fix(openclaw): bound managed inference compaction

[HOYALIM]

Summary

Mitigates #4781.

This bounds each managed remote-inference compaction attempt on inference.local routes to 120 seconds, enables OpenClaw lifecycle notifications, uses safeguard compaction with an aggressive history budget, preserves one recent turn verbatim, disables quality-regeneration retries, and requests transcript rotation after successful compaction.

The existing Local Ollama small-context policy from #5468 is preserved.
ollama-local continues to use the lowered reserve-token policy and is excluded from this managed remote-inference safeguard.

This configuration prevents an unbounded seven-minute wait and improves visibility, but it does not by itself prove successful compaction, transcript rotation, or reduced context tokens.
Issue #4781 should remain open until an exact-head live run demonstrates those outcomes.

The original implementation is by @HOYALIM.
The maintainer follow-up retains that credit through Co-authored-by: Ho Lim <subhoya@gmail.com>.

Validation

• npx vitest run --project integration test/ollama-local-openclaw-config-propagation.test.ts test/generate-openclaw-config.test.ts (137 tests passed)
• npx @biomejs/biome check scripts/generate-openclaw-config.mts test/ollama-local-openclaw-config-propagation.test.ts
• npm run test-size:check
• npm run check:diff

Summary by CodeRabbit

• New Features
  • Added an OpenClaw safeguard-based context compaction policy for managed inference routes to help preserve conversation quality as context grows.
• Bug Fixes
  • Refines compaction selection so safeguards are only applied within the intended managed-inference environment, preventing unintended overrides in other setups.
• Tests
  • Updated and expanded coverage to validate managed-route safeguard behavior and to confirm safeguards are disabled when outside the expected managed-inference base URL.
• Documentation
  • Added a guide to “OpenClaw Context Compaction,” including behavior, limits, and how compaction is applied.

Signed-off-by: Ho Lim subhoya@gmail.com

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[coderabbitai]

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: f5e511e0-ab45-401f-9a89-1a1e1abd6f71

📥 Commits

Reviewing files that changed from the base of the PR and between 50731fe and 0106fa1.

📒 Files selected for processing (2)

• docs/inference/inference-options.mdx
• scripts/generate-openclaw-config.mts

✅ Files skipped from review due to trivial changes (1)

• docs/inference/inference-options.mdx

🚧 Files skipped from review as they are similar to previous changes (1)

• scripts/generate-openclaw-config.mts

---

📝 Walkthrough

Walkthrough

Adds a managed-inference safeguard compaction policy in the OpenClaw config generator, wires it into agents.defaults.compaction, and updates tests and docs for the new managed-route compaction behavior.

Changes

Managed inference safeguard compaction

Layer / File(s)	Summary
Safeguard compaction constants and builder scripts/generate-openclaw-config.mts	Adds managed-inference routing constants and a base safeguard compaction object, plus a function that enables the safeguard only for the local managed-inference route and clones the qualityGuard when building the compaction object.
Compaction assignment in buildConfig scripts/generate-openclaw-config.mts	Computes managed-inference compaction and assigns it to agentDefaults.compaction when defined.
Tests and docs for managed-route compaction test/ollama-local-openclaw-config-propagation.test.ts, docs/inference/inference-options.mdx	Updates the compaction test suite to expect the safeguard policy for managed inference, adds negative cases for non-matching inputs, and documents the OpenClaw context compaction behavior.

Estimated code review effort: 2 (Simple) | ~15 minutes

Possibly related issues

• [Ubuntu 24.04][Agent&Skills] /compact runs for minutes without completing and increases context tokens instead of decreasing #4781: Adds managed-inference compaction safeguards that match the issue’s /compact behavior.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly matches the main change: bounding managed inference compaction in OpenClaw.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

🧹 Nitpick comments (1)

test/ollama-local-openclaw-config-propagation.test.ts (1)

192-208: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Test title doesn't describe both asserted cases.

The title "does not enable managed-inference safeguards outside inference.local" only matches the first assertion (hostname mismatch). The second call (providerKey: "nvidia-prod") is actually testing a provider-key mismatch while the hostname is inference.local — it's not "outside inference.local". Consider splitting into two it() blocks with precise titles, and adding a local issue reference suffix consistent with the neighboring test.

As per coding guidelines: "Write behavior-oriented test titles, and put local issue references in a final (#1234) suffix."

♻️ Suggested split

-  it("does not enable managed-inference safeguards outside inference.local", () => {
-    expect(
-      buildManagedInferenceSafeguardCompaction(
-        "inference",
-        "nvidia-prod",
-        "https://integrate.api.nvidia.com/v1",
-      ),
-    ).toBeUndefined();
-    expect(
-      buildManagedInferenceSafeguardCompaction(
-        "nvidia-prod",
-        "nvidia-prod",
-        "https://inference.local/v1",
-      ),
-    ).toBeUndefined();
-  });
+  it("does not enable managed-inference safeguards outside inference.local (`#4781`)", () => {
+    expect(
+      buildManagedInferenceSafeguardCompaction(
+        "inference",
+        "nvidia-prod",
+        "https://integrate.api.nvidia.com/v1",
+      ),
+    ).toBeUndefined();
+  });
+
+  it("does not enable managed-inference safeguards for a non-inference provider key (`#4781`)", () => {
+    expect(
+      buildManagedInferenceSafeguardCompaction(
+        "nvidia-prod",
+        "nvidia-prod",
+        "https://inference.local/v1",
+      ),
+    ).toBeUndefined();
+  });

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@test/ollama-local-openclaw-config-propagation.test.ts` around lines 192 -
208, The test title in buildManagedInferenceSafeguardCompaction’s coverage is
too broad and only describes the hostname-mismatch case, while the second
assertion covers a provider-key mismatch with inference.local. Split the current
it() into two behavior-focused tests with precise titles that match each
assertion, and add the local issue reference as a final (`#1234`) suffix
consistent with the surrounding test names.

Source: Coding guidelines

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Nitpick comments:
In `@test/ollama-local-openclaw-config-propagation.test.ts`:
- Around line 192-208: The test title in
buildManagedInferenceSafeguardCompaction’s coverage is too broad and only
describes the hostname-mismatch case, while the second assertion covers a
provider-key mismatch with inference.local. Split the current it() into two
behavior-focused tests with precise titles that match each assertion, and add
the local issue reference as a final (`#1234`) suffix consistent with the
surrounding test names.

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Path: .coderabbit.yaml

Review profile: CHILL

Plan: Enterprise

Run ID: f544dd14-a511-45d2-82af-0590f3d8b5a2

📥 Commits

Reviewing files that changed from the base of the PR and between 05ff005 and 4e17982.

📒 Files selected for processing (2)

• scripts/generate-openclaw-config.mts
• test/ollama-local-openclaw-config-propagation.test.ts

[apurvvkumaria]

Requesting changes for acceptance coverage, not for the test-title nit. This PR currently proves only the generated JSON shape. Issue #4781 reports live manual compaction taking more than seven minutes, increasing the displayed token count, and returning no completion signal. The proposed settings do not cap that path to the requested one-to-two-minute budget because timeoutSeconds is omitted, and pinned OpenClaw v2026.5.27 defaults it to 900 seconds. They also omit notifyUser, whose pinned default is false. No behavior-level test or exact-head live evidence shows that the successor transcript rotates, the displayed token count falls, or manual compaction completes within the expected budget. Please reproduce on the pinned OpenClaw version, then add a behavior-level regression or exact-head live evidence for those acceptance criteria. Set and test timeoutSeconds and notifyUser, or document with evidence why the manual path meets the criteria without them. This PR has no release tag, so I will leave the contributor 24 hours from this review, until 2026-07-09 20:42 UTC, before any maintainer takeover.

[apurvvkumaria]

Maintainer salvage pushed as 50731fec2, retaining the original implementation with Co-authored-by: Ho Lim <subhoya@gmail.com>. The generated safeguard now caps each compaction attempt at 120 seconds and enables lifecycle notifications; comments and PR scope no longer claim guaranteed success or token reduction. Focused validation passed 137/137 tests, Biome, test-size, scoped pre-commit/pre-push hooks, and CLI typecheck. The PR remains contributor-blocked until @HOYALIM adds the required PR-body Signed-off-by: declaration matching their identity. Issue #4781 intentionally stays open pending exact-head live evidence of successful compaction, transcript rotation, and reduced context. No merge performed.

[apurvvkumaria]

The exact head is now DCO-compliant, conflict-free, and green across all required CI; CodeRabbit has no unresolved threads. The timeout, user notification, and narrowed mitigation scope address the earlier review. Two exact-head advisor runs still recommend merge_after_fixes, however, for one remaining boundary:

1. Document the source-of-truth boundary and removal condition near MANAGED_INFERENCE_SAFEGUARD_COMPACTION: the invalid behavior is in the pinned OpenClaw compaction runtime, NemoClaw is applying a generated-config mitigation, and this override should be removed when the pinned runtime fixes that behavior.
2. Add targeted runtime/schema evidence that the pinned OpenClaw build accepts the generated agents.defaults.compaction object for a managed inference.local route. The current tests prove only JSON emission. Full proof that /compact reduces tokens can remain in #4781, which should stay open, but this PR needs at least config-acceptance/runtime-override coverage so a future OpenClaw schema change cannot silently invalidate the mitigation.

The security review is otherwise clean: the exact hostname/provider guard is narrow, Local Ollama remains excluded, no secrets or authorization boundary changes are present, and negative route tests pass.

Addressed at exact head 0106fa1: source/removal boundary is documented, pinned OpenClaw 2026.6.10 runtime config validation/readback passed, all 35 checks are green, and trusted manual advisor run 29000033108 found no required blocker.

[cv]

Approved at exact head 0106fa1f43.

• All 35 GitHub checks pass; DCO is present and all five commits are Verified.
• The pinned OpenClaw 2026.6.10 production image passed the exact-head runtime-overrides live test (1/1), including openclaw config validate --json and full compaction-object readback.
• The source/runtime ownership boundary and removal condition are now documented; #4781 correctly remains open for live /compact completion and token-reduction proof.
• Trusted manual fork review run 29000033108: GPT-5.5 recommends merge_as_is with zero findings. Nemotron has no required-before-merge finding; I explicitly disposition its residual merge_after_fixes warnings as already covered, intentionally deferred to #4781, or incorrect (same-host case normalization/explicit ports should remain valid, and Local Ollama precedence already has coverage).
• CodeRabbit has no actionable exact-head finding; security review is clean.