AgentEngine deployment changes - #1162

)

(cherry picked from commit 2bcf2c7)

…919, #786, #1212, #1162) (#1282)

* feat: enable h2c on web launcher listener (#826)

* feat(web): enable h2c on launcher listener

- Configure the web launcher HTTP server with net/http.Protocols so the same listener accepts both HTTP/1 and cleartext HTTP/2 traffic.
- This keeps existing REST, Web UI, A2A, and trigger routes working over HTTP/1 while allowing custom web sublaunchers to register handlers that require h2c, including native gRPC servers. The change avoids any public launcher API changes  and uses the standard library instead of adding an http2/h2c dependency.
- Add a regression test covering existing HTTP/1 route handling through the base router.

* feat(launcher): add opt-in h2c support for web server

* test(launcher): cover opt-in h2c behavior

Test HTTP/1.1 and prior-knowledge h2c requests with the flag enabled and disabled. Clarify h2c security, upgrade, and streaming timeout behavior in the flag documentation.

* fix(launcher): handle response body close errors in h2c tests

(cherry picked from commit e6562fe)

* fix(skilltoolset): document load_skill param as name, not skill_name (#919)

* fix(skilltoolset): document load_skill param as name, not skill_name

The default system instruction told the model to call the load_skill tool
with skill_name="<SKILL_NAME>", but LoadSkillArgs marshals its field as
"name". The first load_skill call in each session was therefore rejected for
an unknown field and only recovered via retry. Use name= to match the schema.

* fix(skilltoolset): make the new instruction test compile and pass lint

The test file added here imported the pre-v2 module path, which no longer
exists, so the whole skilltoolset package failed to build:

    tool/skilltoolset/instruction_test.go:8:2: no required module provides
    package google.golang.org/adk/tool/skilltoolset/internal/skilltool

Point the import at google.golang.org/adk/... and add the Apache 2.0
header that goheader requires on every new Go file.

No change to the test's logic or to the instruction fix.

---------

Co-authored-by: João Westerberg <westerberg@google.com>

(cherry picked from commit 9e6a3db)

* docs(llmagent): correct OutputSchema comment — tools still work via set_model_response (#786)

The previous comment incorrectly stated that OutputSchema prevents the use of
tools. The framework injects a set_model_response tool transparently so all
tool types (function tools, RAG, agent transfer) continue to work alongside
structured output. Updated the comment to reflect actual behavior.

Fixes #756

Co-authored-by: João Westerberg <westerberg@google.com>

(cherry picked from commit d1fd187)

* docs: list the Kotlin and TypeScript ADK implementations in README (#1212)

The Important Links header offers Python ADK and Java ADK but predates the
Kotlin and TypeScript implementations, so readers looking for a sibling
language find an incomplete list. Add both, and drop a stray trailing space
on the Java line.

(cherry picked from commit 5344562)

* deploy(agentengine): drop the hardcoded GOOGLE_API_KEY secret env (#1162)

(cherry picked from commit 2bcf2c7)

---------

Co-authored-by: Jan Krynauw <jan@alisx.com>
Co-authored-by: Parafee41 <parafee_2041@qq.com>
Co-authored-by: nuthalapativarun <nuthalapativarun@gmail.com>
Co-authored-by: Karol Piotrowicz <karol.piotrowicz@gmail.com>

…removed dependency on GOOGLE_API_KEY secret (google#1162)

Switches the sample to gemini-3.5-flash. gemini-flash-latest is an AI
Studio alias and returns 404 NOT_FOUND on Vertex, which the README
offers as Option B, so a reader following that path hit an error. The
15 samples still on the alias were all written in one batch on the same
day; every example touched since #1162 uses gemini-3.5-flash, and that
commit made the switch specifically to drop the API-key dependency.

Replaces the flowchart with a sequence diagram, as suggested in review.
The sample exists to show WHICH participant emits the event carrying the
output, and a flowchart cannot express authorship — the sequence diagram
shows the reply leaving drafter rather than assistant. It also carries
the branch natively through alt/else, so the step numbering and the
solid-versus-dotted legend the flowchart needed are both gone.
Precedent: examples/agentregistry/{bind,a2a}.

The example session was re-captured from a live run on the new model
rather than edited by hand.

* docs(examples/workflow): add a WithUseAsOutput sample

The workflow examples cover dynamic orchestration and AgentNode
composition, but nothing shows output delegation: handing a dynamic
node's terminal output to a child so the child's event carries it and
the parent emits none.

The orchestrator branches in plain Go before any model call, so both
paths are reachable from the console and the documented session
reproduces. The in-scope branch delegates to an LlmAgent child with
WithUseAsOutput, the out-of-scope branch returns its own string.

* docs(examples/workflow): correct the use_as_output README

Review of the sample turned up four claims that do not match what the
code does.

The "needs LLM on the delegating branch only" bullet was wrong in the
way that costs a newcomer the most time: the model client is built at
startup, so without credentials the sample dies with "api key is
required" before printing a prompt, including on the branch that never
calls the model.

"Reaches the caller exactly once" was true of events but not of the
console, which prints the reply once either way, so a reader deleting
WithUseAsOutput to see the promised duplicate would have concluded the
option does nothing.

Nothing said why the orchestrator has no successor. A node chained
after a delegating dynamic node receives the zero value rather than the
delegated text, silently, so the omission was the one thing a reader
most needed explained.

The unknown-agent warning fires on turns after a delegated one, when
the runner walks the history, not on the delegated turn itself.

Also numbers the diagram edge that was skipped on the delegating path,
reconciles the legend with the arrows actually drawn, and switches the
transcript to US spelling. The transcript was re-captured from a live
run rather than edited by hand.

* docs(examples/workflow): address review on the use_as_output sample

Switches the sample to gemini-3.5-flash. gemini-flash-latest is an AI
Studio alias and returns 404 NOT_FOUND on Vertex, which the README
offers as Option B, so a reader following that path hit an error. The
15 samples still on the alias were all written in one batch on the same
day; every example touched since #1162 uses gemini-3.5-flash, and that
commit made the switch specifically to drop the API-key dependency.

Replaces the flowchart with a sequence diagram, as suggested in review.
The sample exists to show WHICH participant emits the event carrying the
output, and a flowchart cannot express authorship — the sequence diagram
shows the reply leaving drafter rather than assistant. It also carries
the branch natively through alt/else, so the step numbering and the
solid-versus-dotted legend the flowchart needed are both gone.
Precedent: examples/agentregistry/{bind,a2a}.

The example session was re-captured from a live run on the new model
rather than edited by hand.

…removed dependency on GOOGLE_API_KEY secret (google#1162)