feat(adk): Go runtime conversation-history compaction - #2025

## Problem

Since 0.10.0 the Agent CRD defaults `spec.declarative.runtime` to `go`,
and the Go runtime never read `spec.declarative.context.compaction`. The
controller compiles the field into the agent's `config.json`
(`context_config.compaction`) and the Python runtime consumes it, but a
Go agent grew its prompt with every turn regardless of what the operator
configured; the only signal was the `UnsupportedFeatures` condition
(kagent-dev#2728). Measured on an internal installation running 0.10.1: an agent
session grew from 4k to 190k input tokens per model call over 119 calls,
and every Go agent there ran without compaction because the runtime
ignored the field.

`google.golang.org/adk/v2` v2.3.0 (2026-08-31) ships native compaction
(`session/compaction`: sliding window, tail retention, LLM summarizer)
and the `runner.Config.Compaction` hook, so the runtime can consume it
instead of carrying its own port of the algorithm, which is what kagent-dev#1971
and kagent-dev#2025 attempted before the primitives existed.

## Change

- **Dependency:** `google.golang.org/adk/v2` v2.2.0 → v2.3.0. The
OpenTelemetry exporters and contrib instrumentation move to the matching
v1.45.0 / v0.21.0 / v0.70.0 set, which the new ADK version requires for
`otel/log`. otel/log v0.21.0 hands log bodies as attribute.Value; the
telemetry test on main
follows that in the same commit. No other kagent code needed changes for
the bump.
- **Runtime:** `agent.CompactionConfig` translates
`context_config.compaction` into a `compaction.Config` and
`runner.CreateRunnerConfig` hands it to the ADK runner.
`compaction_interval` / `overlap_size` drive the sliding window,
`token_threshold` / `event_retention_size` tail retention; either
strategy is off when its settings are unset, so an agent without
`context_config` runs exactly as before.
- **Half-configured settings** are ignored with a warning rather than
rejected: `token_threshold` without `event_retention_size` (and the
reverse), `overlap_size` without an interval. That mirrors the Python
runtime, where adk-python runs tail retention only when both settings
are present. What remains is validated at startup with the runner's own
rules, so a configuration the runner would refuse (negative values, no
strategy at all) fails fast with the reason instead of on the first
turn.
- **Summarizer:** `summarizer_model` builds a `compaction.LLMSummarizer`
through the same model factory as the agent (`CreateLLM`), with the
model's generation config and a 60 s bound per call (the same bound the
runner applies to its default summarizer). A `prompt_template` gets the
tool-name instruction the Python runtime appends (kagent-dev#1907), so summaries
keep namespaced tool names; a template without a summarizer model runs
on the agent's own model. With neither, the runner summarizes with the
agent's model and ADK's default prompt, matching the Python runtime.
- **Observability:** startup logs `context compaction enabled` with the
resolved values and the summarizer model, or `context compaction not
configured`, so an operator can see from the agent log what the runtime
does with the field.
- **API:** `main` had no field for compaction in the v1alpha3 API (the
0.10 line has `spec.declarative.context`), so nothing produced
`context_config`. `spec.kagent.compaction` on the `Harness` adds it,
next to `spec.kagent.memory`: `compactionInterval` / `overlapSize` for
the sliding window, `tokenThreshold` / `eventRetentionSize` for tail
retention, and an optional `summarizer` with a `modelConfigRef` (a
`ModelConfig` in the Harness namespace) and a `promptTemplate`. It is
runtime policy rather than portable agent behaviour (specific to the ADK
runner; the Claude and Codex harnesses have nothing to apply it to),
which is why it lives on the Harness and not on the `AgentTemplate`. The
CRD rejects at admission what the runtime would refuse at startup: no
strategy, `tokenThreshold` without `eventRetentionSize` (and the
reverse), `overlapSize` without `compactionInterval`, a `promptTemplate`
without `{conversation_history}`.
- **Controller:** the kagent compiler translates the Harness's
compaction policy into `context_config` in `config.json` of every root
agent the Harness runs. A summarizer `ModelConfig` other than the
agent's own is resolved from the Harness namespace like the memory
model, so its credentials and egress join the revision and a change to
it compiles a new revision; the agent's own model stays implicit because
the runtime already summarizes with it by default.

Compaction records travel on `EventActions.Compaction`; the local
DurableDir session store persists `Actions` as a whole (ADK's
`session/database`), so a summary survives an actor restart and the same
range is not summarized again.

The 0.10 line, where the field already exists on the Agent CRD and the
reconciler flagged it as unsupported, is kagent-dev#2791.

## Testing

- `go test ./adk/pkg/agent/`: `TestCompactionStrategies` (table: each
strategy alone and combined, unset settings, the three half-configured
cases, no strategy, negative values), `TestCompactionConfig` (not
configured → nil; default summarizer; dedicated summarizer model; custom
prompt on the agent model; prompt without `{conversation_history}`
rejected; prompt without any model rejected; the `context_config` block
as the controller compiles it),
`TestCompactionSlidingWindowFiresAfterInterval` (a scripted model drives
the ADK runner: no record after turn 1, one record after turn 2 whose
content is the summarizer's reply, and turn 3's request shows the
summary and none of the compacted turns),
`TestCompactionTailRetentionFiresOverTokenThreshold`.
- `go test ./core/internal/translator/...`: `TestApplyCompaction` (not
configured → no `context_config`; all four settings and a prompt on the
agent model → mapped with no `summarizer_model`; a dedicated summarizer
`ModelConfig` → its model in `config.json`, its credential in the
environment, its host in the egress, itself in the provenance models; a
missing summarizer `ModelConfig` → error).
- `go test ./api/v1alpha3/` (envtest): the five new CEL cases reject
each half configuration with its message, and the fully configured
Harness is accepted.
- `make -C go test`: all packages pass.
- `make -C go lint`: 0 issues.

## End-to-end

`TestAgentInstanceContextCompaction` (in the `test-e2e` job) runs the
whole chain with `mockllm` as the summarizer: a clone of the suite's
`kagent` Harness with `compactionInterval: 2` and a dedicated summarizer
`ModelConfig`, admitting a plain `AgentTemplate`; both models point at
one mock LLM behind a recording proxy on the test host; a default header
on each `ModelConfig` tells the two models apart. Three turns through
the public A2A API, then the recorded requests show what the runtime
sent: the summarizer model was called exactly once, after the second
turn, with both turns in the transcript and the custom prompt; the
second turn's request still carried the first turn raw; the third turn's
request carries the summary and none of the four compacted messages. Run
before pushing against a local kind cluster with Substrate 0.0.26 and
the controller and Go ADK images built from this branch: passes in 4.9
s.

The real-model run (kind, images built from the branch, Anthropic
`claude-haiku-4-5`, a five-turn tool-using conversation: turn 5 needs
2,524 prompt tokens with compaction against 50,247 without) was done on
the 0.10 backport kagent-dev#2791. The runtime code is the same on both branches
apart from the logging package (`log/slog` here, logr on 0.10).

Refs kagent-dev#2728 (closed by kagent-dev#2791 for the 0.10 line; on `main` the field
returns with the v1alpha3 API).

---------

Signed-off-by: Timo Derstappen <[email protected]>

## Problem

Since 0.10.0 the Agent CRD defaults `spec.declarative.runtime` to `go`,
and the Go runtime never read `spec.declarative.context.compaction`. The
controller compiles the field into the agent's `config.json`
(`context_config.compaction`) and the Python runtime consumes it, but a
Go agent grew its prompt with every turn regardless of what the operator
configured; the only signal was the `UnsupportedFeatures` condition
(kagent-dev#2728). Measured on an internal installation running 0.10.1: an agent
session grew from 4k to 190k input tokens per model call over 119 calls,
and every Go agent there ran without compaction because the runtime
ignored the field.

`google.golang.org/adk/v2` v2.3.0 (2026-08-31) ships native compaction
(`session/compaction`: sliding window, tail retention, LLM summarizer)
and the `runner.Config.Compaction` hook, so the runtime can consume it
instead of carrying its own port of the algorithm, which is what kagent-dev#1971
and kagent-dev#2025 attempted before the primitives existed.

## Change

- **Dependency:** `google.golang.org/adk/v2` v2.2.0 → v2.3.0. The
OpenTelemetry exporters and contrib instrumentation move to the matching
v1.45.0 / v0.21.0 / v0.70.0 set, which the new ADK version requires for
`otel/log`. otel/log v0.21.0 hands log bodies as attribute.Value; the
telemetry test on main
follows that in the same commit. No other kagent code needed changes for
the bump.
- **Runtime:** `agent.CompactionConfig` translates
`context_config.compaction` into a `compaction.Config` and
`runner.CreateRunnerConfig` hands it to the ADK runner.
`compaction_interval` / `overlap_size` drive the sliding window,
`token_threshold` / `event_retention_size` tail retention; either
strategy is off when its settings are unset, so an agent without
`context_config` runs exactly as before.
- **Half-configured settings** are ignored with a warning rather than
rejected: `token_threshold` without `event_retention_size` (and the
reverse), `overlap_size` without an interval. That mirrors the Python
runtime, where adk-python runs tail retention only when both settings
are present. What remains is validated at startup with the runner's own
rules, so a configuration the runner would refuse (negative values, no
strategy at all) fails fast with the reason instead of on the first
turn.
- **Summarizer:** `summarizer_model` builds a `compaction.LLMSummarizer`
through the same model factory as the agent (`CreateLLM`), with the
model's generation config and a 60 s bound per call (the same bound the
runner applies to its default summarizer). A `prompt_template` gets the
tool-name instruction the Python runtime appends (kagent-dev#1907), so summaries
keep namespaced tool names; a template without a summarizer model runs
on the agent's own model. With neither, the runner summarizes with the
agent's model and ADK's default prompt, matching the Python runtime.
- **Observability:** startup logs `context compaction enabled` with the
resolved values and the summarizer model, or `context compaction not
configured`, so an operator can see from the agent log what the runtime
does with the field.
- **API:** `main` had no field for compaction in the v1alpha3 API (the
0.10 line has `spec.declarative.context`), so nothing produced
`context_config`. `spec.kagent.compaction` on the `Harness` adds it,
next to `spec.kagent.memory`: `compactionInterval` / `overlapSize` for
the sliding window, `tokenThreshold` / `eventRetentionSize` for tail
retention, and an optional `summarizer` with a `modelConfigRef` (a
`ModelConfig` in the Harness namespace) and a `promptTemplate`. It is
runtime policy rather than portable agent behaviour (specific to the ADK
runner; the Claude and Codex harnesses have nothing to apply it to),
which is why it lives on the Harness and not on the `AgentTemplate`. The
CRD rejects at admission what the runtime would refuse at startup: no
strategy, `tokenThreshold` without `eventRetentionSize` (and the
reverse), `overlapSize` without `compactionInterval`, a `promptTemplate`
without `{conversation_history}`.
- **Controller:** the kagent compiler translates the Harness's
compaction policy into `context_config` in `config.json` of every root
agent the Harness runs. A summarizer `ModelConfig` other than the
agent's own is resolved from the Harness namespace like the memory
model, so its credentials and egress join the revision and a change to
it compiles a new revision; the agent's own model stays implicit because
the runtime already summarizes with it by default.

Compaction records travel on `EventActions.Compaction`; the local
DurableDir session store persists `Actions` as a whole (ADK's
`session/database`), so a summary survives an actor restart and the same
range is not summarized again.

The 0.10 line, where the field already exists on the Agent CRD and the
reconciler flagged it as unsupported, is kagent-dev#2791.

## Testing

- `go test ./adk/pkg/agent/`: `TestCompactionStrategies` (table: each
strategy alone and combined, unset settings, the three half-configured
cases, no strategy, negative values), `TestCompactionConfig` (not
configured → nil; default summarizer; dedicated summarizer model; custom
prompt on the agent model; prompt without `{conversation_history}`
rejected; prompt without any model rejected; the `context_config` block
as the controller compiles it),
`TestCompactionSlidingWindowFiresAfterInterval` (a scripted model drives
the ADK runner: no record after turn 1, one record after turn 2 whose
content is the summarizer's reply, and turn 3's request shows the
summary and none of the compacted turns),
`TestCompactionTailRetentionFiresOverTokenThreshold`.
- `go test ./core/internal/translator/...`: `TestApplyCompaction` (not
configured → no `context_config`; all four settings and a prompt on the
agent model → mapped with no `summarizer_model`; a dedicated summarizer
`ModelConfig` → its model in `config.json`, its credential in the
environment, its host in the egress, itself in the provenance models; a
missing summarizer `ModelConfig` → error).
- `go test ./api/v1alpha3/` (envtest): the five new CEL cases reject
each half configuration with its message, and the fully configured
Harness is accepted.
- `make -C go test`: all packages pass.
- `make -C go lint`: 0 issues.

## End-to-end

`TestAgentInstanceContextCompaction` (in the `test-e2e` job) runs the
whole chain with `mockllm` as the summarizer: a clone of the suite's
`kagent` Harness with `compactionInterval: 2` and a dedicated summarizer
`ModelConfig`, admitting a plain `AgentTemplate`; both models point at
one mock LLM behind a recording proxy on the test host; a default header
on each `ModelConfig` tells the two models apart. Three turns through
the public A2A API, then the recorded requests show what the runtime
sent: the summarizer model was called exactly once, after the second
turn, with both turns in the transcript and the custom prompt; the
second turn's request still carried the first turn raw; the third turn's
request carries the summary and none of the four compacted messages. Run
before pushing against a local kind cluster with Substrate 0.0.26 and
the controller and Go ADK images built from this branch: passes in 4.9
s.

The real-model run (kind, images built from the branch, Anthropic
`claude-haiku-4-5`, a five-turn tool-using conversation: turn 5 needs
2,524 prompt tokens with compaction against 50,247 without) was done on
the 0.10 backport kagent-dev#2791. The runtime code is the same on both branches
apart from the logging package (`log/slog` here, logr on 0.10).

Refs kagent-dev#2728 (closed by kagent-dev#2791 for the 0.10 line; on `main` the field
returns with the v1alpha3 API).

---------

Signed-off-by: Timo Derstappen <[email protected]>