chore(release): activeagent and actionagent 1.6.2

[TonsOfFun]

Preview of the 1.6.2 cut: release/1.6.0 (already carrying 1.6.0 and 1.6.1, which main does not yet have) plus every open PR that merges cleanly and passes, with the version bump and changelog on top. Merging this also brings main up to the release line, as #442 did for 1.5.2.

What is in it

PR	Change
#453	Agent releases — digest, revision, pinning on traces/runs/evaluation runs, add_agent_releases migration, rake tasks
#452	Tools tab reads agent.tools for schema tools, and can change it
#454	Evaluations run on MySQL (nil config/criteria), judge class memoized once, sign_in_path/sign_out_path documented
#456	Dashboard API hardening — container-valued query params coerced instead of 500ing, compare validates providers, a recording's show scrubs cookies and web storage from the handoff state
#398	405 for unsupported MCP transport methods
#358, #357, #348	Actions bumps: action-gh-release v3, upload-artifact v7, setup-node v7

Plus its own commits: the merged PRs' changelog entries moved out of ## [1.6.0] (where a merge onto the release branch places them by context) into the new ## [1.6.2] section, the version bump, a telemetry test that restores the global configuration after swapping the tracer, the ## [1.6.2] entries for #456, and a test-helper fix for an order-dependent failure the suite already had (below).

Not in it, and why

• build(deps): update json requirement from < 3 to < 4 #423 (json < 4) — edits only the root Gemfile; the ten gemfiles/*.gemfile CI uses keep < 3, so it changes nothing CI runs and leaves its own "json 3 breaks db:migrate" comment in place. Suggest closing.
• chore(deps-dev): update minitest requirement from ~> 5.0 to ~> 6.0 #349 (minitest 6) — gemfiles/*.gemfile still pin ~> 5.0; Bundler refuses the conflicting requirement.
• feat: add Atlas Cloud provider #352, feat: Add allowed_tools and require_approval to option transforms #328, Add llms.txt for machine-readable docs index #313 — feature / red CI since July / conflicting.

Blockers before this stops being a preview

1. feat(agents): releases — a digest of what the model is given, cut on deploy, pinned to every trace, run and evaluation #453's test isolation. Resolved: feat(agents): releases — a digest of what the model is given, cut on deploy, pinned to every trace, run and evaluation #453 landed on main carrying the fix (the two real-generation tests instrument their own agent class and clear what they stored), main is merged in here, and CI on the branch head is green.
2. Version. By the reasoning in the 1.6.1 release note ("no new public surface … so a patch under semver"), feat(agents): releases — a digest of what the model is given, cut on deploy, pinned to every trace, run and evaluation #453 adds public surface — ActiveAgent::Release, a migration, rake tasks — which reads as 1.7.0 rather than 1.6.2. Named 1.6.2 here as requested; renaming is a one-commit change.
3. chore(deps): bump softprops/action-gh-release from 2 to 3 #358 and chore(deps): bump actions/upload-artifact from 4 to 7 #357 change release.yml, which the gem publish depends on; worth watching the first tag that runs through them.
4. Changelog date is today's; set at cut time.

Test isolation fixed on this branch

The rails-main job on 8ebff51 failed with 28 errors, every one Gem::LoadError: Error loading the 'redis' Action Cable pubsub adapter out of AgentRun#broadcast_update. Action Cable reads config/cable.yml for the current environment the first time its server class loads, and the dummy app does not eager load in test, so that first time was whichever test first broadcast; when the seed put engine_integration_test's staging case first, its stubbed Rails.env had no section in cable.yml, the config came back empty and the adapter fell back to redis for the rest of the process. Reproduced on Rails 8 with a forced order; test/test_helper.rb now loads the config under the real test environment before any test runs (f995cc9). Independent of #456; any seed could hit it on any Rails.

Worth watching

The rails-main job on f995cc9 errored once, on Integration::OpenAI::Responses::CommonFormat::ToolsTest "tool_choice_auto Request Building": NoMethodError: undefined method 'each_key' for nil inside WebMock's global-stub bookkeeping (DynamicResponse#evaluate), during the second, WebMock-stubbed replay of the cassette. Green on re-run; #456's own rails-main run saw the Anthropic sibling of the same test fail as Connection error. The 343 integration replay tests pass three times in a row locally on Rails 8, and no file either failure touches changes on this branch, so it is filed here rather than fixed.

Verified

• Every PR merged onto release/1.6.0 without conflict; Coerce container-valued query params and scrub handoff state on read #456 merged onto this branch without conflict.
• CI on f995cc9: all eight checks green (four test matrices, lint, API gems, both solid_agent integrations).
• bin/test (Ruby 3.4.9, gemfiles/rails8.gemfile): 2014 runs, 0 failures on the branch head before Coerce container-valued query params and scrub handoff state on read #456; after it and the helper fix, 2016 runs, 0 failures (local errors only from provider credentials the container lacks).
• actionagent/frontend: 25 tests, 25 pass.
• bin/lint: no offenses (549 files); rubocop clean on the files Coerce container-valued query params and scrub handoff state on read #456 and the helper fix touch.

🤖 Generated with Claude Code

[TonsOfFun]

CI reading for this preview: lint, API gems and both solid_agent integrations green; 3.4/rails8 green; 3.3/rails8, 3.2/rails7 and 3.4/railsmain red.

Everything red traces to #453's test isolation (details and the rails7 database is locked evidence on #453 (comment) and the follow-up there), except one error each on 3.3/rails8 and railsmain: an Anthropic Connection error in a live-network example test, unrelated to this branch. Once #453 is updated its merge here gets redone and CI re-run.

[TonsOfFun]

A second, older source of the dashboard order-sensitivity is fixed on this branch in 49a57cbf — the one the 1.6.1 release note called "one ordering-sensitive dashboard test that fails identically with and without these changes".

telemetry_correlation_test's swap_global_tracer set the global configuration's enabled and local_storage, and both callers' ensure called the same helper to "restore" — so after either test, telemetry stayed on with local storage for the rest of the process. With GenerationInstrumentation on ActiveAgent::Base (installed, legitimately, by InstrumentationInstallTest), every later VCR generation was traced, delivered by the batching reporter's thread on its own connection, and committed — surviving the test transaction and showing up as observed agents in whichever dashboard test ran next (the Providers::OllamaProviderTest::EfficientAgent.ask and …ArrayInstructionsAgent.basic_request agents in this PR's last CI diff).

Measured with a probe on every ingest and on row counts after each test, same seed, before and after: before — background-thread ingests during dozens of unrelated tests, rows surviving rollback; after, on a clean database — 0 failures, 0 leaked rows, 0 background ingests, 0 rows left on disk. One thing to know for local runs: the leaked rows had been committed to test/dummy/storage/test.sqlite3, so a suite run after a leaky one starts dirty until the tables are cleared.