Aleph Skill v2.0.1

Aleph Skill 2.0.1

Aleph 2.0.1 is a portability, protocol-correctness, security, and release-integrity update. Schema and numerical formula identifiers remain at 2.0.0, so there is no schema-number migration. The validator intentionally tightens several 2.0 contracts; existing workspaces must be revalidated and may require artifact regeneration before they pass.

Highlights

• Promotes current native Agent Skills discovery for supported IDEs and CLIs, with exact project/user paths and non-global fallbacks.
• Makes every documented command resolve from an explicit installed skill root instead of assuming the process working directory.
• Reconciles the likelihood contract: uncalibrated branches use relative_weight; calibrated probabilities require the calibration and hindcast gates.
• Adds a provenance-preserving, assurance-limited research fallback when D Research is unavailable.
• Adds resumable research checkpoints and an honest unsaturated handoff when a host execution window ends.
• Makes lifecycle acceptance part of the release gate and enforces genuine strict mypy checks.
• Keeps self-tests non-mutating by placing lint, type-check, and coverage caches outside the attested package tree.
• Reports every refused or failed installer transaction with ok: false, including Windows symlink privilege failures.
• Builds the official ZIP from the verified distribution manifest only, verifies byte-for-byte reproducibility, and publishes GitHub build-provenance attestations.
• Verifies the remote annotated tag object independently of actions/checkout, binds its commit to both the checked-out source and main, then rechecks the exact tag object and target immediately before publication.

Compatibility notes

• schema_version remains 2.0.0.
• formula_version remains 2.0.0.
• This is not input-compatible with every artifact accepted by the v2.0.0 validator. Revalidate existing workspaces before relying on them.
• In uncalibrated workspaces, node/branch/actor likelihood must use the exact relative_weight contract and must not retain probability claims.
• Material-actor packets, roleplay outputs, execution rows, and receipts must be regenerated when they do not satisfy the new packet-only semantic binding and anti-replay checks; do not hand-edit or copy hashes.
• Re-render reports that use the old “probability updates” heading, then recompile/replay numerical artifacts whose source hashes changed and finalize again to regenerate receipts.
• migrate_workspace.py remains the 1.x-to-2.0 schema migrator; it is not a substitute for this v2.0.1 revalidation/repair pass.
• The seven bundled domain packs remain explicitly experimental; they do not authorize probability claims.
• External CLI JSON profiles describe a portable adapter contract. They are not a claim of turnkey orchestration for every third-party host.

Check an existing workspace in draft mode before changing it:

python "$ALEPH_SKILL_ROOT/scripts/validate_simulation_artifacts.py" \
  --workspace <run-dir> --mode draft --write-report

After repairing reported contract errors, rerun the workspace's original compile/run/replay commands, render the report, finalize, and require a clean final validation. Preserve an untouched backup throughout the process.

Verification

The tagged release workflow must pass the full Linux, Windows, and macOS verification matrix before publishing. It runs the cross-contract release gate on Python 3.10 through 3.13, plus strict static checks, lifecycle acceptance, adversarial validation, deterministic replay, adapter drift checks, domain-pack validation, and manifest-exact packaging.

Manual cross-host forward check

The reproducible release ZIP was also exercised read-only on 2026-07-14 with Grok Build 0.2.101 (grok-4.5), OpenCode 1.17.18 through Ollama Cloud (glm-5.2, kimi-k2.6, and minimax-m3), and Cline 3.0.40 through all ten Cline Pass models exposed by the installed binary (deepseek-v4-flash, deepseek-v4-pro, glm-5.2, kimi-k2.6, kimi-k2.7-code, mimo-v2.5, mimo-v2.5-pro, minimax-m3, qwen3.7-max, and qwen3.7-plus). Every model recovered the exact likelihood, sealed-roleplay, fallback-assurance, absolute-root, package-version, and schema-version contracts and observed passing preflight and fixture validation results.

Every listed model passed the semantic audit, including a second adversarial prompt that required it to reject uncalibrated probability, roleplay networking, verified fallback assurance, and current-working-directory helper resolution. Exact single-object JSON adherence varied by host, model, and prompt: Grok prepended progress text; OpenCode GLM changed the requested schema on the adversarial form while citing all four correct violations; and members of the MiMo/MiniMax Cline Pass set intermittently wrapped or reshaped otherwise correct output. These are recorded as model instruction-following warnings, not silently counted as strict-format success. The checks demonstrate semantic interpretability of the packaged instructions; they do not grant simulation assurance or calibrate any domain pack.

Download the ZIP, distribution-manifest.json, and SHA256SUMS.txt into the same directory, then verify both downloadable files:

sha256sum -c SHA256SUMS.txt

The build-provenance attestation is stored by GitHub's attestation service rather than attached as a separate release file. With GitHub CLI installed, verify the ZIP against this repository:

gh attestation verify aleph-skill-v2.0.1.zip \
  --repo d-init-d/aleph-skill \
  --signer-workflow d-init-d/aleph-skill/.github/workflows/release.yml \
  --source-ref refs/tags/v2.0.1

Repository-level immutable releases are enabled before the tag is pushed, so GitHub locks the published tag and assets. The workflow also refuses to replace or edit an existing v2.0.1 release.

See CHANGELOG.md for the detailed change list and retained engineering limits.