release(0.5.8): finalize enforcement release

Author: coreyt

CHANGELOG.md

@@ -7,10 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## [Unreleased]
 
-### Internal 0.5.7 release train (not published)
+## [0.5.8] — 2026-08-02
 
-This odd-numbered release train is build- and CI-validated but intentionally has
-no package version bump, tag, or PyPI publication.
+Published release containing the build- and CI-validated internal 0.5.7 train,
+the GPT-5.6 enforcement work, and configuration-applicability clarity.
 
 ### Added
 
@@ -21,6 +21,9 @@ no package version bump, tag, or PyPI publication.
 - `X-Airlock-Model-Alias` makes configured generic aliases transparent: clients see
   the actual served LiteLLM body and a directly callable newer/current-generation
   alias without any routing change.
+- The TUI Config screen now marks controls that require a proxy restart and identifies
+  controls that its Apply action does not change. Apply reports whether each actual
+  change is effective immediately or pending a restart, and reports no-op Applies.
 
 ### Fixed
 

airlock/__init__.py

@@ -1,3 +1,3 @@
 """Airlock — Enterprise LLM proxy built on LiteLLM."""
 
-__version__ = "0.5.6"
+__version__ = "0.5.8"

airlock/callbacks/tracing.py

@@ -45,7 +45,7 @@ def _get_tracer() -> Any:
         provider = TracerProvider(resource=resource)
         trace.set_tracer_provider(provider)
 
-    return trace.get_tracer("airlock", "0.5.6")
+    return trace.get_tracer("airlock", "0.5.8")
 
 
 _tracer = _get_tracer()

airlock/fast/guardian.py

@@ -39,7 +39,10 @@
     AirlockProviderBlocked,
     sanitize_reason,
 )
-from airlock.reasoning_effort import normalize_reasoning_effort, validate_reasoning_effort
+from airlock.reasoning_effort import (
+    normalize_reasoning_effort,
+    validate_reasoning_effort,
+)
 from airlock.transparency import detect_dropped_params, record_mutation
 from airlock.client_identity import (
     client_id_from_api_key,
@@ -312,7 +315,9 @@ async def async_pre_call_hook(
                 )
                 suggestions = alias_table.suggest(model_name)
                 suggestions.sort(
-                    key=lambda item: 0 if item.get("model") == cross_tier.suggested else 1
+                    key=lambda item: (
+                        0 if item.get("model") == cross_tier.suggested else 1
+                    )
                 )
                 if not suggestions:
                     suggestions = [

airlock/proxy_errors.py

@@ -89,9 +89,7 @@ def __init__(
 class AirlockInvalidReasoningEffort(Exception):
     """A known-invalid OpenAI reasoning effort rejected before provider dispatch."""
 
-    def __init__(
-        self, requested: str, model: str, supported: frozenset[str]
-    ) -> None:
+    def __init__(self, requested: str, model: str, supported: frozenset[str]) -> None:
         self.requested = requested
         self.model = model
         self.supported = tuple(sorted(supported))
@@ -262,7 +260,9 @@ def install_airlock_error_handlers_on_proxy_app() -> bool:
     if not getattr(app.state, "airlock_model_not_found_handler_installed", False):
         app.add_exception_handler(AirlockModelNotFound, airlock_model_not_found_handler)
         app.state.airlock_model_not_found_handler_installed = True
-    if not getattr(app.state, "airlock_invalid_reasoning_effort_handler_installed", False):
+    if not getattr(
+        app.state, "airlock_invalid_reasoning_effort_handler_installed", False
+    ):
         app.add_exception_handler(
             AirlockInvalidReasoningEffort, airlock_invalid_reasoning_effort_handler
         )

airlock/reasoning_effort.py

@@ -201,7 +201,9 @@ def _supported_efforts(model: str | None) -> frozenset[str] | None:
     return frozenset(levels)
 
 
-def _validation_status(model: str | None, requested: str) -> tuple[bool, frozenset[str]] | None:
+def _validation_status(
+    model: str | None, requested: str
+) -> tuple[bool, frozenset[str]] | None:
     """Return ``(known_invalid, supported)`` or ``None`` when validation is unsafe.
 
     GPT-5.6's ``max`` support is documented by OpenAI even though the pinned LiteLLM

dev/plans/0.5.8-plan.md

@@ -754,6 +754,6 @@ The open GitHub backlog is intentionally **not** an implicit 0.5.8 queue:
 - [x] P-2 and P-2b enforcement reviewed and implemented; post-release reporting stays
       enabled.
 - [x] P-6a and P-6b resolved (implemented or explicitly recorded as no-change).
-- [ ] P-8 implemented and reviewed, including its focused TUI tests.
+- [x] P-8 implemented and reviewed, including its focused TUI tests.
 - [ ] P-9 complete: final validation recorded, #22 audited/closed-or-deferred, version
       and release notes prepared, and the normal 0.5.8 release tag created.

dev/plans/runs/0.5.8-release-evidence-2026-08-02.md

@@ -0,0 +1,62 @@
+# 0.5.8 release evidence — 2026-08-02
+
+## Candidate
+
+- Package version: `0.5.8` in `pyproject.toml`, `airlock.__version__`, and the tracing
+  version. `scripts/check-version-consistency.py --tag v0.5.8` passed.
+- Dependency authority: `uv.lock` refreshed after the manifest version change;
+  `uv lock --check` passed. Docker resolves the locked LiteLLM `1.94.1` and validates
+  it against Airlock's declared `>=1.94.1,<2` floor at image build time.
+
+## P-8 / #34 validation
+
+`uv run pytest tests/test_tui.py -q -k config` passed: **12 passed**.
+
+The ConfigPane has one field-applicability contract. It marks restart-required and
+unapplied controls, reports no-op/live/restart/mixed Applies without values, and leaves
+`config.local.yaml` loading semantics unchanged.
+
+## GitHub issue #22 audit — PASS
+
+| Required decision | Evidence |
+|---|---|
+| Keep `/v1/chat/completions` OpenAI-compatible | `airlock/docs.py` conceptual/API documentation and `tests/test_docs.py` |
+| Accept Gemini request controls without a new body response contract | `airlock/gemini_interface.py` and `tests/test_gemini_interface.py` |
+| Surface Gemini mode/state through response headers | `airlock/callbacks/model_override_headers.py` and `tests/test_model_override_headers.py` |
+| Preserve structured provider detail for operators | `airlock/callbacks/request_event.py`, enterprise JSONL projection, `tests/test_enterprise_logger.py`, and the TUI Guards projection |
+
+Focused audit command:
+
+```text
+uv run pytest tests/test_gemini_interface.py tests/test_model_override_headers.py \
+  tests/test_enterprise_logger.py tests/test_tui.py -q -k 'gemini or config'
+# 24 passed, 143 deselected
+```
+
+The issue is documentation/decision scope, not a request for a new provider-native
+response body. It is eligible to close with this evidence.
+
+## Local release gate
+
+| Gate | Result |
+|---|---|
+| Full non-live suite | `2768 passed, 107 deselected, 1 xpassed` (98 deprecation warnings) |
+| Ruff check / format | clean: `airlock/` and `tests/` |
+| Mypy | `airlock/fast/` clean (`--ignore-missing-imports`) |
+| Dependency audit | no known vulnerabilities; local project and spaCy model are intentionally not PyPI-auditable |
+| Docs | `mkdocs build --strict` passed |
+| Build | `uv build` produced `airlock_llm-0.5.8.tar.gz` and wheel |
+| Clean wheel install | fresh temporary venv imported `airlock 0.5.8`; `airlock --help` passed |
+| Docker | image built with locked LiteLLM `1.94.1`; isolated container returned `"I'm alive!"` from `/health/liveliness` with `AIRLOCK_HOST=0.0.0.0` and optional services disabled; container stopped |
+
+Live-provider tests remain opt-in and were intentionally deselected. The direct funded
+OpenAI `reasoning_effort=max` probe remains unavailable because no funded key is
+present; documented GPT-5.6 support is covered by the model-specific contract test.
+
+## Remaining release actions
+
+1. Commit this release candidate.
+2. Close GitHub issue #22 with this audit evidence.
+3. Push the candidate and require the GitHub Actions CI workflow to pass.
+4. Create and push `v0.5.8`; the release workflow then performs PyPI trusted
+   publishing and creates the GitHub release.

dev/plans/runs/STATUS-0.5.8.md

@@ -0,0 +1,32 @@
+# STATUS — 0.5.8  (live state board)
+
+> The current state is derived from release evidence and commits; this board is a
+> compaction-safe pointer, not a substitute for them.
+
+_Last updated: 2026-08-02 · release candidate: `v0.5.8` (local candidate)_
+
+## Current state
+
+- **In flight:** P-9 release hygiene.
+- **Next action:** commit the release candidate, close the audited documentation issue,
+  push for GitHub Actions, then create the `v0.5.8` publishing tag only after CI is
+  green.
+
+## Scope scoreboard
+
+| Item | State | Witness |
+|---|---|---|
+| P-2 reasoning-effort enforcement | CLOSED | `5a7c714`, `5b20c4e`; tests in `tests/test_reasoning_effort.py` |
+| P-2b cross-tier fuzzy refusal | CLOSED | `5a7c714`; tests in `tests/test_model_suggestion.py` |
+| P-6 alias disclosure / P-6a / P-6b | CLOSED | `1062343`; header and interface tests |
+| P-7 documentation | CLOSED | routing/configuration docs and 0.5.8 plan |
+| P-8 config applicability (#34) | CLOSED | `5b86675`; 12 ConfigPane tests green |
+| P-9 release hygiene / #22 audit | IN PROGRESS | `0.5.8-release-evidence-2026-08-02.md` |
+
+## Constraints
+
+- `config.local.yaml` is a deliberate, default-empty extension mechanism and any
+  machine-local override remains unstaged.
+- Liveness validation uses `/health/liveliness` only—never `/health`.
+- The release tag triggers PyPI trusted publishing; it follows, rather than precedes,
+  a green GitHub Actions run.

dev/user-needs.md

@@ -968,9 +968,9 @@ multiples on every request until someone notices a bill.
    token. *(0.5.6)*
 4. The refusal response tells the caller what to send instead: a 404 whose
    `error.message` is self-sufficient prose, plus ranked suggestions with cost tier.
-   *(0.5.7 F-3 — not yet shipped)*
+   *(0.5.8)*
 5. Suggestions never include a model outside the caller's permitted catalog.
-   *(0.5.7 F-3)*
+   *(0.5.8)*
 
 ---
 
@@ -1016,3 +1016,24 @@ only one of them is honest.
    these; Airlock inherits them, confirmed by inspection and local calculation)*
 5. Per-request-priced providers either record their cost or are explicitly documented
    as unrecorded with a reason. *(0.5.7 F-4 step 3)*
+
+---
+
+## UN-31: Truthful Runtime Configuration Applicability
+
+**As an** Airlock operator,
+**I need** the TUI to tell me which configuration edits take effect immediately and
+which need a proxy restart,
+**so that** I do not either interrupt traffic unnecessarily or assume an inactive
+setting is protecting a running proxy.
+
+### Acceptance Criteria
+
+1. Every ConfigPane control has an explicit applicability classification; controls not
+   changed by Apply are identified as such rather than presented as live.
+2. Restart-required controls are visibly marked, including provider credentials,
+   listener address, failover map, and logging destinations.
+3. Apply reports no change, live-only changes, restart-required changes, and mixed
+   changes accurately without exposing a secret or configuration value in its status.
+4. Existing live controls—enforcement mode, PII/keyword controls, and MCP tool
+   allow/block lists—continue to update the runtime environment immediately.