[ViperJuice]

Executive merge status — 2026-10-19: pending corrected Gate A and review.

The current head is 7b340871dbbcf0ef0999e596177aa97a4504982d. Separate clean-room run 39292281085 failed on the prior head because the full-suite venv lacked the optional SDK parser dependency. That failure is retained. The correction installs this same wheel's claude-sdk extra only after the existing base-install probes, without changing runtime code, test selection or assertions. The two local Gate A unit checks and shell syntax check passed.

Corrected full Gate A run 39303547973 and a fresh presidential delta review are pending. The previously accepted original full-suite and scoped live evidence retain their original commit bindings; their carry to this CI-only correction is under review. Earlier statements of six green required checks and merge readiness describe the earlier snapshot, not a completed current-head gate.

The PR remains draft. The owner authorized executive merge once current gates hold; no merge, default promotion or panel activation has occurred. The broader agent-harness#1391 remains open.

Adds explicit PHASE_LOOP_CLAUDE_ROUTE=sdk / claude_sdk selection for solo Claude authoring, execution and repair through the optional claude-agent-sdk==0.2.167 extra and the operator's installed Claude Code CLI. The launcher reuses its permission policy, context bundles, supervisor and closeout validation. Existing defaults, print/Channel/Agent View routes and panel SDK refusal retain their behavior.

claude_sdk_adapter.py admits supported personal Pro/Max contexts, binds approved configuration for CLI checks and inference, and emits normalized terminal output and metadata-only route receipts. The candidate preserves tested literal email addresses, fenced/inline code and Markdown separators; final review identified a remaining instruction-source admission defect. Context-file delivery keeps the project writable and binds only the context snapshot read-only. The README documents installation, limits and rollback. uv.lock is regenerated, and the test group installs the SDK so credential-free tests exercise its actual subprocess transport.

The owner-approved execution addendum includes the two required helper-launch inventory entries and the contributor changelog note. Both launches use the existing scratch-relocation policy; no exception or acceptance waiver is added.

Validation on the current candidate:

• Supported runner: locked dependency refresh and uv lock --check pass; the expanded prescribed targeted suite reports 446 passed, 1 skipped, 18 deselected. The skip is the opt-in live qualification test.
• Python 3.10 wheel built and tested outside the checkout with PYTHONPATH unset: 70 passed, 1 skipped. Core-only launcher import leaves the SDK absent/unimported; both refusal/dry-run checks and installed helper --help pass.
• Lint, entry-document, whitespace and owned-diff secret checks pass. The SDK and launch-inventory tests independently report 95 passed, 1 skipped.
• Full standalone verification has not passed. Local attempts encountered fixture ownership and temporary-directory prerequisites; failed runner artifacts are retained. Five required PR checks passed on 3be72b6; the separately dispatched Gate A and Python matrix are still running. Final round-three review returned one PARTIALLY AGREE with an unresolved instruction-source admission blocker, alongside native AGREE and two carried round-two AGREE judgments. The three-round cap is exhausted: review has not converged and no implementation ratification was issued. Implementation review is through T3 under the owner's override, capped at three rounds, and is not represented as AH-runtime approval.
• Live qualification remains unmet. An earlier candidate reached initialization and was rate limited. The current qualification context refuses an unsupported instruction layout before authentication. No user settings or instructions were removed to force qualification, and no successful receipt or account-charge measurement is claimed.

This remains a blocked draft, halted for operator follow-up under the declared review cap. The admission finding, full standalone verification, final implementation review and a non-skipped authoring/execution/cancellation qualification receipt must be resolved before landing. No default promotion or panel SDK activation is included.

Follows the Stage 3a detailed plan and subscription amendment ratified by agent-harness#1732, with the owner-approved execution addendum. Tracks agent-harness#1391 without closing its carried work.

---

^{Need help on this PR? Tag @codesmith-bot with what you need.}

Implementation continuation update:

Adds explicit PHASE_LOOP_CLAUDE_ROUTE=sdk / claude_sdk selection for solo Claude authoring, execution and repair using optional claude-agent-sdk==0.2.167 and the operator's installed Claude Code CLI. The launcher reuses its permission policy, context bundles, supervision and closeout validation. Existing route defaults and panel SDK refusal retain their behavior; selecting an existing route is the rollback.

The helper binds approved configuration and instructions for CLI checks and inference, admits supported personal Pro/Max contexts, and emits normalized terminal output and metadata-only receipts. The latest delta aligns instruction admission with project ancestry and preserves supported plain instructions. Context delivery binds only the frozen context file and keeps the project writable. The README explains installation and supported contexts; uv.lock is regenerated. The owner-approved plan addendum includes the required process-launch inventory entries and CHANGELOG note, with no scratch-policy exception or acceptance waiver.

Observed verification on this candidate:

• Supported runner: locked test/SDK dependency refresh and uv lock --check pass; prescribed expanded targeted suite 450 passed, 1 skipped, 18 deselected. The skip is opt-in live qualification.
• SDK and launch-inventory suite: 99 passed, 1 skipped. Four new admission regressions observed 2 failed, 2 passed before the helper fix and pass afterward.
• Python 3.10 wheel built and tested outside the checkout with PYTHONPATH unset: 74 passed, 1 skipped. Core-only install keeps SDK absent/unimported, imports the launcher and passes 2 dry-run/refusal checks. Installed helper --help, lint, whitespace and owned-diff secret checks pass.
• Full prescribed standalone verification has not passed. Failed local runner artifacts are retained; workspace-ownership and temporary-directory prerequisites were diagnosed. Gate A and the Python matrix were dispatched on parent 3be72b6; those results do not silently replace the prescribed full-run requirement. Five required PR checks passed on that parent.
• Live qualification remains unmet. Updated bound preflight on claw still refuses an unsupported instruction layout before authentication. An older candidate reached initialization and received HTTP 429. No successful authoring/execution/cancellation receipt or account-charge measurement is claimed; no user instructions were removed to force qualification.

The owner explicitly authorized continued review to convergence beyond the original three-round cap and required a president. T3/manual review continues under the owner's override while AH review is broken. The dissenting seat is reviewing this delta; prior standing AGREE judgments are carried with their original scope. Claude president intake was rate limited, so GPT-6 Astra is the availability-based president; a fresh final ruling will follow convergence on the observed implementation HEAD. No AH-runtime approval is claimed.

This remains a draft. Code-review convergence, the president ruling, successful prescribed full verification and a non-skipped live qualification receipt are required before landing. No default promotion or panel SDK activation is included.

Follows the Stage 3a plan/amendment ratified by agent-harness#1732 and its owner-approved execution addendum. Tracks agent-harness#1391 without closing its carried work.

Final code review update:

Adds explicit PHASE_LOOP_CLAUDE_ROUTE=sdk / claude_sdk selection for solo Claude authoring, execution and repair using optional claude-agent-sdk==0.2.167 and the operator's installed Claude Code CLI. The launcher reuses its permission policy, context bundles, supervision and closeout validation. Existing route defaults and panel SDK refusal retain their behavior; selecting an existing route is the rollback.

The helper binds approved configuration and instructions for CLI checks and inference, admits supported personal Pro/Max contexts, and emits normalized terminal output and metadata-only receipts. The latest delta aligns instruction admission with project ancestry and preserves supported plain instructions. Context delivery binds only the frozen context file and keeps the project writable. The README explains installation and supported contexts; uv.lock is regenerated. The owner-approved plan addendum includes the required process-launch inventory entries and CHANGELOG note, with no scratch-policy exception or acceptance waiver.

Observed verification on this candidate:

• Supported runner: locked test/SDK dependency refresh and uv lock --check pass; prescribed expanded targeted suite 450 passed, 1 skipped, 18 deselected. The skip is opt-in live qualification.
• SDK and launch-inventory suite: 99 passed, 1 skipped. Four new admission regressions observed 2 failed, 2 passed before the helper fix and pass afterward.
• Python 3.10 wheel built and tested outside the checkout with PYTHONPATH unset: 74 passed, 1 skipped. Core-only install keeps SDK absent/unimported, imports the launcher and passes 2 dry-run/refusal checks. Installed helper --help, lint, whitespace and owned-diff secret checks pass.
• Full prescribed standalone verification has not passed. Failed local runner artifacts are retained; workspace-ownership and temporary-directory prerequisites were diagnosed. A genuine prescribed full command is now running through the supported runner on the same candidate in an isolated non-root Python 3.10 environment on claw, with valid workspace ownership, direct temporary storage and sandbox prerequisites. An initial isolated collection failure from missing jq was retained; the dependency was installed only in the test image before the distinct new run. Gate A and the Python matrix were dispatched on final candidate 0c3d5f9 (run 39056551969); their results do not silently replace the prescribed full-run requirement. Final CI remains under assessment.
• Live qualification remains unmet. Updated bound preflight on claw still refuses an unsupported instruction layout before authentication. An older candidate reached initialization and received HTTP 429. No successful authoring/execution/cancellation receipt or account-charge measurement is claimed; no user instructions were removed to force qualification.

The owner explicitly authorized continued review to convergence beyond the original three-round cap and required a president. T3/manual review continues under the owner's override while AH review is broken. Code review has converged: Grok round four is fresh AGREE; native Astra round three and both Gemini round-two judgments are carried AGREE with their original scope. Claude president intake was rate limited, so GPT-6 Astra served as the availability-based president. The fresh president independently inspected observed HEAD 0c3d5f9 and all frozen input hashes, ruling AGREE on code ratification, DISAGREE on landing/activation admission until full verification, CI and live qualification are satisfied. No AH-runtime approval is claimed.

This remains a draft. Code review and the president code ruling are complete; successful prescribed full verification, final CI and a non-skipped live qualification receipt remain required before landing. No default promotion or panel SDK activation is included.

Follows the Stage 3a plan/amendment ratified by agent-harness#1732 and its owner-approved execution addendum. Tracks agent-harness#1391 without closing its carried work.

Main integration and verification update:

Adds explicit PHASE_LOOP_CLAUDE_ROUTE=sdk / claude_sdk selection for solo Claude authoring, execution and repair using optional claude-agent-sdk==0.2.167 and the operator's installed Claude Code CLI. The launcher reuses permission policy, context bundles, supervision and closeout validation. Existing route defaults and panel SDK refusal retain their behavior; selecting an existing route is the rollback.

The helper binds approved configuration/instructions for CLI checks and inference, admits supported personal Pro/Max contexts, and emits normalized terminal output and metadata-only receipts. Context delivery freezes the context file while keeping the project writable. The README documents installation, supported contexts and rollback; the lockfile includes the optional dependency. Required scratch and static reference inventories classify the helper's existing probes and I/O sites under existing policy; the plan records the verification inventory and contributor CHANGELOG note.

This update integrates current main at e980717d5 following agent-harness#1738. Both changelog entries are preserved. Upstream adapter-reference delivery changes are retained, and nine SDK-helper entries complete the static launch-reference inventory. The SDK helper and its regression-test bytes are unchanged from the code-ratified candidate 0c3d5f9.

Observed validation on the combined candidate:

• Supported runner: actual locked test/SDK dependency refresh, uv lock --check and the expanded SDK, legacy-route, supervisor, adapter-reference and inventory suite pass: 466 passed, 1 skipped, 18 deselected. The skip is opt-in live qualification. The additional static inventory was observed failing before its missing declarations and passing afterward.
• Python 3.10 wheel built and tested outside the checkout with PYTHONPATH unset: 74 passed, 1 skipped. Core-only import keeps SDK absent/unimported and passes 2 dry-run/refusal checks. Installed helper --help, whitespace checks and the staged-diff secret scan pass.
• Full prescribed standalone verification remains unmet. A genuine isolated supported-runner attempt is still running on the earlier 0c3d5f9 candidate; its outcome will be recorded at that scope. Earlier failed artifacts are retained. Final-candidate CI and Gate A must be assessed after publication; CI does not silently replace the prescribed full command.
• Live qualification remains unmet. Claw's existing linked user-instruction file is outside the current supported layout and is refused before authentication. No instructions were removed to force qualification. The older HTTP 429 supplies no successful authoring, schema-valid execution or cancellation receipt, and no account-charge measurement is claimed.

The owner authorized review to convergence beyond the initial cap and required a president. Four standing code judgments agreed on 0c3d5f9 (Grok fresh; native Astra and two Gemini judgments carried at their recorded scope). A fresh Astra president independently ratified that code, while rejecting landing/activation admission until verification and live gates pass. Claude president intake was rate limited and cast no vote. The president's assessment will be refreshed on the observed synchronization commit; prior approval is not relabeled as a new-head approval. Review is manual/T3 under the owner's override while AH review is broken; no AH-runtime usable assertion is claimed.

This remains a draft. Successful prescribed full verification, final CI, refreshed final-candidate review and non-skipped live qualification are required before landing. No default promotion or panel SDK activation is included.

Follows the plan/amendment ratified by agent-harness#1732 and the owner-authorized execution continuation. Tracks agent-harness#1391 without closing carried work.

Python 3.12 integration and verification update

The optional solo SDK route retains explicit selection, the pinned optional
dependency, existing supervised execution, and unchanged legacy routes. This
revision incorporates upstream agent-harness#1549 and preserves its Python 3.12
minimum, CI lane and chronology controls. Hosted, Dagger and local suite installs
include the SDK extra; core-only installation remains supported.

Current Python 3.12 verification: supported targeted runner 640 passed, 1 live
skipped, 18 deselected, with successful lock check and locked dependency refresh.
The sdist and wheel built. Isolated installed-wheel checks outside the checkout:
SDK extra 75 passed, 1 live skipped; core-only 28 passed, 48 skipped, including
clean SDK dependency-absence behavior. Pinned lint and staged audit passed.

Cross-vendor code review converged on the unchanged SDK implementation, and the
presidential integration intake found no code blocker. This remains a draft:
fresh committed-candidate presidential binding, current-candidate CI, the original
full standalone suite, and non-skipped live authoring/execution/cancellation
qualification remain required. Earlier failed full-suite artifacts are retained;
the corrected Python 3.12 environment must produce a new result. The live context
layout currently refuses before authentication. No activation or landing approval
is claimed. Rollback remains selecting an existing Claude route.

Main ancestry integration

This update records the incorporated upstream main as an ancestor and resolves
the lockfile merge conflict. Its complete source tree is byte-identical to the
previously reviewed and published candidate. The observed Python 3.12 targeted
and installed-wheel results above carry at that identical source scope; no tests
are reclassified as a new commit's CI result. The supported publication adapter
publishes the genuine merge commit in its existing prebuilt mode.

The PR remains a draft. Fresh committed-head presidential binding, current-head
CI/Gate A, full standalone verification and non-skipped live qualification remain
required. No landing or activation approval is claimed.

Current CI verification — 63ac991

All six required PR checks passed. The separately dispatched workflow on this
exact commit also completed successfully, including Python 3.12 pytest, the
standalone-from-wheel Gate A clean room, hosted fallback and suite gate:
https://github.com/Consiliency/agent-harness/actions/runs/39072141160

The prescribed full standalone runner, including its original chronology scope,
is still running on claw. Live subscription qualification remains unmet. The
president approved the code at this exact commit; landing and activation remain
unapproved while those acceptance requirements are outstanding.

Full standalone verification — container prerequisite repair

The original full standalone run completed with 166 failures, 213 setup errors,
13,569 passes, 712 skips and 598 deselections. Lock checking and locked dependency
refresh succeeded; the suite exit status is 1 and the failed artifact is retained.
The container's successful wrapper exit is not reported as a suite pass.

The isolated image lacked ip and socat, producing the dominant helper-admission
errors. A corrected image now validates all 14 required helpers and creates a real
filtered network namespace. A focused rerun of the 379 failed/error cases is in
progress before another full attempt; no source guard or fixture was weakened.
There were no SDK runner failures. Current-candidate hosted CI and Gate A remain
successful, while the prescribed full-run and live-qualification requirements
remain unmet. Presidential diagnosis of the residual groups is in progress.

Focused residual verification update

The corrected network environment reran the 379 failed/error cases: 327 passed,
52 failed, no setup errors. The original full-suite failure remains binding.
A fresh isolated diagnostic now runs the 52 residual cases with real historical
inputs, the installed command on PATH, the required toolchain, Python 3.12.14,
visual dependencies and an adequate descriptor limit. Presidential diagnosis
identified both missing prerequisites and inherited code/test issues; remaining
failures are not assumed to be environmental. No SDK source, guard or acceptance
criterion changed, and no live qualification or full-suite pass is claimed.

Current implementation and review update

The explicit solo Claude SDK route now preserves supported linked user
instructions and their explicit dependencies as a complete private snapshot.
Unsupported or changing instruction layouts refuse before provider execution.
The optional extra pins SDK 0.2.167 and parser 4.2.0; core-only installation,
existing route defaults, panel behavior and rollback through existing routes
remain supported. The source incorporates the 0.7.27 release and reviewed
upstream provider changes; the SDK addition remains Unreleased.

Current supported targeted verification: 1,320 passed, 1 live skipped,
18 deselected, including a successful lock check and locked dependency
refresh. Clean installed wheel: 615 passed, 1 live skipped with the extra;
11 passed for core-only absence controls. Additional upstream checks:
140 passed for the imported publisher and inventory. Earlier provider and
network checks passed 294 tests, 1 skipped on the unchanged shared files.
Build, lint, package metadata and lazy-import checks
passed. A metadata-only offline capture preserved the complete approved
instruction pair. Independent Claude and presidential source reviews agree
on the frozen implementation, with the previously recorded compatibility
limits retained.

This PR remains a draft. The original full-suite failure is retained, and a
fresh run of that full command, non-skipped live authoring/execution/cancellation
qualification and presidential binding of the actual published commit remain
required. Current targeted and offline results do not satisfy those gates.
No landing, default promotion or panel activation approval is claimed.

Entry documentation correction

Clarify that the SDK instruction dependency belongs to the selected user configuration. The exact entry-documentation check now passes for all eight entry docs. Runtime code is unchanged; the full-suite run remains bound to the preceding implementation commit, and live qualification remains unmet after a provider rate-limit refusal. This PR remains a draft.

Current main integration and verification

The SDK branch incorporates main at 7b49e86 through ordinary merge 2da1322.
Only the incoming repository configuration and changelog differ from d677c5e;
the accepted configuration blob is retained exactly. SDK/runtime source, tests,
dependencies, lockfile and plan bytes are unchanged. Fresh independent Daybreak
delta review agrees with no corrections.

New supported commands-only validation passed: locked dependency refresh, lock
check, exact head/tree/blob identity, whitespace, all eight entry documents and
documentation freshness. This is not a fresh local test-suite result. The earlier
d677c5e targeted result (1,322 passed / 1 live skipped / 18 deselected) and
publishing result (165 passed) retain their historical bindings.

Fresh required CI, scrub and Gate A are pending at exact 2da1322:

• PR test: https://github.com/Consiliency/agent-harness/actions/runs/39377999912
• Scrub: https://github.com/Consiliency/agent-harness/actions/runs/39377999986
• Dispatched Gate A: https://github.com/Consiliency/agent-harness/actions/runs/39378014586

The prior d677c5e scrub run 39371031254 remains recorded as failed; incoming-main
source identity does not replace a fresh successful scan on this head. Earlier
Gate A failure 39292281085 and parent success 39303547973 retain their exact
historical bindings. The r19 full success at ab3ca88 (14,861 passed / 639 skipped /
598 deselected / 1 xfailed) and scoped original187 frontier live receipt also
remain historical. No new live execution or billing measurement is claimed.

The PR remains draft. Final presidential evidence binding and the executive's
merge/closeout procedure await genuine current-head results. This author has
not merged main, changed route defaults or activated panel SDK execution.

[ViperJuice]

Disposition under OD-2 (specs/phase-plans-seats-v1.md, INVENTORY phase):

Salvage the SDK adapter, the claude-sdk extra pin and the Gate A clean-room install fix into CLAUDE lane C. The solo-runner route re-files on agent-harness#1391.

Closed by: EC-CLAUDE-3.