Endofthestars · Endofthestars · Jul 31, 2026 · Jul 31, 2026
@@ -5,6 +5,11 @@ versioned schemas plus explicit migration notes for compatibility.
 
 ## 0.3.0rc2 — The Invitation
 
+- Fixed `benchwork_open_task` minting every Task Capsule for the `codex` Host.
+  It now accepts the optional `host_session` argument the completion tools
+  already took, so a non-Codex Host can record truthful provenance. The default
+  remains `codex` and an unregistered Host is refused.
+- Recorded the Claude Code CLI Tier 1 Host PASS from its 2026-07-31 trial.
 - Added the auditable POSIX installer and exact release/channel manifests.
 - Added uv-first isolated installation, pipx fallback, plugin staging,
   installer state, Host diagnostics, repair, rollback, and uninstall.

@@ -20,7 +20,8 @@ tier.
 
 Tier 1 is evaluated per declared release Host. Codex CLI is the primary Phase
 2 Host. Host-neutral compatibility with another CLI is not the same as an
-interactive PASS for that Host.
+interactive PASS for that Host; Claude Code CLI earned its own PASS from its
+own trial, not from the shared contract.
 
 ## Current acceptance
 
@@ -29,7 +30,7 @@ interactive PASS for that Host.
 | Core Runtime | 0 | PASS | Required gate satisfied |
 | MCP STDIO | MCP | PASS | Required gate satisfied |
 | Codex CLI | 1 | PASS | Primary CLI gate satisfied |
-| Claude Code CLI | 1 | PENDING_HOST_VALIDATION | Host-neutral contract only; not a claimed interactive PASS |
+| Claude Code CLI | 1 | PASS | Second CLI gate satisfied by the 2026-07-31 trial |
 | Codex IDE extension | 2 | BLOCKED_BY_ENVIRONMENT | Accepted exception `HOST-IDE-001` |
 | Claude Desktop | 2 | OPTIONAL_NOT_RUN | No Phase 2 release impact |
 | External Review | disclosure | WAITING_FOR_DISCLOSURE_AUTHORIZATION | Correctly denied; not a failure |
@@ -57,12 +58,37 @@ Reproduction and closure:
 6. Record Host version, plugin version, tool count, Receipt identifier, and
    any exception without copying secrets or conversation reasoning.
 
+## 2026-07-31 Claude Code CLI trial
+
+Fresh ephemeral Claude Code CLI `2.1.220` conversations loaded the Benchwork
+MCP server with `--strict-mcp-config` against a temporary project outside the
+source repository. The server reported `connected` and the Host discovered all
+38 registry tools. `benchwork_status` and `benchwork_next_actions` each
+returned `ok: true` with an `mcp-tool-result/1.0` envelope, both from an
+explicit request and from the broad request "What should we do next for this
+study?" with 26 non-MCP tools still available.
+
+The bounded Task loop initially failed. `benchwork_open_task` minted every Task
+Capsule with a hardcoded `codex` Host, so a truthful `claude-code` provenance
+was refused with `VALIDATION_REJECTED`. Athanor was correct; the MCP boundary
+gave a non-Codex Host no way to describe itself. After `benchwork_open_task`
+was given the same optional `host_session` argument the completion tools
+already accepted, the loop completed: Capsule `TK-B280B5160682` recorded host
+`claude-code`, Ward returned `PASS`, and Athanor issued Receipt
+`RC-47D74468807A` for the accepted `code-inspection-result/1.0` output. The
+source repository never acquired a `.benchwork/` directory.
+
+Evidence is published under
+[`examples/phase2-final/host-claude-code`](../../examples/phase2-final/host-claude-code/README.md).
+
 ## Evidence policy
 
 PASS requires a dated, reproducible Host trial. Contract inspection or shared
-configuration alone is insufficient. `BLOCKED_BY_ENVIRONMENT` requires an
-identifier, owner, reason, impact, reproduction steps, and the boundary before
-which it must be closed.
+configuration alone is insufficient. The Claude Code CLI trial demonstrates
+why: the adapter, the Capsule schema, and Ward all already admitted
+`claude-code`, and only a live Host reached the one line that did not.
+`BLOCKED_BY_ENVIRONMENT` requires an identifier, owner, reason, impact,
+reproduction steps, and the boundary before which it must be closed.
 
 See the [Phase 2 acceptance matrix](PHASE2_ACCEPTANCE.md) and
 [Acceptance Exception Policy](plugins/acceptance-exception-policy.md).
@@ -27,7 +27,7 @@ states.
 | Tier 0 — Kernel | PASS | Required | Chronicle, Schema, Ward, Capability, replay, and deterministic suites |
 | MCP | PASS | Required | in-memory and spawned-STDIO discovery, envelopes, pagination, Task and canonical tool tests |
 | Tier 1 — Codex CLI Host | PASS | Required before Phase release | fresh Codex CLI explicit/implicit Skill trials and MCP calls |
-| Tier 1 — Claude Code CLI | PENDING_HOST_VALIDATION | Not a claimed Phase 2 Host PASS | host-neutral contract is implemented; independent interactive evidence is pending |
+| Tier 1 — Claude Code CLI | PASS | Second CLI Host gate satisfied | fresh Claude Code CLI MCP discovery, read-only calls, and a bounded Task loop with an Athanor Receipt |
 | Tier 2 — IDE Host | BLOCKED_BY_ENVIRONMENT (accepted) | Does not block Core/MCP/CLI | `HOST-IDE-001`; extension installed, no graphical extension Host available |
 | External Review | WAITING_FOR_DISCLOSURE_AUTHORIZATION | Does not block local review or Core/MCP/CLI | Disclosure Gate denied an unapproved diff review |
 
@@ -79,6 +79,33 @@ an accepted `bench.study.audit` proposal, replacement Protocol preview/commit,
 registered implementation and Runs, Alembic analysis, Assessment, and a
 human-sealed `REPAIR` Decision through Chronicle replay.
 
+## 2026-07-31 Claude Code CLI Tier 1 trial
+
+- Fresh ephemeral Claude Code CLI `2.1.220` conversations loaded the Benchwork
+  MCP server with `--strict-mcp-config` against a temporary project outside the
+  source repository. The server reported `connected` and all 38 registry tools
+  were discovered.
+- `benchwork_status` and `benchwork_next_actions` both returned `ok=true` with
+  `mcp-tool-result/1.0`, from an explicit request and from the broad request
+  "What should we do next for this study?" while 26 non-MCP tools remained
+  available.
+- The first bounded Task loop was blocked. `benchwork_open_task` minted the
+  Capsule with a hardcoded `codex` Host, so truthful `claude-code` provenance
+  was refused with `VALIDATION_REJECTED`. The Host reported the rejection
+  rather than resubmitting a false provenance.
+- `benchwork_open_task` now accepts the optional `host_session` argument that
+  the completion tools already took, defaulting to `codex` and refusing an
+  unregistered Host. The repaired loop completed `bench.code.inspect` as
+  `TK-B280B5160682` with Capsule host `claude-code`, Ward `PASS`, and Receipt
+  `RC-47D74468807A`.
+- The Host used only `Read` alongside the Benchwork tools, matching the Capsule
+  Circle. Deep Doctor reported 2 verified Chronicle events with an intact
+  Receipt chain, and the source repository had no `.benchwork/` directory.
+- Evidence is published under
+  [`examples/phase2-final/host-claude-code`](../../examples/phase2-final/host-claude-code/README.md).
+- IDE execution and External Review were not exercised; `HOST-IDE-001` and
+  `WAITING_FOR_DISCLOSURE_AUTHORIZATION` are unchanged.
+
 ## 2026-07-30 M17 rc1 trial
 
 - The Python wheel and source distribution built as `0.3.0rc1`; Twine accepted

@@ -127,10 +127,14 @@ audit-to-`REPAIR` acceptance; migration and release documentation; and
 release-gated CI coverage.
 
 Phase 2 is sealed with Kernel, MCP, Plugin, and Codex CLI gates passing.
-`HOST-IDE-001` remains an accepted `BLOCKED_BY_ENVIRONMENT` Tier 2 exception,
-External Review remains `WAITING_FOR_DISCLOSURE_AUTHORIZATION`, and Claude Code
-CLI remains `PENDING_HOST_VALIDATION`. These states are not reclassified as
-failures or fabricated PASS results.
+`HOST-IDE-001` remains an accepted `BLOCKED_BY_ENVIRONMENT` Tier 2 exception
+and External Review remains `WAITING_FOR_DISCLOSURE_AUTHORIZATION`. These
+states are not reclassified as failures or fabricated PASS results.
+
+Claude Code CLI was `PENDING_HOST_VALIDATION` at the rc1 seal and recorded its
+own Tier 1 PASS on 2026-07-31, after its trial found and repaired a Task
+Capsule Host defect. See the
+[Host Support Matrix](HOST_SUPPORT_MATRIX.md).
 
 ## M18: The Invitation
 

@@ -22,8 +22,10 @@ An unavailable Tier 2 environment is recorded as `BLOCKED_BY_ENVIRONMENT`, not
 the release boundary before which validation becomes required.
 
 Tier 1 is accepted per Host. Codex CLI is the Phase 2 primary Host and has a
-recorded PASS. Claude Code CLI remains `PENDING_HOST_VALIDATION`; the
-host-neutral tool contract does not manufacture an interactive Host result.
+recorded PASS. Claude Code CLI recorded its own PASS on 2026-07-31 from its own
+interactive trial; the host-neutral tool contract does not manufacture an
+interactive Host result, and that trial found and repaired a Task Capsule Host
+defect the shared contract had hidden.
 
 The current IDE exception is:
 

@@ -5,8 +5,9 @@ canonical: true
 
 # Benchwork 0.3.0rc2 — The Invitation
 
-`0.3.0rc2` adds the auditable macOS/Linux/WSL installer without changing the
-M17 scientific control-plane contracts.
+`0.3.0rc2` adds the auditable macOS/Linux/WSL installer. The M17 scientific
+control-plane contracts are unchanged apart from one additive optional field,
+recorded under Fixed below.
 
 ## Added
 
@@ -19,6 +20,23 @@ M17 scientific control-plane contracts.
 - installer-owned repair, rollback, PATH blocks, and uninstall
 - release assets, SBOM, provenance, GitHub Pages publication, and installer CI
 
+## Fixed
+
+`benchwork_open_task` minted every Task Capsule for the `codex` Host, so a
+non-Codex Tier 1 Host could not record truthful provenance: Athanor correctly
+refused the Agent Result with `VALIDATION_REJECTED`. The tool now accepts the
+same optional `host_session` argument `benchwork_complete_task` and
+`benchwork_fail_task` already took, and mints the Capsule for that Host.
+
+This is an additive optional field under the
+[MCP API Policy](../MCP_API_POLICY.md). The 38 frozen tool names are unchanged,
+the default remains `codex` so existing clients behave exactly as before, and
+an unregistered Host is now refused at the boundary instead of being minted and
+later rejected by Ward.
+
+The Claude Code CLI Tier 1 trial that found this is published under
+[`examples/phase2-final/host-claude-code`](../../../examples/phase2-final/host-claude-code/README.md).
+
 ## Safety
 
 The installer does not require root, mutate system Python, create research

@@ -0,0 +1,99 @@
+# Claude Code CLI Tier 1 Host trial
+
+This directory records the 2026-07-31 Tier 1 Host validation for Claude Code
+CLI, which closed the `PENDING_HOST_VALIDATION` state left open by the Phase 2
+acceptance. It contains exported, non-canonical evidence only. No `.benchwork/`
+state file and no conversation reasoning is included here.
+
+## Method
+
+Each trial was a fresh, ephemeral Claude Code CLI `2.1.220` conversation
+started with `--strict-mcp-config`, so the Benchwork MCP server was the only
+one loaded. Every trial ran against a temporary project outside this
+repository; the source repository never acquired a `.benchwork/` directory.
+
+Transcripts were captured as `stream-json` and distilled into
+[`trial-log.json`](trial-log.json). The Host session identifier was pinned with
+`--session-id`, so the `invocation_id` recorded in Chronicle can be checked
+against the transcript that produced it.
+
+## What the trial found
+
+Read-only discovery passed immediately. The MCP server reported `connected`
+and the Host discovered all 38 registry tools. `benchwork_status` and
+`benchwork_next_actions` each returned `ok: true` with an
+`mcp-tool-result/1.0` envelope, both when the tools were named explicitly and
+when the Host was given only the broad request "What should we do next for this
+study?" with 26 non-MCP tools still available.
+
+The bounded Task loop did not pass. `benchwork_open_task` minted every Task
+Capsule with a hardcoded `codex` Host and exposed no parameter for the acting
+Host, so a truthful `claude-code` provenance was refused:
+
+```text
+VALIDATION_REJECTED: Agent Result provenance Host does not match its Task Capsule
+```
+
+Athanor was right to refuse. The defect was that the MCP boundary gave a
+non-Codex Host no way to describe itself, in a module whose own docstring
+claims to be host-neutral. The Host reported the rejection rather than
+resubmitting a false `codex` provenance to force a green run.
+
+This is what the evidence policy means when it says contract inspection is
+insufficient. `hosts.py` already exported a `ClaudeCodeHostAdapter`, the Task
+Capsule schema already enumerated `claude-code`, and Ward already admitted it.
+Only a live Host trial reached the one line that did not.
+
+## The repair
+
+`benchwork_open_task` now accepts the same optional `host_session` argument
+that `benchwork_complete_task` and `benchwork_fail_task` already took, and
+mints the Capsule for that Host. The default remains `codex`, so existing
+Codex flows are unchanged, and an unregistered Host is refused rather than
+minted and later rejected by Ward.
+
+## Result
+
+The repaired loop completed end to end. Chronicle recorded the Capsule and the
+accepted Agent Result under the acting Host:
+
+| Item | Value |
+| --- | --- |
+| Program | `RP-001`, Receipt `RC-3913CD0AC6A6` |
+| Task | `TK-B280B5160682`, capability `bench.code.inspect` |
+| Capsule Host | `claude-code` |
+| Circle | `read` only, no network, 900 s |
+| Ward | `PASS` |
+| Completion | `ok: true`, Receipt `RC-47D74468807A` |
+| Output | `code-inspection-result/1.0` |
+
+The Host used only `Read` alongside the Benchwork tools, matching the Capsule
+Circle. Deep Doctor reported 2 verified Chronicle events with an intact Receipt
+chain. The canonical provenance excerpt is in
+[`agent-result-provenance.json`](agent-result-provenance.json).
+
+## Verify
+
+```bash
+sha256sum --check examples/phase2-final/host-claude-code/SHA256SUMS
+
+jq -e '.verdict.tier_1_claude_code_cli == "PASS"' \
+  examples/phase2-final/host-claude-code/trial-log.json
+
+jq -e '.agent_result.provenance.host == "claude-code"' \
+  examples/phase2-final/host-claude-code/agent-result-provenance.json
+
+python3 -m unittest \
+  tests.mcp.test_runtime.MCPRuntimeTest.test_declared_host_owns_its_capsule_and_completes_the_task \
+  tests.mcp.test_runtime.MCPRuntimeTest.test_capsule_defaults_to_codex_and_rejects_a_foreign_host_result \
+  tests.mcp.test_runtime.MCPRuntimeTest.test_unknown_task_host_is_refused
+```
+
+## Scope
+
+This trial closes Tier 1 for Claude Code CLI only. `HOST-IDE-001` remains
+`BLOCKED_BY_ENVIRONMENT` and External Review remains
+`WAITING_FOR_DISCLOSURE_AUTHORIZATION`; neither was exercised here. The
+inspection output also recorded a residual risk worth carrying forward: the
+Circle preservation requirement in the Host adapter README is documentation
+prose, not an enforced mechanism.
@@ -0,0 +1,2 @@
+f032311827150fa5a728df72ea1842b0aa32f6c02e978d4b0b7f038a4e144f45  examples/phase2-final/host-claude-code/agent-result-provenance.json
+56e8b4a0b59e5d4eb22606fd7a09aff6237d24bcb7755f4c461de836d0bf250a  examples/phase2-final/host-claude-code/trial-log.json
@@ -0,0 +1,25 @@
+{
+  "agent_result": {
+    "outputs": [
+      {
+        "blob_sigil": "sha256:f718355fa44902306d2b0680e940769b99c19a91b3f42af56ae4a74faea10860",
+        "schema": "code-inspection-result/1.0",
+        "uri": ".benchwork/mcp/task-results/TK-B280B5160682-f718355fa449.json"
+      }
+    ],
+    "provenance": {
+      "host": "claude-code",
+      "invocation_id": "44444444-4444-4444-8444-444444444444",
+      "model": "claude-opus-5",
+      "provider": "anthropic",
+      "runtime": "claude-code-cli/2.1.220"
+    },
+    "schema_version": "agent-result/1.1",
+    "status": "COMPLETED"
+  },
+  "capsule_host": "claude-code",
+  "event_id": "CE-8E61DDA3265E",
+  "exported_from": "Chronicle event payload of the round B Task completion",
+  "object_id": "TK-B280B5160682",
+  "occurred_at": "2026-07-31T08:46:31.240296+00:00"
+}
Original file line number	Diff line number	Diff line change
		@@ -0,0 +1,2 @@
		f032311827150fa5a728df72ea1842b0aa32f6c02e978d4b0b7f038a4e144f45 examples/phase2-final/host-claude-code/agent-result-provenance.json
		56e8b4a0b59e5d4eb22606fd7a09aff6237d24bcb7755f4c461de836d0bf250a examples/phase2-final/host-claude-code/trial-log.json