Skip to content

srstack v0.1.3

Choose a tag to compare

View all tags
@tomismeta tomismeta released this 18 Sep 22:56
v0.1.3
9d82fcd

Easier installation and direct helper commands

srstack v0.1.3 improves the ergonomics of the independent, read-only Standard Reserve skill. Financial models, canonical parameters, planner schema/model 1 and live-reader targets are unchanged.

What improves for users and agents

  • Install a complete, verified bundle. The README now separates normal pinned-release installation from reviewed-commit/isolated review. The normal workflow checks the runtime ZIP checksum, safely stages and verifies it, preserves the old installation outside discovery roots, and restores it if final verification fails. It does not require downloading the audit ZIP.

  • Use explicit helper arguments instead of stdin plumbing. New commands cover protocol/auction/charter snapshots, price and gross-value reads, and the fixed fictional example. No-argument JSON stdin remains supported; calculation and result schemas are unchanged.

    python3 -B -I scripts/snapshot.py charter --id 1
python3 -B -I scripts/price.py --amount-standard 2220.9
python3 -B -I scripts/scenario.py --example

    These are example inputs, not a claim about your position. Live reads and helper execution still require normal host permission.

  • Verify first, smoke only when requested. python3 -B -I scripts/verify.py checks complete runtime membership and hashes without launching a child or fetching data. --offline adds only the fictional example and rejects live flags. --charter 1 --price explicitly checks one live charter plus its gross pending-balance valuation. Stage status, bounded failure details and elapsed time distinguish integrity failures from helper failures; incomplete checks never pass.

  • Know what a quick test should do. The docs cover the three-route menu, offline explanations, direct inspection and the common charter USD question: one charter snapshot, then one price read using the exact pending amount. Stage timing is not a promise about model, discovery or approval latency.

Update your installed copy

Use the attached runtime ZIP, verify its entry in SHA256SUMS, and follow the normal pinned-release installation instructions. Preserve customizations in the retained backup; do not overlay an old install, install a full repository/source-code archive, or install the audit ZIP. Confirm the actual loaded path/version in a fresh host conversation.

The verifier establishes integrity against its manifest, not authenticity. Review/pin the release and trust the outer checksum before executing package code. There are no wallet operations, dependency installers, credential lookups, telemetry or permission bypasses in the runtime helpers.

Scope and verification

Release commit: 9d82fcd. The runtime contains 45 files, 12 source records, 115 parameters and 14 entities.

  • CI: all 136 tests per job passed on Python 3.10 and 3.14, including verifier containment/failure boundaries and deterministic packaging. Final candidate run.
  • Runtime and installer: default/offline diagnostics passed with network denied; nine installer success/rejection/rollback cases passed using the actual README shell block with local release URLs. The fictional example's complete output matched v0.1.2 byte-for-byte. Live snapshots, quotes, cross-check, exact-amount gross valuation and the combined diagnostic were exercised. An initial protocol RPC HTTP failure and its later successful same-command invocation are both retained.
  • Native OMP: menu and packaged research required no live helpers; charter USD used one snapshot plus one price call with the exact amount; the fictional example ran once; the denial scenario refused wrappers/PTY/YOLO without tools. The host truncated the fictional numerical display, so the model did not invent outcomes or rerun it; complete helper output was independently checked. These targeted checks are not full Hermes/OpenClaw model-host certification.
  • Pinned repeat scans: Cisco Skill Scanner 2.1.0 strict static/behavioral analysis reported 0 high/critical and 3 medium findings; all 45 files were independently inventoried, with recursive-reference depth limits retained. The normal unmodified Hermes community guard allowed installation without force, with 3 medium warnings; discovery and all 45 resources in default/raw/preprocessed modes passed. Local ClawHub moderation was clean; its nonfatal Bun diagnostic is retained. Gitleaks 8.30.1 found no secrets in the runtime, all 21 reachable commits, or the final scanned evidence/assets.

The initial candidate was blocked by Hermes for README installer environment reads. The final installer takes explicitly selected paths as positional arguments instead; the unchanged guard then passed. The rejected attempt, final findings and dispositions remain in the audit evidence. No findings were suppressed. Scanner tooling ran with isolated state, verified source/dependencies, and kernel-enforced home/write/network boundaries; this is bounded release evidence, not a smart-contract audit or sandbox guarantee.

Release assets

Tagged SKILL.md is available for inspection; importing that one file is not a complete installation. Published v0.1.2 remains unchanged.

Assets 5
Loading