simjecture 0.3.2__tar.gz → 0.4.0__tar.gz

{simjecture-0.3.2 → simjecture-0.4.0}/.github/workflows/ci.yml

@@ -50,7 +50,16 @@ jobs:
       - name: Install uv
         run: python -m pip install uv
       - name: Install locked environment
-        run: uv sync --frozen --extra tui
+        run: uv sync --frozen --extra tui --extra dsh
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "22.23.1"
+          cache: npm
+          cache-dependency-path: integrations/dsh/package-lock.json
+      - name: Install pinned DSH test runtime
+        run: npm ci --prefix integrations/dsh
+      - name: Test DSH adapter and real MCP profile
+        run: SIMJECTURE_MCP_EXECUTABLE="$PWD/.venv/bin/simjecture-mcp" npm test --prefix integrations/dsh
       - name: Lint
         run: uv run ruff check .
       - name: Test

{simjecture-0.3.2 → simjecture-0.4.0}/.github/workflows/release.yml

@@ -40,7 +40,16 @@ jobs:
             --ro-bind /lib64 /lib64 \
             /usr/bin/true
       - name: Install locked test environment
-        run: uv sync --frozen --extra tui
+        run: uv sync --frozen --extra tui --extra dsh
+      - uses: actions/setup-node@v4
+        with:
+          node-version: "22.23.1"
+          cache: npm
+          cache-dependency-path: integrations/dsh/package-lock.json
+      - name: Install pinned DSH test runtime
+        run: npm ci --prefix integrations/dsh
+      - name: Test DSH adapter and real MCP profile
+        run: SIMJECTURE_MCP_EXECUTABLE="$PWD/.venv/bin/simjecture-mcp" npm test --prefix integrations/dsh
       - name: Lint release source
         run: uv run ruff check .
       - name: Test release source

{simjecture-0.3.2 → simjecture-0.4.0}/.gitignore

@@ -21,3 +21,5 @@ artifacts/
 !demos/agent_roles_validation/record/workspace/artifacts/*.json
 data/
 .private/
+
+integrations/dsh/node_modules/

{simjecture-0.3.2 → simjecture-0.4.0}/CHANGELOG.md

@@ -4,6 +4,30 @@ This project follows semantic versioning. Dates use ISO 8601.
 
 ## Unreleased
 
+## 0.4.0 — 2026-09-20
+
+- Updated the bundled DSH integration to the tested `0.1.5-rc.2` runtime,
+  including current session, tool-call, prompt and structured-output APIs.
+  The Python release and DSH bundle now share version `0.4.0`.
+- Enabled native web search, page retrieval, shell, filesystem, skills,
+  workflows and configured plugins for research agents. Scientific calls retain
+  role restrictions; independent judges retain their frozen-case isolation.
+- Separated the native research workspace from authoritative campaign files.
+  Upgrades start a fresh `.research-v1` conversation, reconcile existing kernel
+  state, and preserve older conversations. Built-in DSH file/shell policies
+  protect records; arbitrary installed plugins remain trusted process code.
+- Added `read_workspace_image` for bounded PNG, JPEG and WebP image inspection
+  through MCP, with dimensions and content hashes. Inspection is not evidence
+  acceptance and requires a vision-capable model.
+- Extended the Simote scientific tool catalog for image reads. Matching Simote
+  changes add AGY workers and judges; Simote remains a separate optional app.
+- Added real-runtime DSH tests for role isolation, context pruning, V2/V3 log
+  migration, durable-job recovery and complete CLI/MCP pause/resume behavior.
+- Verified native search, fetch, file writing and image inspection with real
+  DeepSeek API calls on non-simulation research tasks.
+- Added a Linux launch bundle containing the wheel, source distribution,
+  matching DSH profile, setup/launch scripts and SHA-256 checksums.
+
 ## 0.3.2 — 2026-09-09
 
 - Added `simjecture-call`, a one-shot JSON transport for trusted supervisors,

{simjecture-0.3.2 → simjecture-0.4.0}/CITATION.cff

@@ -2,8 +2,8 @@ cff-version: 1.2.0
 message: "If you use this software, please cite the exact software version and Git commit."
 title: "Simjecture"
 type: software
-version: 0.3.2
-date-released: 2026-09-09
+version: 0.4.0
+date-released: 2026-09-20
 abstract: >-
   An evidence-governed runtime for autonomous computational experimentation and
   falsification over hypothesis trees.

{simjecture-0.3.2 → simjecture-0.4.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.5
 Name: simjecture
-Version: 0.3.2
+Version: 0.4.0
 Summary: Evidence-governed autonomous experimentation and falsification for computational science
 Project-URL: Homepage, https://github.com/tomzhu0225/simjecture
 Project-URL: Documentation, https://github.com/tomzhu0225/simjecture/tree/main/docs
@@ -24,6 +24,7 @@ Requires-Dist: httpx<1,>=0.27
 Requires-Dist: matplotlib<4,>=3.8
 Requires-Dist: numpy<3,>=1.26
 Requires-Dist: pandas<4,>=2.2
+Requires-Dist: pillow<13,>=10
 Requires-Dist: pydantic<3,>=2.9
 Requires-Dist: scipy<2,>=1.12
 Provides-Extra: dsh
@@ -49,7 +50,7 @@ writes its own experiments and diagnostics, and searches deliberately for the
 simplest result that survives independent verification. The harness controls
 what may count as evidence.
 
-> **Research preview:** version 0.3 extends the infrastructure with a
+> **Research preview:** version 0.4 extends the infrastructure with a
 > license-safe FLASH resistive-MHD capability and an explicitly bounded
 > island-coalescence audit. It does not claim to solve arbitrary
 > scientific prose, establish truth about nature from one simulator, or replace
@@ -289,11 +290,11 @@ the human-first projection.
 
 ## DeepSeek Harness integration
 
-The v0.2.2 integration moves the model-facing research loop to DeepSeek Harness
+The v0.4.0 integration moves the model-facing research loop to DeepSeek Harness
 without moving scientific authority out of Simjecture. DSH owns the provider,
 conversation, retry policy, compaction, and resumable session. A native MCP
-boundary implements 22 typed endpoints. A compact persistent Lead Scientist
-sees seven coordination tools and delegates claim work to fresh, scoped
+boundary implements 23 typed endpoints. A compact persistent Lead Scientist
+sees eight scientific coordination/inspection tools plus native research tools and delegates claim work to fresh, scoped
 Falsifier/Experimenter, Blocker Resolver, and Repair Scientist sessions. A separate fresh,
 tool-free Judge reviews surviving claims. The Python campaign kernel continues to own hypotheses,
 evidence contracts, commissioning, skills, simulation capabilities, sandboxing,

{simjecture-0.3.2 → simjecture-0.4.0}/README.md

@@ -13,7 +13,7 @@ writes its own experiments and diagnostics, and searches deliberately for the
 simplest result that survives independent verification. The harness controls
 what may count as evidence.
 
-> **Research preview:** version 0.3 extends the infrastructure with a
+> **Research preview:** version 0.4 extends the infrastructure with a
 > license-safe FLASH resistive-MHD capability and an explicitly bounded
 > island-coalescence audit. It does not claim to solve arbitrary
 > scientific prose, establish truth about nature from one simulator, or replace
@@ -253,11 +253,11 @@ the human-first projection.
 
 ## DeepSeek Harness integration
 
-The v0.2.2 integration moves the model-facing research loop to DeepSeek Harness
+The v0.4.0 integration moves the model-facing research loop to DeepSeek Harness
 without moving scientific authority out of Simjecture. DSH owns the provider,
 conversation, retry policy, compaction, and resumable session. A native MCP
-boundary implements 22 typed endpoints. A compact persistent Lead Scientist
-sees seven coordination tools and delegates claim work to fresh, scoped
+boundary implements 23 typed endpoints. A compact persistent Lead Scientist
+sees eight scientific coordination/inspection tools plus native research tools and delegates claim work to fresh, scoped
 Falsifier/Experimenter, Blocker Resolver, and Repair Scientist sessions. A separate fresh,
 tool-free Judge reviews surviving claims. The Python campaign kernel continues to own hypotheses,
 evidence contracts, commissioning, skills, simulation capabilities, sandboxing,

{simjecture-0.3.2 → simjecture-0.4.0}/docs/conf.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 project = "Simjecture"
 author = "Bowen Zhu"
 copyright = "2026, Bowen Zhu and contributors"
-release = "0.3.0"
+release = "0.4.0"
 
 extensions = [
     "myst_parser",

{simjecture-0.3.2 → simjecture-0.4.0}/docs/how-to/deepseek-harness.md

@@ -1,16 +1,21 @@
 # Run a Simjecture campaign under DSH
 
 The DSH integration is a native-tool profile plus a small resumable driver for
-a durable Simjecture campaign. Version 0.2.2 uses a persistent Lead Scientist
+a durable Simjecture campaign. Version 0.4.0 uses a persistent Lead Scientist
 and fresh, claim-scoped Falsifier, Repair Scientist, and tool-free Judge
 sessions. Fresh workers see bounded kernel state rather than inherited chat.
-The Python MCP process is the only scientific tool surface: it exposes campaign
-state, claims, evidence contracts, skills, literature metadata, workspace files,
-and bounded jobs. Generic shell, filesystem bypass, and unguarded `finish` tools are
-disabled by the profile wherever the host DSH version supports disabling them.
-The profile also sets DSH's approval policy to `never`: all remaining scientific
-side effects already cross the typed CampaignKernel policy and Bubblewrap
-boundary, so no second routine permission loop is inserted around MCP calls.
+Native web search, page fetch, shell, filesystem, skills, workflows and configured
+plugins remain available to research agents. Only authoritative Simjecture tools
+are filtered by scientific role. The Python MCP process governs campaign state,
+contracts, evidence acceptance, and tracked experiments. Its image endpoint
+returns bounded PNG/JPEG/WebP pixels for vision-capable models.
+
+The profile uses DSH's `workspace-write` file policy around a separate research
+directory and approval `never` (no prompts; escalation requests are rejected).
+Experiment execution retains Simjecture's separate Bubblewrap sandbox.
+
+For the upstream version review and migration requirements, see the
+[DSH upgrade assessment](dsh-upgrade-assessment.md).
 
 ## Provision the two runtimes
 
@@ -59,8 +64,8 @@ resulting local bundle into the isolated harness profile:
 ```bash
 SIMJECTURE_DSH_PROFILE="$(simjecture dsh-profile)"
 npm pack "$SIMJECTURE_DSH_PROFILE" --pack-destination /tmp
-dsh plugin --profile simjecture add @deepseek-ai/dsh-headless@0.1.1-rc.2
-dsh plugin --profile simjecture add /tmp/simjecture-dsh-bundle-0.2.2.tgz
+dsh plugin --profile simjecture add @deepseek-ai/dsh-headless@0.1.5-rc.2
+dsh plugin --profile simjecture add /tmp/simjecture-dsh-bundle-0.4.0.tgz
 ```
 
 For a checkout-only development install, use the directory directly when the
@@ -71,11 +76,13 @@ dsh plugin --profile simjecture add "$PWD/integrations/dsh"
 ```
 
 The bundle pins the tested DSH prerelease and MCP-client prerelease in its
-`package.json`. Keep the lockfile produced by `npm pack`/the DSH plugin manager
-with the deployment artifact; do not silently widen either range.
+`package.json`. The source bundle includes a `package-lock.json` for `npm ci` and runtime tests.
+Keep the separate lockfile produced by the DSH plugin manager with the
+deployment artifact; `npm pack` does not create a lockfile. Do not widen the pins.
 
-DSH `0.1.1-rc.2` requires Node.js `^22.19.0` or `>=24.0.0`. Install the DSH CLI
-according to its upstream instructions and verify that `dsh` is on `PATH`
+DSH `0.1.5-rc.2` requires Node.js `^22.19.0` or `>=24.0.0`. Install the DSH CLI
+at that exact version (`npm install -g @deepseek-ai/dsh@0.1.5-rc.2`)
+and verify that `dsh` is on `PATH`
 before installing this isolated profile.
 
 ## Inspect the resolved configuration
@@ -91,8 +98,8 @@ Confirm that the native MCP client is named `simjecture`, starts
 `simjecture-mcp`, carries the `SIMJECTURE_*` environment, and has
 `failOnStartupError: true`. Also confirm `approval.policy: never`, the
 `simjecture` permission preset (`workspace-write` plus `never`), and that the
-generic tool rows are disabled. A startup or handshake failure should stop the
-scientific profile rather than silently fall back to a bypass tool.
+native research tools remain enabled. A startup or handshake failure must stop
+the profile before its first model request.
 
 Completed large execution and workspace-write exchanges stay fully visible for
 one model request. At a later step the profile may replace the balanced exchange
@@ -138,11 +145,12 @@ simjecture-mcp --workspace "$SIMJECTURE_WORKSPACE" \
 
 That command speaks MCP over stdio and expects an MCP client; it is not a
 human-facing REPL. Use DSH's initialize/list-tools/call-tool trace to verify the
-22 explicit MCP endpoints before submitting work. The persistent lead sees seven
-tools: two durable reads, fresh falsification, blocker resolution, and repair
-delegation, independent adjudication, and guarded finalization. Each worker receives a narrower tool
-surface plus a claim-scoped argument guard. Raw judge prepare/commit operations
-are private to the adjudication composite.
+23 explicit MCP endpoints before submitting work. The lead has eight scientific
+coordination/inspection tools alongside native research tools. Workers receive
+role-specific scientific subsets and claim guards, while retaining native tools.
+Native delegated children inherit their parent's scientific restrictions. The
+independent judge sees only its frozen case and DSH's structured-output handoff;
+raw judge prepare/commit operations remain private to the composite.
 
 `observation_sufficient=true` means that a linked artifact satisfies its
 selected prospective contract; it is not a researcher-issued support verdict.
@@ -191,3 +199,65 @@ a `repairs` child; an honestly complete blocker may close as
 the immutable hand-off artifact. If later review finds an error, append a
 corrective audit record rather than editing the original ledger, transcript, or
 report.
+
+## Upgrade an existing deployment
+
+Stop its campaign supervisor and back up the entire campaign directory, including
+`operator_input/dsh_sessions`, before upgrading the CLI and isolated profile.
+Install DSH `0.1.5-rc.2`, replace the headless bundle with the matching version,
+and install Simjecture's `0.4.0` profile using the commands above. Inspect
+`--dump-config` before resuming. Do not mix the `0.1.6` alpha packages into this
+profile. Provider/model selection stays in your DSH configuration.
+
+DSH publishes a new V3 session generation when opening an older supported log
+for writing; it retains the older generation. A rollback must use the complete
+pre-upgrade campaign backup and the matching older runtime/profile. Do not point
+an older DSH runtime at a campaign that has continued under the newer version.
+
+To reproduce adapter validation from this checkout:
+
+```bash
+uv sync --extra dsh
+npm ci --prefix integrations/dsh
+SIMJECTURE_MCP_EXECUTABLE="$PWD/.venv/bin/simjecture-mcp" npm test --prefix integrations/dsh
+uv run pytest -q tests/test_dsh_bundle.py tests/test_dsh_engine.py
+```
+
+The JavaScript tests use the pinned real DSH agent loop, tools, child runtime,
+and persistence, with a deterministic model adapter. The CLI test also starts
+the real Python MCP server and exercises snapshot, finalization refusal,
+pause/resume, and mandatory-startup failure. It skips when the Python executable
+is unavailable. These tests make no paid model requests and run no scientific
+simulation. The test profile uses filesystem polling to avoid host inotify
+limits; this is a test setting, not a change to deployed profiles.
+
+## Research workspace and record boundary
+
+The runner creates a native research directory outside `SIMJECTURE_WORKSPACE`.
+Its default is the sibling `.simjecture-research/<session-id>` directory;
+`SIMJECTURE_DSH_RESEARCH_ROOT` can select another directory. Overlapping research
+and campaign roots are rejected. Native DSH file and shell tools may write in
+this research directory; they cannot directly overwrite campaign files under
+the configured file policy. Network tools remain enabled. Copy experiment source
+into the scientific workspace using `write_workspace_file` before submitting a
+contracted run. Native notes and calculations do not automatically become evidence.
+
+Bundle 0.4.0 uses the deterministic `<launch-session-id>.research-v1` DSH identity.
+This deliberately avoids resuming pre-0.4.0 sessions whose immutable working
+directory was the campaign itself. On first upgrade, the researcher starts a
+fresh conversation and reconciles the existing kernel snapshot; previous logs
+are retained. Later pause/resume uses the new stable identity normally.
+
+This is a boundary for DSH's enforcing built-in tools, not an OS security boundary
+around arbitrary plugin code. Installed extensions are trusted code with the
+DSH process's authority; upstream 0.1.5 workflow workers likewise are not a
+security sandbox. Deploy the scientific service under a separate account or on
+another host if protection against hostile extensions is required. This profile
+does not install that account/service separation.
+
+`read_workspace_image(path)` accepts campaign-relative PNG, JPEG, and WebP files
+up to 4 MiB and 16 megapixels. MCP clients receive actual image content plus
+path, dimensions, byte count and SHA-256 metadata. A vision-capable model is
+required to interpret the pixels. Image reading is non-mutating and does not
+promote an artifact to scientific evidence. The independent judge still receives
+only its frozen adjudication packet, not unrestricted browsing or image tools.

simjecture-0.4.0/docs/how-to/dsh-upgrade-assessment.md

@@ -0,0 +1,73 @@
+# DSH upgrade assessment — 2026-09-20
+
+Migration implemented: Simjecture profile `0.2.3` pins DSH `0.1.5-rc.2`.
+The scientific CampaignKernel and its evidence contract remain unchanged.
+The adapter changes and validation described below replace the earlier
+recommendation to retain `0.1.1-rc.2` pending compatibility work.
+
+The official npm registry currently tags `0.1.5-rc.2` as `latest` and `next`,
+and `0.1.6-alpha.2` as `alpha`. These are all prereleases. The recommended first
+migration target is the default-channel `0.1.5-rc.2`, keeping the alpha work
+separate. Versions were checked with `npm view @deepseek-ai/dsh dist-tags --json`.
+
+## Compatibility findings addressed
+
+The audit inspected upstream tag `dsh-v0.1.5-rc.2`, commit
+`fb2c4b9e698e30edb738bca4cf0618587db7d203`.
+
+| Boundary | Previous adapter | Upstream target | Migration requirement |
+| --- | --- | --- | --- |
+| Session history | `runner.js` and `context-elider.js` read `session.events` | `Session` exposes `snapshotEvents()` and `eventAt()` instead | Migrate history access and exercise resume, summarization, and pruning against real DSH |
+| Surface replacement | `context-elider.js` writes `{ op: 'replace', start, end }` | `SurfaceOp` requires `startSeq` and `endSeq` | Update replacement events and test balanced tool exchanges and preserved original history |
+| Durable sessions | Existing campaign-local DSH logs | Target includes session-format V3 and V2-to-V3 migration | Test upgrade on copies of old logs, including restart and failed migration; retain originals for rollback |
+| Profile composition | Pinned headless profile, disabled generic tools, seven lead tools and scoped workers | Plugin composition and runtime have evolved | Inspect resolved configuration and actual tool catalogs; test mandatory MCP startup failure and every role restriction |
+
+Primary source files:
+
+- [Session history implementation](https://github.com/deepseek-ai/deepseek-harness/blob/dsh-v0.1.5-rc.2/packages/core/session/src/index.ts)
+- [Surface event types](https://github.com/deepseek-ai/deepseek-harness/blob/dsh-v0.1.5-rc.2/packages/core/session/src/types.ts)
+- [V2-to-V3 migration](https://github.com/deepseek-ai/deepseek-harness/blob/dsh-v0.1.5-rc.2/packages/session/session-format-v2-to-v3/src/migration.ts)
+- [Official releases](https://github.com/deepseek-ai/deepseek-harness/releases)
+
+The `0.1.6` release notes additionally describe headless stdin/session/JSON-event
+support, MCP SDK v2, asynchronous agent initialization, deprecation of synchronous
+history readers, and changes to plugin loading. Those features may eventually
+simplify our custom driver, but do not by themselves replace its campaign budget,
+control-file, activity projection, and reconciliation responsibilities. An alpha
+migration should use its asynchronous history API rather than adopting a newly
+deprecated synchronous interface.
+
+## Why the scientific kernel does not need to change
+
+DSH owns model requests, retries, conversation history, compaction, and child
+sessions. CampaignKernel owns contracts, claim disposition, evidence provenance,
+sandboxing, durable simulation jobs, and the deterministic finish gate. The
+confirmed upstream breaks are in the JavaScript adapter's DSH APIs, not the
+scientific MCP contract. No kernel schema migration is justified by these changes.
+
+The version-change acceptance criteria require a real-runtime adapter smoke test in addition
+to the existing bundle tests: fresh campaign, each scoped role, tool-free judge,
+context pruning, pause/resume on migrated logs, interrupted durable job recovery,
+and finalization only through the kernel gate. Existing source-string/fake tests
+alone cannot qualify a DSH version change. Install and test an isolated profile;
+do not migrate active campaign logs in place during evaluation.
+
+## Completed validation
+
+The adapter now uses `snapshotEvents()` / `eventAt()`, V3 surface replacement
+fields, the renamed `ToolCallId` export, and `personaPrefix`. Removed profile
+rows no longer produce missing-entry warnings. Claim guards attach at
+`agent/created`, before a child receives its first prompt. The Blocker Resolver
+allows DSH's schema-validated `structured_output` handoff. The Judge uses the
+supported schema subset and explicitly rejects inconsistent decision/disposition
+pairs before calling the kernel.
+
+The checked-in Node tests run real DSH services with deterministic model output.
+They cover original-history preservation, V2-to-V3 migration, fresh/resumed lead
+sessions, each worker's first-call mutation guard, independent judge tools and
+usage, and job reconciliation without model polling or resubmission. A complete
+CLI/profile/Python-MCP test verifies the exact lead tool catalog, immutable
+snapshot access, rejection of premature finalization, pause/resume, and no model
+request after mandatory MCP startup failure. No paid-provider or long-running
+scientific campaign was used for qualification. Production profiles and existing
+campaign logs were not modified during this work.

{simjecture-0.3.2 → simjecture-0.4.0}/docs/how-to/simote-agent-roles.md

@@ -1,6 +1,6 @@
 # Scientific roles with Simote agent sessions
 
-Simote can run Codex and Grok CLI agents as claim-scoped Simjecture workers.
+Simote can run Codex, Grok, and AGY (Antigravity) CLI agents as claim-scoped Simjecture workers.
 Simjecture remains the authority for evidence, claim disposition, jobs, budgets,
 and campaign finalization. The native API runner and DSH profile remain available.
 
@@ -18,7 +18,7 @@ and role launches; existing assigned work can finish and records are preserved.
 ## Run a shared campaign
 
 1. Configure a campaign-owner bot and one or more worker bots in the same Simote
-   team section, using the same named SSH compute machine. Select Codex or Grok
+   team section, using the same named SSH compute machine. Select Codex, Grok, or AGY
    agent engines for the workers, with their existing local logins.
 2. Ask the owner to use `simjecture_open_campaign`. The owner keeps the campaign
    under its workspace. Scientific worker tasks share that campaign; they do not
@@ -49,7 +49,7 @@ and role launches; existing assigned work can finish and records are preserved.
 5. For independent review, the owner or assigned lead calls
    `simjecture_adjudicate` with `campaign_id`, `judge_bot_id`, `operation_id`,
    `claim_id`, `contract_version`, and `case_for_sufficiency`. The selected
-   Codex/Grok instance receives a fresh session, empty working directory, no
+   Codex/Grok/AGY instance receives a fresh session, empty working directory, no
    conversation history or Simote integrations, and only the frozen case.
    Tool activity invalidates the review. Only a successful text-only response
    reaches Simjecture's existing verdict validation and scientific gates.
@@ -64,6 +64,13 @@ and the final report. You can create or attach a campaign, assign workers,
 request independent review, cancel jobs, and finalize through the same kernel
 gates. The existing Simjecture web dashboard is also available.
 
+AGY workers require the instance's full-auto setting, because its headless CLI
+cannot approve MCP calls interactively. Simote mounts named scientific MCP tools
+for each turn and removes them before a tool-free judge starts. AGY uses a global
+MCP configuration, so Simote serializes AGY child lifetimes to prevent one task
+from receiving another task's credentials. User-configured MCP entries remain
+preserved; any observed judge tool use rejects the review, as for other engines.
+
 ## Recovery and boundaries
 
 Campaign creation first probes Bubblewrap namespace support. A GPU container

{simjecture-0.3.2 → simjecture-0.4.0}/docs/index.md

@@ -6,7 +6,7 @@ It accepts a natural-language hypothesis, lets an agent commission and use
 computational instruments, and preserves an independently inspectable path from
 proposal to claim disposition.
 
-This documentation describes version 0.3.2, a research preview. The project
+This documentation describes version 0.4.0, a research preview. The project
 began in computational plasma physics. The same evidence harness is now ready
 to extend to other simulation-gated fields. The software has completed real
 autonomous CPU and CUDA campaigns, but it does not claim unrestricted
@@ -75,6 +75,7 @@ how-to/guided-commissioning
 how-to/add-a-capability
 how-to/deploy-runtimes
 how-to/deepseek-harness
+how-to/dsh-upgrade-assessment
 how-to/simote-agent-roles
 ```
 

{simjecture-0.3.2 → simjecture-0.4.0}/integrations/dsh/README.md

@@ -7,7 +7,7 @@ final response. Simjecture remains authoritative for the hypothesis DAG,
 evidence contracts, commissioning, provenance, sandbox, and durable simulation
 jobs.
 
-Version 0.2.2 separates scientific work into four scopes:
+Version 0.4.0 targets DSH 0.1.5-rc.2 and separates scientific work into four scopes:
 
 - a persistent, compact **Lead Scientist** reads durable state and delegates;
 - a fresh **Falsifier/Experimenter** commissions and tests exactly one open
@@ -52,26 +52,24 @@ the hypothesis file must exactly match the immutable root stored there.
 
 ## Pack and install an isolated profile
 
-The integration is pinned to DSH `0.1.1-rc.2`. `simjecture dsh-profile` resolves
+The integration is pinned to DSH `0.1.5-rc.2`. `simjecture dsh-profile` resolves
 the same bundle from either a source checkout or an installed wheel. Pack it
 outside the repository or remove the generated tarball after installation:
 
 ```bash
 SIMJECTURE_DSH_PROFILE="$(simjecture dsh-profile)"
 npm pack "$SIMJECTURE_DSH_PROFILE" --pack-destination /tmp
-dsh plugin --profile simjecture add @deepseek-ai/dsh-headless@0.1.1-rc.2
-dsh plugin --profile simjecture add /tmp/simjecture-dsh-bundle-0.2.2.tgz
+dsh plugin --profile simjecture add @deepseek-ai/dsh-headless@0.1.5-rc.2
+dsh plugin --profile simjecture add /tmp/simjecture-dsh-bundle-0.4.0.tgz
 ```
 
-The dedicated profile is important: its base/headless composition lets this
-bundle disable generic shell, filesystem, web, skill, workflow, and subagent
-tools. Installing only the host row into DSH's standard Web agent preset would
-also leave that preset's coding tools available.
-
-The profile sets DSH approval to `never`; this does not bypass Simjecture's
-policy. The only remaining scientific side effects are typed MCP actions checked
-by CampaignKernel and executed in its sandbox. Provider/model choice is still a
-DSH setting and is not hard-coded by this bundle.
+The dedicated profile composes DSH's native research tools with Simjecture's
+scientific tools. Web, shell, files, skills, workflows, and configured plugins
+remain available. Role restrictions apply to scientific calls, and ordinary
+native file/shell operations are confined to a separate research workspace.
+Approval `never` disables prompts and rejects privilege escalation. Experiment
+execution and evidence acceptance remain governed by CampaignKernel.
+Provider/model choice remains a DSH setting.
 
 Two context policies are included. Completed large tool exchanges remain
 verbatim for one model request and may then be replaced on the model-facing
@@ -90,7 +88,7 @@ dsh --profile simjecture --dump-config
 
 The dump must show `mode: native`, `approval.policy: never`, the matching
 `simjecture` permission preset, `mcp-simjecture`, the explicit `SIMJECTURE_*`
-environment, and the bypass rows as disabled. Missing workspace or hypothesis
+environment, and native research tools as enabled. Missing workspace or hypothesis
 variables fail profile activation.
 
 After installation, let Simjecture create the immutable launch contract and
@@ -107,10 +105,9 @@ log beside the campaign, and reopens the same session on **Resume**.
 
 The MCP child exposes no generic shell and no unguarded `finish` tool. Its
 `finalize_campaign` endpoint can write a report only after CampaignKernel's
-scientific finish gate passes. The lead sees only snapshot, claims, the three
-role composites, and finalization. Each child receives only the tools required
-for its assignment; the two raw judge prepare/commit endpoints remain private
-to the adjudication composite.
+scientific finish gate passes. Research roles retain native tools alongside
+role-scoped scientific operations, including image inspection. The two raw
+judge prepare/commit endpoints remain private to the adjudication composite.
 
 Capability jobs, including configured WarpX capabilities, still return a
 durable job ID immediately at the kernel boundary. The DSH waiter performs the
@@ -125,7 +122,7 @@ invalid receipt remains `outcome_unknown` and cannot become evidence.
 
 `simjecture-mcp` speaks MCP over standard input/output and is not a human REPL.
 Use an MCP client to perform initialize, list-tools, and call-tool. A healthy
-server advertises 22 explicit endpoints. The persistent lead sees seven tools;
+server advertises 23 explicit endpoints. The persistent lead sees eight scientific tools plus native research tools;
 the fresh role composites expose narrower task-specific subsets, while raw
 judge operations stay private. A direct MCP client must still reconcile a job
 with `job_status`; the DSH profile performs that wait automatically.
@@ -135,3 +132,22 @@ its selected contract; it does not declare scientific support. The Falsifier
 can make that bounded statement, but only the fresh Judge can authorize support.
 Adjudication is rejected before a Judge is started when no qualifying link could
 pass the kernel's deterministic closure gate.
+
+## Adapter validation
+
+From the checkout, run `npm ci --prefix integrations/dsh` and
+`npm test --prefix integrations/dsh`. Tests exercise the real pinned DSH runtime
+with a deterministic model adapter, including V2-to-V3 history migration,
+pruning, role isolation, and durable-job reconciliation. The full CLI/MCP test
+requires the repository's Python environment with the `dsh` extra; set
+`SIMJECTURE_MCP_EXECUTABLE` to its `simjecture-mcp` executable if needed.
+Back up campaigns before upgrading an existing deployment and retain matching
+CLI/headless/profile versions for rollback.
+
+Native research files live outside campaign records (`SIMJECTURE_DSH_RESEARCH_ROOT`
+can override the default sibling directory). Version 0.4.0 starts a new
+`.research-v1` conversation when upgrading from older campaign-root sessions;
+existing scientific state is reconciled and old logs remain available. DSH's
+built-in file/shell policies protect records from ordinary native writes.
+Installed plugins remain trusted process code; separate-account isolation against
+hostile extensions is not provided by this profile.

{simjecture-0.3.2 → simjecture-0.4.0}/integrations/dsh/adjudicator.js

@@ -8,7 +8,7 @@
  */
 
 import { appendFileSync } from 'node:fs'
-import { CallId } from '@deepseek-ai/dsh-llm'
+import { ToolCallId } from '@deepseek-ai/dsh-llm'
 
 export const name = 'simjecture-adjudicator'
 export const inject = ['agentDefaultModel', 'subagents', 'tools']
@@ -40,25 +40,6 @@ const VERDICT_SCHEMA = {
     evidence_gaps: { type: 'array', items: { type: 'string' } },
     next_test: { oneOf: [{ type: 'string' }, { type: 'null' }] },
   },
-  oneOf: [
-    {
-      properties: {
-        decision: { type: 'string', enum: ['sufficient'] },
-        scientific_disposition: {
-          type: 'string',
-          enum: ['supported', 'falsified', 'instrument_limited', 'unresolved'],
-        },
-      },
-      required: ['decision', 'scientific_disposition'],
-    },
-    {
-      properties: {
-        decision: { type: 'string', enum: ['insufficient'] },
-        scientific_disposition: { type: 'null' },
-      },
-      required: ['decision', 'scientific_disposition'],
-    },
-  ],
   required: [
     'claim_id',
     'contract_version',
@@ -135,7 +116,7 @@ function errorText(result) {
 
 async function internalMcpCall(ctx, exec, suffix, name, args) {
   const result = await ctx.tools.execute({
-    callId: CallId(`${String(exec.callId)}:simjecture:${suffix}`),
+    callId: ToolCallId(`${String(exec.callId)}:simjecture:${suffix}`),
     rootCallId: exec.rootCallId,
     name,
     arguments: args,
@@ -159,7 +140,7 @@ async function internalMcpCall(ctx, exec, suffix, name, args) {
 
 function usageFrom(run) {
   const totals = {}
-  for (const event of run.localAgent?.session?.events ?? []) {
+  for (const event of run.localAgent?.session.snapshotEvents() ?? []) {
     if (event.type !== 'assistant/message' || event.data.usage === undefined) continue
     for (const [key, value] of Object.entries(event.data.usage)) {
       if (typeof value === 'number' && Number.isFinite(value)) {
@@ -289,6 +270,12 @@ export function apply(ctx) {
         )
       }
 
+      // DSH's structured-output subset disallows type + oneOf at the same
+      // schema node. Preserve the cross-field invariant before kernel commit.
+      const verdict = result.structured
+      if ((verdict.decision === 'insufficient') !== (verdict.scientific_disposition === null)) {
+        throw new Error('independent judge returned an inconsistent decision and scientific disposition')
+      }
       const selection = ctx.agentDefaultModel.currentSelection()
       const recorded = await internalMcpCall(
         ctx,