[AronAxe]

Owner-approved patch release: v0.11.1

Aron authorized merging and releasing PR #20 after the formatting-only repair made the candidate green. The previous published version is v0.11.0; no existing release/tag is modified.

Reviewed compatibility fix

• Add _coerce_threshold_tokens_cap required by Hermes live compression.
• Honor lazy context-window and threshold invalidation, explicit pin removal, installed-host route metadata and per-model thresholds.
• Refresh the engine window from the current host select_context budget without resetting usage/retry state.
• Preserve isolated clone settings, internal-call exclusion, exact evidence and the strict tokenizer gate. No summarizer or destructive history rewrite.
• Nine optional real-Hermes regression cases. Commit 58b6d6c fixes only Ruff formatting, with unchanged test semantics.

Release-only preparation

Python and Rust metadata are aligned to 0.11.1. README changes are limited to the patch note, current version/install references and correcting its stale FAQ: explicitly selected TT can replace LCM as the Hermes ContextEngine; middleware mode keeps the host engine. Changelog, migration and focused release notes are updated. Schematics, artwork, runtime logic and tests remain unchanged from the green formatted candidate (apart from the version constant). Temporary preparation workflow is removed from the final tree; existing publication workflows are unchanged.

Verification

The formatted candidate 58b6d6c passed all five workflows, including Python 3.10–3.13 and Windows lint/format/pytest, Rust, packaging, learned-policy fitting/benchmark, dashboard, ContextEngine and Context IR checks. Learning artifact: 611 passed, three skips, zero failures. One skip is the entire optional live-Hermes test module, not nine tests counted as run; those cases were previously run locally against installed Hermes 188a1a5 on Python 3.14. Final release-head checks will be recorded before merging. No paid inference testing is needed for this patch.

Boundaries

This publishes the PR #20 implementation, not the separate deployed-adapter snapshot branch. Earlier incident deployment comments remain historical; this release operation will not touch a running Hermes installation or perform another restart.

Unknown-target tokenizer support and the separate host directory-discovery race are not fixed by this patch. Exact-tokenizer unavailability still causes ContextEngine/IR to pass through. Publishing or successfully loading the adapter is not proof of current-session reduction. Managed/multiplex users must update the environment their launcher actually imports, then use the normal idle-backend restart path and inspect actual current-model request metrics. No histories, vaults, keys or live configuration are altered by publication.

Release notes: docs/releases/v0.11.1.md. Existing main-push workflows publish the GitHub wheel/source/checksum assets and version-aligned Rust interoperability companion.

[AronAxe]

Live deployment update — a controlled Desktop backend restart has now been completed

The earlier statement that no live deployment was possible is superseded for the installed adapter snapshot recorded at 914021b884e5bc9fc23eb9e40c8f863b1f0fe819, branch fix/hermes-live-compression-deployed.

The main compatibility branch advanced concurrently to eebfd9d while that deployment was being recorded. Its non-fast-forward update was refused; no force push or overwrite was performed. The deployed snapshot is intentionally retained separately for reconciliation rather than claiming the two implementations are identical. It includes positive-int/None cap coercion, live cache invalidation, host threshold arithmetic including output reservation, route metadata and isolated clone state. PR #20 additionally has host-hook budget refresh/per-model threshold changes that were not part of this deployed snapshot.

Actions actually completed

• Backed up and atomically replaced the adapter in the managed plugin source actually selected by Hermes bootstrap, and in the older venv fallback copy.
• Deployed source SHA-256: 844f92fa673a361ec29d992f0e76955991e7c669a99c0c3912a1ba9a278f336f. The snapshot commit is formatting-equivalent; deployment files retain their own checksum.
• Ran 11 regression checks against the installed Hermes 188a1a5 live-config function, including output reservation, cap/pin removal, cache refresh and exact-history retention. All passed, then passed again after loading the installed patched adapter. No generative compressor was constructed.
• Verified idle admission, performed one targeted Desktop backend teardown, and let Desktop's existing supervisor start its replacement. New backend health and TT's authenticated plugin summary endpoint both return HTTP 200. Separate messaging gateway/relay/dashboard processes were not restarted. Do not repeat a restart on the assumption that deployment is still pending.
• No conversation/vault deletion, model switch, configuration rewrite or paid generation call was performed. Local backups and the deployment/test report were retained on the authorized machine.

Remaining obstruction — not a claim of full session recovery

The current chat target is not recognized by the installed exact-tokenizer mapping. TT returns tokens=None and its ContextEngine exits before reduction. The compatibility patch and healthy restart do not resolve that independent gate. Historical savings rows from supported targets were explicitly not counted as new evidence of recovery. No arbitrary tokenizer override was installed or labelled exact.

A separate explicit context-length override also disagrees with the host's automatically resolved window; it was not silently changed. The underlying host directory-discovery race remains unpatched. API availability and a successful isolated load are not proof the host-wide race has been eliminated.

The additional real-Hermes regression script and deployment boundaries are recorded in the snapshot branch. This update is not a merge, release, full-CI result, or claim that the affected current-model provider request has shrunk.

[AronAxe]

Formatting-only CI repair completed

Commit 58b6d6c2af2b9014bce658caf2f6fb0119d0f759 (style: ruff format hermes live-compression tests) changes only tests/test_hermes_live_compression.py, applying exactly the two formatting hunks from Aron's bug brief. The before/after Python ASTs and non-whitespace token sequences are identical. The fetched original blob was ddecbd5bbcda060ae2ff4dfcdd3e5a2339bcdc9f; formatted blob is 7387ababf6d74e0da4cc1c58234daf348053af30.

No changes to hermes_engine.py, test logic, dependency pins, workflows, or benchmarks/quality_ab/cases. No merge, release, deployment or backend restart.

Actual new-head CI results

• CI: success. Python 3.10, 3.11, 3.12, 3.13 and Windows each passed ruff check, ruff format --check and the actual pytest step. Package and Rust jobs also passed.
• Learned policy verification: success, including lint/format, the training-enabled suite, the three-repeat benchmark and artifact upload.
• ContextEngine verification: success.
• Dashboard verification: success.
• Context IR verification: success.

Downloaded and SHA-256 verified learning artifact 11283223742 (2e744d84a217a48bd4d04876960732a766b89d30c6c6900e47adaf8572984419). It contains both previously missing files. JUnit reports 614 collected: 611 passed, 3 skipped, 0 failures, 0 errors. One skip is the entire Hermes-dependent live-compression test module (requires a current Hermes checkout); the other two require external Hermes/RTK integrations. This is not a claim that the nine new live-config tests executed in the learning job. The existing dedicated Hermes workflow passed separately.

The benchmark JSON contains 18 completed deployment trials, zero paid calls, no enlargement, and passing protection/recovery/retry checks. These remain fixture results, not live-session repair evidence.

Local work for this change was limited to source/blob validation, applying the supplied formatter diff, AST/token equivalence and inspecting CI artifacts. The Ruff/full-suite runs reported above were executed by GitHub Actions, not a newly provisioned local venv. The original red status was the formatting blocker described in the brief; no additional functional changes were needed for these reruns.

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-04T04:17:17.715651Z	a0ac08e	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.

[chatgpt-codex-connector]

💡 Codex Review

Here are some automated review suggestions for this pull request.

Reviewed commit: a0ac08e2d6

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review".

If Codex has suggestions, it will comment; otherwise it will react with 👍.

Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".

[chatgpt-codex-connector]

Update the bundled Rust documentation for 0.11.1

Publishing this crate as 0.11.1 still uses rust/README.md as its package README, but that file identifies the crate as v0.11.0 and tells users to run cargo add token-terminator@0.11.0; docs/RUST_CRATE.md likewise contains only 0.11.0 commands and links. Consequently, the crates.io/docs guidance for the new release directs users back to the previous version instead of installing 0.11.1, so these bundled docs should be version-aligned with this bump.

Useful? React with 👍 / 👎.

[AronAxe]

Released v0.11.1 — October 4, 2026

Owner-approved merge completed. Reviewed release head: a0ac08e2d6b52dc27a4ef8644b72c9b0aaef4728. Merge and independently checked release-tag commit: f7d5cea0658ffcf62c48c521307b73c6e22b0b1f.

• GitHub v0.11.1 is published, neither draft nor prerelease, with wheel, source distribution and SHA256SUMS. Release ID: 402835163.
• Rust companion 0.11.1 is independently confirmed in the public registry: version ID 3398020, published 2026-10-04T04:17:30.383164Z, not yanked. Rust behavior is unchanged.
• Both GitHub publication and crate publication jobs completed successfully.

Final pre-merge verification

All five workflows on the exact release head completed successfully before merging:

• CI, including Python 3.10–3.13 and Windows lint/format/pytest, Rust and packaging.
• ContextEngine.
• Learned policy.
• Dashboard.
• Context IR.

Downloaded and SHA-256 checked final learning artifact 11292987547 (4bc45e9fbe1e7016a6e12e9ac631896032ae3f37803b720230791557de48a26f). Its actual JUnit report is 614 collected, 611 passed, 3 skipped, 0 failures/errors. The entire optional real-Hermes module is one skip; the nine local live-config regressions are not miscounted as executed in that job. Benchmark JSON is present and explicitly fixture-only. No paid inference calls.

Final package artifact 11293551998 was checksum-verified (9c2d96f3ae9926b50ac561b050a1930e62a1e3b1c5662f501676014c691363f6), extracted and inspected. Both wheel/sdist passed the release verifier locally as well. Relative to green candidate 58b6d6c, 46 runtime/source assets (excluding only the version constant), 37 test files, 12 scripts, 35 benchmark files and all 5 graphics are byte-identical. The three release-document unit tests also passed locally. Existing changelog/migration history was preserved.

README scope

Only the patch note, current release/install references and the stale FAQ were changed. The FAQ now correctly distinguishes TT replacing LCM when explicitly selected from middleware mode keeping the upstream engine. JEV badge, artwork and schematics are unchanged. Focused changelog, migration and release notes document the managed/multiplex environment requirement. Temporary preparation workflow is absent from the final tree; existing publication workflows were not changed.

Published integrity

The publicly retrieved checksum manifest matches the GitHub release asset digests:

3faef282b72c97c6659e67647ba5cb154631ff52cafe38f9777bd658f2089707  token_terminator-0.11.1-py3-none-any.whl
88fe2bda4b1733b84a3cadbd8eb6cc37edde089a6d1d6c37d5e199c68775ff45  token_terminator-0.11.1.tar.gz

This publishes PR #20's adapter, not the separate deployed-adapter branch. No live machine installation, restart, configuration or evidence-store change was performed during this release. Unknown-model tokenizer support and the upstream discovery race remain separate limitations and are explicitly identified in the release notes. Existing v0.11.0 and older tags/releases are untouched. The results above are final pre-merge checks and publication checks, not a claim that a live session has reduced or that the later post-merge test run has finished.