Hermeneutic 0.1.12 — native Gemini CLI response gate

Adds a native Gemini CLI AfterAgent extension that checks the proposed response with Hermeneutic’s deterministic drift gate and requests one bounded automatic retry for medium- or high-risk output. The extension was exercised from a clean copied install against Gemini CLI 0.32.1; malformed input and internal failures remain explicitly fail-open. The source distribution now includes the complete Gemini extension contract.

Hermeneutic 0.1.11 — native Hermes Agent output gate

What's Changed

• Add an opt-in native Hermes Agent final-output gate plugin in #23
• Prepare the 0.1.11 patch release in #24

Install hermeneutic into the same Python environment as Hermes Agent, then enable it with hermes plugins enable hermeneutic.

The plugin runs locally on Hermes Agent's transform_llm_output hook. Clean and low-severity responses are unchanged; medium/high evidence obligations append a bounded advisory before delivery.

Full Changelog: v0.1.10...v0.1.11

Hermeneutic 0.1.10 — cited correction evidence

What's Changed

• ci: pin actions and restrict workflow permissions by @roli-lpci in #14
• test: guard CI workflow hardening by @roli-lpci in #15
• Add assertion-level evidence citations to compile output by @roli-lpci in #16
• docs: link Hermeneutic product page by @roli-lpci in #17
• feat: add deterministic version command and LintLang SARIF dogfooding by @roli-lpci in #18
• build(deps): bump actions/checkout from 7.0.0 to 7.0.1 by @dependabot[bot] in #19
• build(deps): bump github/codeql-action/upload-sarif from 4.37.6 to 4.37.9 by @dependabot[bot] in #20
• build(deps): bump hermes-labs-ai/lintlang from 0.4.1 to 0.5.0 by @dependabot[bot] in #21
• chore: prepare Hermeneutic 0.1.10 by @roli-lpci in #22

Full Changelog: v0.1.9...v0.1.10

Hermeneutic v0.1.9 — Apache-2.0 successor release

What's Changed

• Correct Hermeneutic research metadata relationships by @roli-lpci in #10
• chore: migrate license to Apache-2.0 by @roli-lpci in #11
• Report malformed compile-index triples cleanly by @roli-lpci in #12
• Prepare Hermeneutic 0.1.9 Apache release by @roli-lpci in #13

Full Changelog: v0.1.8...v0.1.9

v0.1.8 — verified packaging and release maintenance

Hermeneutic v0.1.8 is a maintenance release for installation and packaging reliability. It does not change gate rules, correction mining, retrieval, Router behavior, CLI semantics, or integration support.

What improved

• CI and release jobs now exercise the installed wheel and the exact source distribution.
• Release actions are immutably pinned, and the workflow rejects a release tag that does not match the package version.
• GitHub Actions maintenance versions and the package Documentation link are current.
• Package, runtime, citation, Zenodo, SBOM, security, and installation metadata agree on 0.1.8.

Upgrade

python3 -m pip install --upgrade hermeneutic

Hermeneutic remains a local-first review aid: eight deterministic English surface checks, correction-log mining, optional local retrieval, and an explicitly configured Python Router. It is not proof of correctness or a replacement for human review.

v0.1.7 — the observable gate + forward-deployed harness

Publication truth corrections

• Reframed the product around its actual distinct surfaces: personal correction mining, optional personal retrieval, a fixed deterministic English gate, and a separately configured Python Router.
• Replaced the Claude compile hook's user-visible systemMessage with the documented UserPromptSubmit hookSpecificOutput.additionalContext contract. The installer now registers an exec-form command bound to the running Python interpreter, handles paths with spaces, migrates or cleans its exact legacy registration, preserves unrelated hooks, and avoids destructive malformed-settings behavior.
• Added exact compile-hook contract coverage for success JSON, empty/missing results, unavailable Ollama, malformed input/state, compiler failures, timeouts, idempotency, foreign-file protection, legacy cleanup, coexistence, and uninstall.
• Replaced the leave-one-out evaluator's duplicated selection approximation with calls through the shipped compile_prompt() production path. Current same-bucket results on the frozen 346-triple corpus are 88/104 (84.6%) at CLI/hook defaults and 94/104 (90.4%) at Python defaults.
• Added a privacy-safe current gate-coverage receipt: 115/346 (33.24%) prior assistant replies in the frozen correction corpus fire at least one current rule. This is retrospective derivation-set coverage, not held-out accuracy.
• Moved the prior 83.7% leave-one-out and 98/100 discrimination figures into historical-only framing because their evaluators did not match the shipped selection path.
• Classified every integration by evidence maturity. The Claude compile hook, Codex plugin mechanics, and forward-deployed tooling are mechanically tested; Cursor/OpenHands are design sketches; current Claude Stop, Cline, and Windsurf adapters are removed from the ready-support matrix.
• Synchronized README, package description, machine-readable summary, examples, theory/intent docs, citation metadata, Zenodo metadata, plugin descriptions, installation commands, optional dependencies, privacy boundaries, and author display.

Added and hardened

• Eight fixed English stage-one rules with a partial-progress contrast guard.
• Correction mining for Claude Code, Codex rollout, and OpenAI-style message logs; loud zero-parse diagnostics; multi-directory input; nested output-directory creation.
• Optional local-Ollama compile index and deterministic bucket-aware warning preambles.
• Local opt-in telemetry, reviewable audit modes, stats, harvest classification, human review, and accepted-record promotion.
• Wheel-built Claude compile-hook installer plus repository/sdist standalone plugin assets and integration documentation.
• Forward-deployed environment/boot/harvest/report/gate tooling, report leak checks, tamper-evident mission state, and optional Codex notification sentinel.
• Trusted-publishing GitHub Actions workflow, wheel/sdist checks, exact-sdist tests, typed-package smoke, release metadata, and software citation files.

Boundaries

• Core library and CLI have zero required Python runtime dependencies.
• Stage-one checks are fixed English surface patterns. Mining and promotion do not generate regex rules.
• Live Ollama, live response-hook hosts, external Router backends, Windows, false-positive rate on ordinary prompts, and downstream model effectiveness are not certified by this release.
• The wheel contains the core, CLI, type marker, and built-in hook installers. Evaluations, standalone plugins, integration docs, and forward-deployed tooling require the repository or extracted sdist.