing
+  `mind.goals`/`Deliberator` machinery. Fixed a latent bug in the
+  `goal_satisfied` early-exit: the coordination unit's (PI oracle +
+  Synthesist) own goal is now excluded from the AND-gate and synthesis is
+  checked via a published claim instead, since that goal has no defined
+  relationship to whether the sub-area experts finished.
+- **Intra-expert research probe gate (#1021).** `_think()` nudges (never
+  filters) the model away from re-researching a sub-area already backed by a
+  healthy density of confirmed facts, via a new `research_hint` on
+  `default_context_builder`. Tunable via `HYPERMIND_KNOWLEDGE_PROBE_THRESHOLD`
+  / `HYPERMIND_KNOWLEDGE_PROBE_MIN_FACTS`.
+
+## [2.2.1] — 2026-07-29
 
 ### Security
 - **CAP_REVOKE now gossips (#943 Phase 1).** A signed `CapRevocationRecord`
@@ -72,6 +106,13 @@ changes only on major-version bumps.
   module docstring for the full scoping rationale.
 
 ### Added
+- **Single-container streamable-HTTP MCP gateway (#1013).** `deploy/Dockerfile.mcp-http`
+  + `deploy/docker-compose.mcp-http.yml` front `hmctl mcp serve` with
+  [`supergateway`](https://github.com/supercorp-ai/supergateway), exposing a
+  plain streamable-HTTP `/mcp` endpoint for platform integrations (gateway-style
+  tool discovery services) that don't speak raw TCP/SSE. Safe-by-default posture
+  matches `deploy/Dockerfile.mcp`: write mode on by default, `MCP_WRITE=0` for
+  read-only. Documented in `deploy/README.md`.
 - **Bi-temporal recall + cross-namespace citation resolution (#942, L3
   follow-up).** `confirmed_hypotheses` entries carried no real timestamp
   before this — only `created_round`, a simulation round index, not a

{hypermind-2.2.1 → hypermind-2.3.0}/CLAUDE.md

@@ -36,7 +36,7 @@ Activate the venv first; everything assumes Python 3.12–3.14 (`requires-python
 .venv/bin/pip install -e ".[test,dev]"
 
 # Run the full suite (always run before commit)
-.venv/bin/python -m pytest -q                                      # ~4476 passing in ~90s
+.venv/bin/python -m pytest -q                                      # ~5093 passing in ~90s
 .venv/bin/python -m pytest --cov=src/hypermind --cov=src/hmctl     # coverage report
 
 # Run a single test
@@ -250,9 +250,9 @@ When changing a public verb, check that every example in `examples/` still runs
 
 ## Active milestone
 
-As of 2026-07-27, the SDK is at **2.1.5** on the **engineering GA** line (engineering GA shipped 2026-07-09 as 1.0.0; SemVer is binding). The protocol core, the full swarm-intelligence / mind layer, the `hypermind.orgs` deep-research engine (decomposition + self-refine, reasoning strategies + structured outcome parsing, hypothesis lifecycle with prediction-matching, architecture-native metrics, evidence retention + citations, cross-sim learned-fit persistence via KernelStore, streaming query, full-run replay, real Relata ledger + TCP federation), and the independent Rust second-language port (`rust/`) have all landed. Remaining gates are **ecosystem, not code** — IETF I-D submission, IANA early-allocation, OpenSSF Sandbox/Incubating acceptance, third-party security audit, full Tamarin/ProVerif proofs, NLnet/Alpha-Omega awards, ≥2 non-ZySec maintainers, ≥3 pilot MOUs, 7-day chaos-test green run, side-channel CT 30d-green. See [`docs/project/roadmap.md`](docs/project/roadmap.md) for item-numbered status.
+As of 2026-07-31, the SDK is at **2.3.0** on the **engineering GA** line (engineering GA shipped 2026-07-09 as 1.0.0; SemVer is binding). The protocol core, the full swarm-intelligence / mind layer, the `hypermind.orgs` deep-research engine (decomposition + self-refine, reasoning strategies + structured outcome parsing, hypothesis lifecycle with prediction-matching, architecture-native metrics, evidence retention + citations, cross-sim learned-fit persistence via KernelStore, streaming query, full-run replay, real Relata ledger + TCP federation, PDF/site-crawl discovery + RelataDB `document_ingest` alignment, goal-driven charter-decomposed swarms), and the independent Rust second-language port (`rust/`) have all landed. Remaining gates are **ecosystem, not code** — IETF I-D submission, IANA early-allocation, OpenSSF Sandbox/Incubating acceptance, third-party security audit, full Tamarin/ProVerif proofs, NLnet/Alpha-Omega awards, ≥2 non-ZySec maintainers, ≥3 pilot MOUs, 7-day chaos-test green run, side-channel CT 30d-green. See [`docs/project/roadmap.md`](docs/project/roadmap.md) for item-numbered status.
 
-> **Recent releases.** v2.0.0 (SemVer-major consolidation: wire-version 3 supported, deprecated-API removal, libp2p first-class transport binding, TAL v0.2 wire messages, Sealed/Citeable lattice, second-language port in progress). v2.1.0–v2.1.5 (2026-07-27): Rust port landed (`rust/`), L1–L5 autonomy stack, FTS5 storage, MoA endpoint, IEEE-defensible eval benchmark, dark-theme HTML report. The 2026-07-25 security-hardening fixes (issues #839-#844) shipped in 2.0 Phase 0. See [`CHANGELOG.md`](CHANGELOG.md) and [`docs/project/v2.0-plan.md`](docs/project/v2.0-plan.md).
+> **Recent releases.** v2.0.0 (SemVer-major consolidation: wire-version 3 supported, deprecated-API removal, libp2p first-class transport binding, TAL v0.2 wire messages, Sealed/Citeable lattice, second-language port in progress). v2.1.0–v2.1.5 (2026-07-27): Rust port landed (`rust/`), L1–L5 autonomy stack, FTS5 storage, MoA endpoint, IEEE-defensible eval benchmark, dark-theme HTML report. v2.2.0 (2026-07-28): autonomy pass (#983) — SSRF redirect/hostname hardening, deployment `profiles`, `coevolution`, `silent_failures`/`health`, `reputation_store`, `CachingResponder`/`DifficultyRouter`, hybrid SQLite search, active grounding gates (defaults changed, API-compatible). v2.2.1 (2026-07-29): CAP_REVOKE gossip + canonical-CBOR gate (#943 P1), TAL v0.2 signed wire messages (#943 P2), bi-temporal recall + cross-namespace Merkle citation proofs (#942), streamable-HTTP MCP gateway (#1013). v2.3.0 (2026-07-31): PDF/site-crawl discovery (#1018), RelataDB `document_ingest` alignment (#1019), goal-driven charter swarms (#1020), intra-expert research probe gate (#1021). The 2026-07-25 security-hardening fixes (issues #839-#844) shipped in 2.0 Phase 0. See [`CHANGELOG.md`](CHANGELOG.md) and [`docs/project/v2.0-plan.md`](docs/project/v2.0-plan.md).
 
 > **Note on doc locations (2026-07-25):** `CHARTER.md`, `GOVERNANCE.md`, `MAINTAINERS.md`, `ROADMAP.md`, and `REPO_MAP.md` were moved from the repo root into [`docs/project/`](docs/project/) (renamed lowercase). Root retains only the GitHub platform files (`README`, `LICENSE`, `SECURITY`, `CODE_OF_CONDUCT`, `CONTRIBUTING`, `CHANGELOG`, `CODEOWNERS`) plus `CLAUDE.md`. See [`docs/INDEX.md`](docs/INDEX.md) for the full doc map.
 

{hypermind-2.2.1 → hypermind-2.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: hypermind
-Version: 2.2.1
+Version: 2.3.0
 Summary: Federated agent intelligence — reference SDK for the HyperMind SCITT contested-claims profile
 Project-URL: Homepage, https://github.com/ZySec-AI/hypermind
 Project-URL: Documentation, https://github.com/ZySec-AI/hypermind/tree/main/docs
@@ -90,8 +90,8 @@ Agents share what they learn, dispute what they doubt, and build reputation over
 [![PyPI](https://img.shields.io/pypi/v/hypermind?color=4c8cff&label=pip%20install%20hypermind)](https://pypi.org/project/hypermind/)
 [![Python](https://img.shields.io/pypi/pyversions/hypermind?color=4c8cff)](https://pypi.org/project/hypermind/)
 [![License](https://img.shields.io/badge/license-Apache--2.0-green.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-4476-brightgreen)](#status)
-[![Version](https://img.shields.io/badge/version-2.1.5-blue)](CHANGELOG.md)
+[![Tests](https://img.shields.io/badge/tests-5093-brightgreen)](#status)
+[![Version](https://img.shields.io/badge/version-2.3.0-blue)](CHANGELOG.md)
 
 **[Quickstart](#quickstart-30-seconds)** ·
 **[MCP tools](#mcp-tools-63-tools)** ·
@@ -378,7 +378,7 @@ hmctl mcp install # wires Claude Desktop/Cursor + auto-adds fetch/Serper/SearchA
 
 ## Status
 
-**Engineering GA** (1.0.0, 2026-07-09; SemVer binding). Current release: **2.1.5** (2026-07-27).
+**Engineering GA** (1.0.0, 2026-07-09; SemVer binding). Current release: **2.3.0** (2026-07-31).
 Remaining gates are **ecosystem, not code** (see [roadmap](docs/project/roadmap.md) §External).
 
 - **Protocol** — 33 numbered spec sections (§00–§32) + independent [`reference_verifier/`](reference_verifier/) proving two-impl interop + second-language [`rust/`](rust/) port (v2.1.0)

{hypermind-2.2.1 → hypermind-2.3.0}/README.md

@@ -12,8 +12,8 @@ Agents share what they learn, dispute what they doubt, and build reputation over
 [![PyPI](https://img.shields.io/pypi/v/hypermind?color=4c8cff&label=pip%20install%20hypermind)](https://pypi.org/project/hypermind/)
 [![Python](https://img.shields.io/pypi/pyversions/hypermind?color=4c8cff)](https://pypi.org/project/hypermind/)
 [![License](https://img.shields.io/badge/license-Apache--2.0-green.svg)](LICENSE)
-[![Tests](https://img.shields.io/badge/tests-4476-brightgreen)](#status)
-[![Version](https://img.shields.io/badge/version-2.1.5-blue)](CHANGELOG.md)
+[![Tests](https://img.shields.io/badge/tests-5093-brightgreen)](#status)
+[![Version](https://img.shields.io/badge/version-2.3.0-blue)](CHANGELOG.md)
 
 **[Quickstart](#quickstart-30-seconds)** ·
 **[MCP tools](#mcp-tools-63-tools)** ·
@@ -300,7 +300,7 @@ hmctl mcp install # wires Claude Desktop/Cursor + auto-adds fetch/Serper/SearchA
 
 ## Status
 
-**Engineering GA** (1.0.0, 2026-07-09; SemVer binding). Current release: **2.1.5** (2026-07-27).
+**Engineering GA** (1.0.0, 2026-07-09; SemVer binding). Current release: **2.3.0** (2026-07-31).
 Remaining gates are **ecosystem, not code** (see [roadmap](docs/project/roadmap.md) §External).
 
 - **Protocol** — 33 numbered spec sections (§00–§32) + independent [`reference_verifier/`](reference_verifier/) proving two-impl interop + second-language [`rust/`](rust/) port (v2.1.0)

{hypermind-2.2.1 → hypermind-2.3.0}/docs/INDEX.md

@@ -1,7 +1,7 @@
 # HyperMind documentation index
 
 > **One-page map** of every doc in `docs/` and the GitHub platform files at the
-> repo root. Last updated 2026-07-27. Rendered site nav lives in
+> repo root. Last updated 2026-07-31. Rendered site nav lives in
 > [`mkdocs.yml`](../mkdocs.yml); this page is the source of truth for *what each
 > doc is for*.
 
@@ -287,7 +287,7 @@ End-to-end docs for the one-command Docker deployment. Target: operators, app bu
 | File | Purpose |
 |---|---|
 | [`migrate/v0.x-to-v1.0.md`](migrate/v0.x-to-v1.0.md) | v0.x → v1.0 migration guide. |
-| [`release-notes/`](release-notes/) | One file per release (`v1.0.0.md` through `v1.11.0.md`). |
+| [`release-notes/`](release-notes/) | One file per release (`v1.0.0.md` through `v2.3.0.md`). |
 | [`paper/00-patent.md`](paper/00-patent.md) | Patent disclosure. |
 | [`paper/hypermind-paper-draft.md`](paper/hypermind-paper-draft.md) | Research-paper draft. |
 | [`simlab/modes.md`](simlab/modes.md) | SimLab mode reference. |

{hypermind-2.2.1 → hypermind-2.3.0}/docs/decide/01-disqualifiers.md

@@ -15,7 +15,7 @@ You should not adopt HyperMind v0 if any of these are true:
 - **All your agents share a database or trust domain.** Use A2A, MCP, or direct RPC.
 - **You cannot afford 4 GB RAM per agent process.** Use lite-tier only (reduced capability) or a different stack.
 - **You need source identity protection for people at physical risk.** Use Tor + SecureDrop; HyperMind only as a corroboration layer afterwards.
-- **You need a working library to import today.** Reference implementation has shipped: `pip install hypermind` (current: v2.1.5 on [PyPI](https://pypi.org/project/hypermind/); engineering GA landed 2026-07-09 as v1.0.0). See [`README.md`](../../README.md) §Quickstart.
+- **You need a working library to import today.** Reference implementation has shipped: `pip install hypermind` (current: v2.3.0 on [PyPI](https://pypi.org/project/hypermind/); engineering GA landed 2026-07-09 as v1.0.0). See [`README.md`](../../README.md) §Quickstart.
 - **Your team has under 6 months of distributed-systems experience.** Prototype only; do not ship to production.
 - **You need real-time response during high-stakes adversarial events** (elections, live conflict intelligence). The ~170-claim Sybil window and timing-analysis de-anonymization are disqualifying. See [../design/03-failure-modes.md](../design/03-failure-modes.md).
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/design/01-architecture.md

@@ -1,6 +1,6 @@
 # Architecture — what HyperMind is
 
-> **Status (2026-07-27):** This doc was written against the v0.x design intent (two-layer P2P sealing protocol with diverse-quorum confirmers, identity tiers, PoW/VDF admission). The shipped SDK at v2.1.5 has evolved substantially: it is an **async-first Python framework for federated agent reasoning** organised around the layers below. The "two-layer protocol" framing here is preserved for historical context and for the design-target sealing system documented in [`docs/paper/00-patent.md`](../paper/00-patent.md) Embodiments 4–5; for the shipped architecture see [`docs/spec/00-overview.md`](../spec/00-overview.md), the [mind layer spec §20](../spec/20-swarm-intelligence.md), and the [deep-research engine spec §31](../spec/31-deep-research-engine.md).
+> **Status (2026-07-31):** This doc was written against the v0.x design intent (two-layer P2P sealing protocol with diverse-quorum confirmers, identity tiers, PoW/VDF admission). The shipped SDK at v2.3.0 has evolved substantially: it is an **async-first Python framework for federated agent reasoning** organised around the layers below. The "two-layer protocol" framing here is preserved for historical context and for the design-target sealing system documented in [`docs/paper/00-patent.md`](../paper/00-patent.md) Embodiments 4–5; for the shipped architecture see [`docs/spec/00-overview.md`](../spec/00-overview.md), the [mind layer spec §20](../spec/20-swarm-intelligence.md), and the [deep-research engine spec §31](../spec/31-deep-research-engine.md).
 >
 > Read [../decide/01-disqualifiers.md](../decide/01-disqualifiers.md) and [../decide/03-incentive-aligned-cases.md](../decide/03-incentive-aligned-cases.md) before this page. If those did not disqualify you, here is the design.
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/design/02-traces.md

@@ -1,6 +1,6 @@
 # Protocol traces — message-level walkthroughs
 
-> **Status (2026-07-27):** **Mixed shipped + design-target.** Traces 1–5 describe the **design-target** sealing flow (ANNOUNCE → CONFIRM → SEAL with diverse-quorum confirmers and 200-claim correctness windows) — none of this ships in the SDK at v2.1.5; the shipped flow is `publish_claim → append_and_cosign → receipt` per [`src/hypermind/agent.py`](../../src/hypermind/agent.py) and [`src/hypermind/knowledge/transparency.py`](../../src/hypermind/knowledge/transparency.py). Traces 6–8 reference DESIGN-PARTIAL primitives. The traces remain useful as **intent for the post-v3.0 sealed/citeable lattice** (spec §04 §5.2 invariants); they should not be read as a description of what the SDK does today.
+> **Status (2026-07-31):** **Mixed shipped + design-target.** Traces 1–5 describe the **design-target** sealing flow (ANNOUNCE → CONFIRM → SEAL with diverse-quorum confirmers and 200-claim correctness windows) — none of this ships in the SDK at v2.3.0; the shipped flow is `publish_claim → append_and_cosign → receipt` per [`src/hypermind/agent.py`](../../src/hypermind/agent.py) and [`src/hypermind/knowledge/transparency.py`](../../src/hypermind/knowledge/transparency.py). Traces 6–8 reference DESIGN-PARTIAL primitives. The traces remain useful as **intent for the post-v3.0 sealed/citeable lattice** (spec §04 §5.2 invariants); they should not be read as a description of what the SDK does today.
 
 > These walkthroughs show the exact protocol message sequence for representative use cases. Every message type maps to a defined message in the [spec](../spec/00-overview.md). Timing figures are estimates on a healthy mesh — see [../operate/03-resource-budget.md](../operate/03-resource-budget.md) for the unmeasured-estimate caveat.
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/design/03-failure-modes.md

@@ -1,6 +1,6 @@
 # Failure modes — adversaries, equilibria, and collective-intelligence pathology
 
-> **Status (2026-07-27):** **Conceptual reference; not all failure modes have shipped defenses.** This page catalogues collective-intelligence failure modes (with and without adversaries). The mechanisms it informs — diverse-quorum sealing, 200-claim correctness windows, anchor-tenant capture pricing — are **design-target** for the post-v3.0 sealed/citeable lattice (spec §04 §5.2 invariants). The shipped SDK at v2.1.5 mitigates several of these failure modes via the Beta-posterior reputation kernel, the dispute FSM, capability tokens, and the cascaded-security L1–L5 model; the unshipped mechanisms remain accurate as intent.
+> **Status (2026-07-31):** **Conceptual reference; not all failure modes have shipped defenses.** This page catalogues collective-intelligence failure modes (with and without adversaries). The mechanisms it informs — diverse-quorum sealing, 200-claim correctness windows, anchor-tenant capture pricing — are **design-target** for the post-v3.0 sealed/citeable lattice (spec §04 §5.2 invariants). The shipped SDK at v2.3.0 mitigates several of these failure modes via the Beta-posterior reputation kernel, the dispute FSM, capability tokens, and the cascaded-security L1–L5 model; the unshipped mechanisms remain accurate as intent.
 
 > Read this before designing anything around HyperMind. These failure modes are not theoretical. They are the reason the protocol has the mechanisms it has — and the reason those mechanisms are insufficient in some categories.
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/design/04-poisoning-defenses.md

@@ -1,6 +1,6 @@
 # Poisoning defenses — innovations beyond §7.4 and the 03-failure-modes inventory
 
-> **Status (2026-07-27):** **Design proposals, not shipped.** None of the six defenses below are implemented in the SDK at v2.1.5, are referenced by a spec section, or are adopted by an ADR. They are tracked as design exploration for the post-v3.0 sealed/citeable lattice.
+> **Status (2026-07-31):** **Design proposals, not shipped.** None of the six defenses below are implemented in the SDK at v2.3.0, are referenced by a spec section, or are adopted by an ADR. They are tracked as design exploration for the post-v3.0 sealed/citeable lattice.
 
 > Read [03-failure-modes.md](03-failure-modes.md) first. This page proposes **new** defenses that address attacks the existing roster does not close. Each section names the specific failure-mode row it targets.
 >

{hypermind-2.2.1 → hypermind-2.3.0}/docs/design/05-deployment-topology.md

@@ -1,6 +1,6 @@
 # Reference Deployment Topology
 
-> **Status (2026-07-27):** **Design-target / forward-looking.** This document describes the deployment topology the spec targets *once* the IPFS / Iroh / libp2p-gossipsub substrate is integrated. As of SDK v2.1.5, **none of those dependencies ship in `pyproject.toml`**. The shipped transport is **TCP federation** (`TCPGossipBus`, per [ADR-0001](../adr/0001-tcp-federation-chosen-as-v1-transport.md)); storage is **SQLite (local) + Relata (federated)** per [ADR-0009](../adr/0009-dual-storage-backend-sqlite-local-default-and-relata-federated.md); **libp2p is the optional external binding** per [ADR-0003](../adr/0003-libp2p-deferred-to-optional-external-binding.md) and lives in `transport-rust/`. Treat the topology table below as the v3.0+ target, not the v2.x production shape. For the shipped shape see [`docs/operate/01-deployment.md`](../operate/01-deployment.md) and [`docs/operate/03-resource-budget.md`](../operate/03-resource-budget.md).
+> **Status (2026-07-31):** **Design-target / forward-looking.** This document describes the deployment topology the spec targets *once* the IPFS / Iroh / libp2p-gossipsub substrate is integrated. As of SDK v2.3.0, **none of those dependencies ship in `pyproject.toml`**. The shipped transport is **TCP federation** (`TCPGossipBus`, per [ADR-0001](../adr/0001-tcp-federation-chosen-as-v1-transport.md)); storage is **SQLite (local) + Relata (federated)** per [ADR-0009](../adr/0009-dual-storage-backend-sqlite-local-default-and-relata-federated.md); **libp2p is the optional external binding** per [ADR-0003](../adr/0003-libp2p-deferred-to-optional-external-binding.md) and lives in `transport-rust/`. Treat the topology table below as the v3.0+ target, not the v2.x production shape. For the shipped shape see [`docs/operate/01-deployment.md`](../operate/01-deployment.md) and [`docs/operate/03-resource-budget.md`](../operate/03-resource-budget.md).
 
 > HyperMind composes existing P2P primitives. Nothing below the SCITT envelope is reinvented.
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/feature-matrix.md

@@ -8,9 +8,13 @@ HyperMind SDK, the `hmctl` CLI, the independent `reference_verifier/`, and the
 second-language `rust/` port. A row exists for every shipped capability. The
 project reached **engineering GA at 1.0.0 (2026-07-09)**, followed by twelve
 1.x minor releases (1.1–1.12), the **v2.0.0 SemVer-major consolidation**
-(2026-07-25), and the **v2.1.x series** (2026-07-27) adding the Rust port,
+(2026-07-25), the **v2.1.x series** (2026-07-27) adding the Rust port,
 L1–L5 autonomy stack, FTS5 storage, MoA endpoint, and the IEEE-defensible eval
-benchmark. Remaining gates are ecosystem/foundation work (see
+benchmark, the **v2.2.0 autonomy pass** (#983: profiles, coevolution,
+hybrid search, responder routing/caching, active grounding gates) +
+**v2.2.1** (CAP_REVOKE gossip, TAL v0.2 wire, bi-temporal recall), and
+**v2.3.0** (site-crawl/PDF discovery, RelataDB `document_ingest` alignment,
+goal-driven charter swarms). Remaining gates are ecosystem/foundation work (see
 [roadmap.md](project/roadmap.md) §External), not code.
 
 ## Maintenance rules
@@ -471,4 +475,4 @@ historical surface still find a row.
 
 ---
 
-*Last updated: 2026-07-27. Version: 2.1.5. Test count: 4476 (per `pytest --collect-only -q`).*
+*Last updated: 2026-07-31. Version: 2.3.0. Test count: 5093 (per `pytest --collect-only -q`).*

{hypermind-2.2.1 → hypermind-2.3.0}/docs/integrations.md

@@ -206,7 +206,7 @@ ledger — no API key, no internet required.
 
 ### MCP tool reference (56 read-only / 63 writable tools)
 
-> **Note.** The mutating set is gated by `hmctl mcp install --write`. The exact count of read-only and mutating tools shifts slightly between releases as new tools are added; the canonical counts are emitted by `hmctl mcp install --dry-run`. README.md and CLAUDE.md cite the current totals (63 in write mode, 56 in read-only mode at v2.1.5).
+> **Note.** The mutating set is gated by `hmctl mcp install --write`. The exact count of read-only and mutating tools shifts slightly between releases as new tools are added; the canonical counts are emitted by `hmctl mcp install --dry-run`. README.md and CLAUDE.md cite the current totals (63 in write mode, 56 in read-only mode at v2.3.0).
 
 **Observability**
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/operate/01-deployment.md

@@ -1,9 +1,9 @@
 # Deployment — cold-start, anchor-tenant, HMAC custody, capability tokens
 
-> **Status (2026-07-27).** **Design-target.** None of the CLI commands in this doc
+> **Status (2026-07-31).** **Design-target.** None of the CLI commands in this doc
 > (`hypermind init`, `hypermind consult`, `hypermind join`, `hypermind rebuild`)
 > are runnable as written; they describe the intended developer experience for
-> the post-v3.0 sealed/citeable lattice. The shipped SDK at v2.1.5 uses `hmctl`
+> the post-v3.0 sealed/citeable lattice. The shipped SDK at v2.3.0 uses `hmctl`
 > (see [`docs/quickstart.md`](../quickstart.md) and `hmctl --help` for the actual
 > surface — `hmctl quickstart`, `hmctl serve-endpoint`, `hmctl mcp install`,
 > `hmctl ingest`, `hmctl dataset export`, `hmctl eval`, `hmctl orgs simulate`,

{hypermind-2.2.1 → hypermind-2.3.0}/docs/operate/data-residency.md

@@ -1,6 +1,6 @@
 # Data Residency
 
-> **Status (2026-07-27):** The shipped SDK at v2.1.5 uses **async TCP federation**
+> **Status (2026-07-31):** The shipped SDK at v2.3.0 uses **async TCP federation**
 > (`TCPGossipBus`) per [ADR-0001](../adr/0001-tcp-federation-chosen-as-v1-transport.md)
 > and is single-region by default. The libp2p binding is the **optional external
 > binding** per [ADR-0003](../adr/0003-libp2p-deferred-to-optional-external-binding.md)

{hypermind-2.2.1 → hypermind-2.3.0}/docs/operate/slo.md

@@ -1,9 +1,9 @@
 # Service-Level Objectives — Production Profile
 
-> **Status (2026-07-27).** **Design-target / forward-looking.** The SLOs below
+> **Status (2026-07-31).** **Design-target / forward-looking.** The SLOs below
 > were drafted against the post-v3.0 sealed/citeable lattice design (with
 > `publish_claim` quorum, sealing pipeline, 200-claim correctness window). The
-> shipped SDK at v2.1.5 has no sealing pipeline; `publish_claim()` lands claims
+> shipped SDK at v2.3.0 has no sealing pipeline; `publish_claim()` lands claims
 > directly in the transparency log via `append_and_cosign()`, so the
 > "Publish latency p99 < 250 ms" objective is trivially met in the absence of
 > quorum traffic. The "Partition-recovery MTTR < 90 s" objective depends on

{hypermind-2.2.1 → hypermind-2.3.0}/docs/paper/hypermind-paper-draft.md

@@ -335,7 +335,7 @@ The design rationale for the conjunction threshold is sound: legitimate reputati
 
 ### 5.1 Overview
 
-The HyperMind reference implementation is a Python SDK, version 2.1.5, supporting Python 3.12–3.14. The SDK collects over 4,476 tests under `pytest` at the time of writing. An independent Rust second-language implementation (`rust/`) shipped in v2.1.0 and is covered by cross-language interop tests.
+The HyperMind reference implementation is a Python SDK, version 2.3.0, supporting Python 3.12–3.14. The SDK collects over 5,093 tests under `pytest` at the time of writing. An independent Rust second-language implementation (`rust/`) shipped in v2.1.0 and is covered by cross-language interop tests.
 
 The test tree is organised by subsystem, with dedicated suites for the wire format (A-ALG-BOUND, WF-DOMAIN), the dispute FSM (I-DISPUTE-PARTITION, TAL-ESCROW-EXCLUSIVE), capability tokens (A-TOMB-REVOKE-DISJOINT and the parent-hash chain invariants), the reputation kernel (Beta posterior, vouch cap, calibration gate), the transparency log (A-SEAL-ATOMIC, receipt verification), the crypto modules (Ed25519, FROST, HLC, domain separation), the mind layer (swarm deliberation, roles, diversity metrics), the transport layer (`_seen` cache bound, anti-entropy Bloom validation), the conformance CLI, the harness scenarios, and integration paths exercised by the numbered `examples/` tutorials. Five numbered self-healing regression suites (`test_self_heal_cycle1.py` through `test_self_heal_cycle5.py`, with two cycle-3 and cycle-4 companion files for the cascade-isolate and deliberator-broadcast fixes) pin specific vulnerability fixes; per-cycle contents are described in §6.1. The two-implementation interop suite (`tests/test_two_impl_interop.py`) covers SDK ↔ `reference_verifier` agreement on wire-format edge cases. Exact per-suite counts shift between releases as new regressions land; the authoritative current figures are emitted by `pytest --collect-only` on the tagged release commit.
 

{hypermind-2.2.1 → hypermind-2.3.0}/docs/project/roadmap.md

@@ -10,9 +10,9 @@ Each milestone is **Python-only** (per the language-scope policy below). Second-
 
 ---
 
-## Status snapshot — 2026-07-26
+## Status snapshot — 2026-07-31
 
-**Working version:** `2.1.5` (`pyproject.toml`, `src/hypermind/__init__.py`) —
+**Working version:** `2.3.0` (`pyproject.toml`, `src/hypermind/__init__.py`) —
 **engineering GA; SemVer now binding** (breaking changes only on major bumps, per
 [`CHANGELOG.md`](../../CHANGELOG.md)). Published on [PyPI](https://pypi.org/project/hypermind/).
 
@@ -295,6 +295,6 @@ trigger was evaluated and the project survived each gate:
 
 **Code lands fast; v1.0 GA is gated on the v1.0 §External & human deliverables list — none of which can be done by automation.**
 
-*Last updated: 2026-07-27. Working version: 2.1.5. Test suite: 4476 tests (`pytest --collect-only -q`).
-All engineering milestones shipped (v0.1 → v1.0 GA → v1.12 → v2.0 → v2.1.x).
+*Last updated: 2026-07-31. Working version: 2.3.0. Test suite: 5093 tests (`pytest --collect-only -q`).
+All engineering milestones shipped (v0.1 → v1.0 GA → v1.12 → v2.0 → v2.3.x).
 Remaining gates: §External & human deliverables only.*

{hypermind-2.2.1 → hypermind-2.3.0}/docs/project/v2.0-plan.md

@@ -8,7 +8,7 @@ This plan is preserved as a historical record of the v2.0 consolidation release.
 
 | v2.0 feature | Status | Where |
 |---|---|---|
-| `version = "2.0.0"` (now 2.1.5) | ✅ shipped | `pyproject.toml:7` |
+| `version = "2.0.0"` (now 2.3.0) | ✅ shipped | `pyproject.toml:7` |
 | `WIRE_VERSION_V3 = 3` added to `WIRE_VERSION_SUPPORTED` | ✅ shipped | `src/hypermind/wire.py:124-133, 188` |
 | `revocation_epoch` in CapToken canonical body (wire break — A-ALG-BOUND covers it) | ✅ shipped | `src/hypermind/capability.py:265-278, 505-523` |
 | `CapRevocationStore` auto-created on every `_NamespaceWorld` | ✅ shipped | `src/hypermind/agent.py:383-394, 419-426` |

hypermind-2.3.0/docs/release-notes/v2.2.0.md

@@ -0,0 +1,47 @@
+# HyperMind v2.2.0 — Release Notes
+
+**Released:** 2026-07-28 · **Minor — backward-compatible (API)** · no removals, no new required arguments, but several defaults now behave differently, deliberately.
+
+> **Upgrade note.** API-compatible, but the grounding gates that shipped as documented no-ops are now active, and a handful of knobs changed defaults. If you depend on previous behavior, the relevant knobs are listed under *Changed defaults* in the [CHANGELOG](../../CHANGELOG.md) §[2.2.0].
+
+## Highlights
+
+### Autonomy pass (#983)
+
+A code-grounded audit of 10 subsystems on the question "does this signal actually change a decision?", then closing every gap it found. Weighted autonomy score 2.9 → 4.6 / 5. See [`docs/audit/autonomy-audit.md`](../audit/autonomy-audit.md).
+
+### Security — SSRF hardening
+
+- **Redirects were never re-validated.** `_validate_url()` checked only the initial URL while `urlopen` followed redirects, so an innocent public URL could `302` to `169.254.169.254`. Every redirect hop is now validated; all HTTP paths route through the validating opener.
+- **Hostname checks were string-based.** Hostnames are now resolved and every resulting address checked via stdlib `ipaddress`, catching encodings a prefix check misses (`http://2130706433/` → `127.0.0.1`; `::ffff:127.0.0.1`).
+
+### New modules
+
+- `hypermind.profiles` — deployment profiles (`coding`, `research`) bundling persona roster, tool allowlist, model tiering, namespace.
+- `hypermind.orgs.coevolution` — dispute-driven prover/disprover fitness (Red Queen dynamics).
+- `hypermind.observability.silent_failures` + `hypermind.observability.health` — countable exception suppression; critical failures become research goals inside the loop.
+- `hypermind.knowledge.reputation_store` — local reputation persistence so trust compounds across runs.
+- `hypermind.responders.cache` (`ResponseCache`, `CachingResponder`) + TTL'd provider-result cache.
+- `DifficultyRouter` — routes on verifiability class (V/G → cheap model, C/S → strong) instead of agent identity.
+- `quality.set_entailment_scorer()` — optional NLI hook blended with the lexical grounding score.
+- `SqliteBackend.store_embedding()` + cosine ranking fused with BM25 via RRF (opt-in hybrid search).
+- `evolve_population()` generational archive with `stepping_stones()` / `archive_lineage()`.
+
+### Notable fixes
+
+A failed tool call was being counted as evidence (`_publish()`'s gate matched only JSON-shaped errors); enabling the evidence gate silently suppressed contradiction detection; reputation slashing ignored `weight`; expert relevance ranked on unbounded raw token overlap; `RelataBackend.hybrid_search()` discarded the server ranking; grounding couldn't distinguish a claim from its own negation.
+
+### Changed defaults
+
+`on_not_ready` `"raise"` → `"learn"`; `MIN_EVIDENCE_FOR_PUBLISH` `0` → `1`, `QUALITY_GATE_ENABLED` `0` → `1`; grounding/hallucination thresholds `0.05` → `0.20`; `MOA_MAX_REFINEMENT_ROUNDS` `1` → `2`; `OrgSim(persist_identities=…)` `False` → `True`; MoA panel size + refinement rounds now class-conditional; population fitness derived from oracle-confirmed `strategy_perf`; hypothesis promotion to `supported` requires a declared falsifier to have been met.
+
+## Breaking changes
+
+None (API). Default-behavior changes are deliberate and listed above.
+
+## Cross-references
+
+- [CHANGELOG.md](../../CHANGELOG.md) §[2.2.0]
+- [Feature matrix](../feature-matrix.md)
+- [Autonomy audit](../audit/autonomy-audit.md)
+- [README](../../README.md) §Quickstart

hypermind-2.3.0/docs/release-notes/v2.2.1.md

@@ -0,0 +1,39 @@
+# HyperMind v2.2.1 — Release Notes
+
+**Released:** 2026-07-29 · **Patch — backward-compatible** · no breaking changes.
+
+> Carries the CAP_REVOKE / TAL v0.2 transport work, the bi-temporal recall layer, the streamable-HTTP MCP gateway, and the #1011 assessment follow-up batch that accumulated between the v2.2.0 autonomy cut and this tag. Cut point is the test-suite stabilization commit.
+
+## Highlights
+
+### Single-container streamable-HTTP MCP gateway (#1013)
+
+`deploy/Dockerfile.mcp-http` + `deploy/docker-compose.mcp-http.yml` front `hmctl mcp serve` with [`supergateway`](https://github.com/supercorp-ai/supergateway), exposing a plain streamable-HTTP `/mcp` endpoint for platform integrations that don't speak raw TCP/SSE. Safe-by-default posture matches `deploy/Dockerfile.mcp`: write mode on by default, `MCP_WRITE=0` for read-only.
+
+### Bi-temporal recall + cross-namespace citation resolution (#942, L3 follow-up)
+
+`Hypothesis` gained `created_at_ms`/`resolved_at_ms`; `KernelStore.recall_as_of(persona, as_of_ms)` filters a persona's `learned_facts`/`confirmed_hypotheses` to what existed by a given cutoff (existence-based filtering). `prove_hypothesis_inclusion()`/`verify_foreign_hypothesis_citation()` add real RFC 6962 Merkle inclusion proofs against the persona's sealed `lattice_root`, so a citation across namespaces is independently verifiable, not just trusted on the citing agent's word.
+
+### Security
+
+- **CAP_REVOKE now gossips (#943 Phase 1).** A signed `CapRevocationRecord` previously had no transport — only the node that directly called `store.record()` enforced it. `revoke_capability_scope()` now gossips the record on the namespace bus and re-gossips it for 30 minutes per the spec's re-gossip obligation; every agent auto-subscribes on `__aenter__()`.
+- **`CapRevocationRecord.from_cbor()` gained the canonical-CBOR re-encoding gate** that `RevocationList.from_cbor()` already had — closes a malleability surface now that untrusted peers can send arbitrary bytes via gossip.
+- Removed a dead, structurally-incompatible duplicate `CapRevocationStore` from `hypermind.capability`; `reference_verifier` gained `parse_cap_revocation_record()`.
+- **TAL v0.2 signed wire messages (#943 Phase 2).** `create_room()`/`join_room()`/`vouch()` optionally emit a real `SignedStatement` (`ROOM_DEFINE`/`ROOM_JOIN`/`VOUCH`, new A-ALG-BOUND record types) and gossip it. New `TAL.revoke_vouch()` backs the new `VOUCH_REVOKE` message. `TAL.apply_signed()` independently re-verifies every inbound message's signature *and* that the envelope's bound issuer matches the payload's claimed identity. Existing callers with no `signing_keypair=` see zero behavior change.
+
+### Also in this cut
+
+- Batch of 15 assessment follow-up fixes (#1011) across intelligence, wiring, storage, and retrieval.
+- Web-capture pipeline fetches through the SSRF-validated opener (#973); checkpoint dir ignores `HYPERMIND_HOME`, off-by-one round-resume fix (#975); `select_experts()` recency + active-dispute suppression (#998); `CachingResponder` wired into the production responder path (#997); `config.py` duplicate constant + dead SWARM-IQ block (#996); retrieval-overlap clustering + dynamic seat allocation + cluster-level Sybil detection (#940).
+- Test-suite stabilization: 4 failures from `load_persistent_env()` env contamination resolved.
+
+## Breaking changes
+
+None.
+
+## Cross-references
+
+- [CHANGELOG.md](../../CHANGELOG.md) §[2.2.1]
+- [Feature matrix](../feature-matrix.md)
+- [Trust & Access Layer spec](../spec/07-trust-access-layer.md)
+- [README](../../README.md) §Quickstart

hypermind-2.3.0/docs/release-notes/v2.3.0.md

@@ -0,0 +1,34 @@
+# HyperMind v2.3.0 — Release Notes
+
+**Released:** 2026-07-31 · **Minor — backward-compatible** (additive on top of v2.2.1; all new parameters default to prior behavior).
+
+## Highlights
+
+Bundles the four epics from the "swarm can research a topic end-to-end" effort: web page/PDF discovery gaps, storage alignment with RelataDB, and making PI-decomposed swarms genuinely goal/objective-driven.
+
+### PDF fetch + site crawl (#1018)
+
+`fetch_page`'s `Content-Type: application/pdf` is no longer silently skipped/garbled — it routes through `pypdf` extraction, shared with the local-file ingest path. New `crawl_site` tool (`orgs/web_crawl.py`): SSRF-validated BFS crawl of same-domain links, hard-capped at 15 pages / depth 4 (default 5/2), covers HTML and PDFs.
+
+### RelataDB `document_ingest` alignment (#1019)
+
+`hypermind.ingest` gained `ExtractedChunk`/`DocumentManifest`/`chunk_to_manifest()`/`to_jsonl()` — a shared chunk/manifest shape matching RelataDB's datagrep extractor protocol. `StorageBackend` gained `ingest_document()`/`search_chunks()`: `RelataBackend` wraps the `relata` SDK's native `RelataClient.ingest_document()`; `SqliteBackend` mirrors it locally (`intel_reports`/`intel_chunks` tables + FTS5) so ingested documents are searchable without a Relata server. `OrgSim(ingest_via_manifest=True)` (opt-in) lands `ingest_dir=` files this way alongside the existing claim-publish path.
+
+### Goal-driven charter-decomposed swarms (#1020)
+
+`charter.plan_target_confidences()` + `OrgSim(unit_goal_confidence=...)` + `DynamicsConfig(mode="deliberator")` let a PI-decomposed swarm stop via real goal satisfaction (per-sub-area confidence + synthesis published) instead of only a fixed round count — bridging `charter.py`'s decomposition to the existing `mind.goals`/`Deliberator` machinery. Fixes a latent bug: the coordination unit's own goal is now excluded from the `goal_satisfied` AND-gate; synthesis is checked via a published claim instead.
+
+### Intra-expert research probe gate (#1021)
+
+`_think()` nudges (never filters) the model away from re-researching a sub-area already backed by a healthy density of confirmed facts, via a new `research_hint` on `default_context_builder`. Tunable via `HYPERMIND_KNOWLEDGE_PROBE_THRESHOLD` / `HYPERMIND_KNOWLEDGE_PROBE_MIN_FACTS`.
+
+## Breaking changes
+
+None (additive).
+
+## Cross-references
+
+- [CHANGELOG.md](../../CHANGELOG.md) §[2.3.0]
+- [Feature matrix](../feature-matrix.md)
+- [Deep-research engine spec](../spec/31-deep-research-engine.md)
+- [README](../../README.md) §Quickstart

{hypermind-2.2.1 → hypermind-2.3.0}/docs/spec/13-observability.md

@@ -267,6 +267,32 @@ hypermind_exploration_fraction_realised{namespace}
 
 Structured-log JSON schema with required fields: `trace_id`, `kid`, `agent_id`, `topic`, `hlc_l`, `hlc_c`. Reference Grafana dashboard JSON ships in `hypermind-conformance/runbooks/`. `hypermind_correctness_rate` is **aggregated by default**; per-agent series exposed only with operator-set auth token (privacy hardening).
 
+#### §13.6.1 Silent-failure observability and health **(v2.2.0, C4-#983)**
+
+A system built to degrade rather than abort suppresses many exceptions by
+design — but a swarm whose Relata writes have been silently failing for an
+hour looks, from the inside, exactly like one that is working. Self-repair is
+impossible without self-observation.
+
+  * `suppressed(*exceptions, op=...)` (`observability/silent_failures.py`) is a
+    drop-in for `contextlib.suppress` that keeps the same never-propagate
+    contract while incrementing a per-`op` counter and logging at DEBUG.
+    `failure_counts()` returns `{op: {count, last}}` for every op that has
+    swallowed something — a question that previously had no answer at any log
+    level. The op key-space is bounded (`_MAX_OPS = 512`, overflow buckets to
+    `_overflow`) so a label built from unbounded input cannot turn it into a
+    leak in a long `run_forever()`. It is deliberately dependency-free and
+    process-local: it must be safe to call from teardown and from inside
+    exception handlers, so it cannot itself raise or perform I/O.
+  * `health_report()` (`observability/health.py`) classifies each op into
+    `healthy` / `degraded` / `critical` by suppressed-failure count
+    (`DEGRADED_THRESHOLD = 10`, `CRITICAL_THRESHOLD = 50`). A `HealthReport`
+    is `healthy` only when nothing is `degraded` or `critical` — isolated
+    failures are normal and expected. `health_research_goals()` turns
+    **critical** subsystems into research goals inside the loop, so a silently
+    broken dependency becomes something the swarm tries to fix rather than
+    quietly works around.
+
 ### §13.7 Operational primitives
 
 ```

{hypermind-2.2.1 → hypermind-2.3.0}/docs/spec/20-swarm-intelligence.md

@@ -340,6 +340,16 @@ The skill evolution loop is the mechanism by which an agent's repeated deliberat
 are automatically promoted into reusable, executable skills — closing the learn→act→improve cycle
 without human intervention.
 
+> **Scope note.** This subsection covers *skill* evolution (work_history →
+> executable `SkillEntry`). The distinct **population / strategy evolution**
+> (retire/clone/mutate ticks, generational archive, stepping stones),
+> **dispute-driven coevolution** (prover/disprover Red Queen dynamics), and
+> **L4 calibration feedback** (`strategy_perf`/`tool_outcomes` → policy
+> mutation) are normatively specified in
+> [§31.11 Evolution and coevolution](31-deep-research-engine.md#3111-evolution-and-coevolution-the-evolve-step-v210--v220).
+> The `KernelStore.strategy_perf` field referenced at §20.2/§20.6 is the
+> topic-blind aggregate that feeds that strategy-gene mutation.
+
 ### §20.7.1 Work History
 
 Every `Deliberator` maintains a rolling `_work_history` deque (max 500 entries). An entry