fix(kilocode): model discovery overstates the usable context window by Yigtwxx · Pull Request #118868 · openclaw/openclaw

Yigtwxx · 2026-08-03T18:48:04Z

AI-assisted.

What Problem This Solves

Fixes an issue where users on the Kilocode provider would get model context
windows larger than the model can actually accept, whenever the primary
provider serving that model publishes a smaller window than the catalog-wide
ceiling.

Kilocode's gateway returns an OpenRouter-shaped catalog. In it, context_length
is the catalog-wide ceiling across all routing candidates, while
top_provider.context_length describes the primary provider that actually
serves the request. Model discovery read only the catalog-wide value.

On the live public catalog this affects 33 of the 335 discovered models. The
overstatement is up to 3.8x, for example:

model                              catalog     primary provider
nvidia/nemotron-3-super-120b-a12b  1000000     262144
minimax/minimax-m3                 1048576     524288
thinkingmachines/inkling           1048576     524288
qwen/qwen3.6-27b                    262144     131072
z-ai/glm-5.1                        204800     200000

Why This Change Was Made

Discovery now prefers top_provider.context_length and falls back to the
catalog-wide context_length, then to the existing default. This is the same
precedence already owned by extensions/openrouter/provider-catalog.ts:197-200,
so the two OpenRouter-shaped catalogs agree.

The same normalization was applied to src/agents/model-scan.ts in
8fa1b82e42a35447482f2722df75128916d0c059 (#110855). The Kilocode reader was
not covered there; this completes it.

Completion-token handling is deliberately left unchanged. The live Kilocode
catalog exposes no top-level max_completion_tokens or max_output_tokens
field at all, so the existing top_provider.max_completion_tokens read is
already the only available source and adding a fallback chain would encode a
field the data never carries.

No config, Plugin SDK, protocol, persistence, migration, or feature surface
changes.

User Impact

Kilocode models are registered with the context window the serving provider
actually offers, so context budgeting and model metadata stop overstating what
a request can use on the 33 affected models. Models whose primary provider
matches the catalog value are unchanged.

Evidence

Live public catalog, no key and no inference call:
curl -H "Accept: application/json" https://api.kilo.ai/api/gateway/models
returned 346 rows; 340 carry top_provider.context_length.

The real catalog response was then fed through the actual
discoverKilocodeModels() implementation, changing only the production file
between the two runs. registered is the contextWindow discovery assigns:

Before (current main):

ROWS=346 MODELS=335 DIVERGENT=33
minimax/minimax-m3       catalog=1048576  provider=524288  registered=1048576
thinkingmachines/inkling catalog=1048576  provider=524288  registered=1048576
stepfun/step-3.7-flash   catalog=262144   provider=256000  registered=262144
qwen/qwen3.6-27b         catalog=262144   provider=131072  registered=262144
z-ai/glm-5.1             catalog=204800   provider=200000  registered=204800

After (this branch):

ROWS=346 MODELS=335 DIVERGENT=33
minimax/minimax-m3       catalog=1048576  provider=524288  registered=524288
thinkingmachines/inkling catalog=1048576  provider=524288  registered=524288
stepfun/step-3.7-flash   catalog=262144   provider=256000  registered=256000
qwen/qwen3.6-27b         catalog=262144   provider=131072  registered=131072
z-ai/glm-5.1             catalog=204800   provider=200000  registered=200000

Tests, run on Node 24.18.1:

extensions/kilocode/provider-models.test.ts: 13/13 pass.
Whole extensions/kilocode/ folder: 35/35 pass across 4 files.
Control run: with the production file reverted to main and the new tests
kept, the precedence test fails with contextWindow 1048576 instead of
524288, so the added coverage is load-bearing rather than self-confirming.

Regression coverage added: primary-provider precedence, and independent
fallback to the catalog-wide value when the provider value is zero, negative,
fractional, non-finite, null, or a string.

oxfmt --check, direct oxlint on both touched files, and git diff --check
are clean.

Scope

Production delta: +6/-1 in extensions/kilocode/provider-models.ts. Tests add
53 lines. Changelog not required: this corrects existing internal catalog
normalization without changing commands or configuration.

The Kilocode gateway publishes an OpenRouter-shaped catalog in which `context_length` is the catalog-wide ceiling while `top_provider.context_length` describes the primary provider that actually serves the request. Discovery read only the catalog-wide value, so models whose primary provider offers a smaller window were registered with a context window larger than requests can use. Prefer the primary-provider window and fall back to the catalog-wide value, matching the precedence already owned by `extensions/openrouter/provider-catalog.ts`. Completion-token handling is left alone on purpose: the live Kilocode catalog exposes no top-level `max_completion_tokens` or `max_output_tokens`, so the existing `top_provider` read is already the only available source.

clawsweeper · 2026-08-03T19:03:57Z

Codex review: needs maintainer review before merge. Reviewed August 3, 2026, 3:02 PM ET / 19:02 UTC.

ClawSweeper review

What this changes

Updates Kilocode model discovery to prefer a valid primary-provider context window over the catalog-wide ceiling, with fallback coverage for unusable provider values.

Merge readiness

⚠️ Ready for maintainer review - 2 items remain

Keep open: current main still registers Kilocode models from the catalog-wide context value, while this narrow patch applies the established primary-provider precedence and adds focused regression coverage. The branch remains necessary and has no actionable correctness or security finding from this review.

Priority: P2
Reviewed head: 1919aab2e8310047c75282eafa7a97cae5cb2044

Review scores

Measure	Result	What it means
Overall readiness	🐚 platinum hermit (4/6)	A focused, contract-aligned repair with concrete live behavior evidence and load-bearing regression coverage.
Proof confidence	🐚 platinum hermit (4/6)	Sufficient (live_output): The PR body includes before/after live public-catalog output fed through the actual discovery implementation, showing the registered context window changes to the primary-provider value; no sensitive data is present in the supplied transcript.
Patch quality	🐚 platinum hermit (4/6)	No actionable review findings were identified.

Verification

Check	Result	Evidence
Real behavior	Verified	Sufficient (live_output): The PR body includes before/after live public-catalog output fed through the actual discovery implementation, showing the registered context window changes to the primary-provider value; no sensitive data is present in the supplied transcript.
Evidence reviewed	7 items	Current behavior: The current Kilocode converter uses only the catalog-wide `context_length` for `contextWindow`, so a smaller valid value under `top_provider` is ignored. Existing canonical sibling policy: The OpenRouter provider catalog already prefers a positive safe-integer `top_provider.context_length`, then falls back to the catalog-wide value and default; the proposed precedence matches that implementation. Related merged implementation: Current `src/agents/model-scan.ts` uses the same independent primary-provider context fallback, and commit `8fa1b82e42a35447482f2722df75128916d0c059` is contained by local main.
Findings	None	None.
Security	None	None.

How this fits together

The Kilocode provider plugin fetches its public model catalog and converts each row into OpenClaw model metadata used for context budgeting. The discovered context window then informs how much input OpenClaw considers usable for a selected Kilocode model.

flowchart LR
  A[Kilocode model catalog] --> B[Kilocode provider discovery]
  B --> C[Validate context limits]
  C --> D[Choose primary-provider limit]
  D --> E[Fallback to catalog limit]
  E --> F[OpenClaw model metadata]
  F --> G[Context budgeting]

Before merge

Resolve merge risk (P1) - build-artifacts and check-lint were still in progress for the reviewed head; merge should wait for their current-head outcomes.
Complete next step (P2) - No repair lane is needed: this is an already-focused contributor fix with no actionable review finding; wait for active checks and ordinary maintainer review.

Agent review details

Security

None.

PR surface

Source +5, Tests +53. Total +58 across 2 files.

View PR surface stats

Area	Files	Added	Removed	Net
Source	1	6	1	+5
Tests	1	53	0	+53
Docs	0	0	0	0
Config	0	0	0	0
Generated	0	0	0	0
Other	0	0	0	0
Total	2	59	1	+58

Review metrics

Metric	Value	Why it matters
Production vs test delta	production +6/-1; tests +53	The small production increase is justified by carrying the established provider-limit precedence into the Kilocode catalog reader, with focused regression coverage for valid and invalid values.

Root-cause cluster

Relationship: canonical
Canonical: #118868
Summary: This PR owns the remaining Kilocode-specific instance of the provider-limit normalization already merged for OpenRouter scans.

Members:

same_root_cause: fix(models): report accurate limits in OpenRouter scans #110855 - The merged scan repair uses the same primary-provider context precedence, but it operates in src/agents/model-scan.ts and does not update Kilocode discovery.

Proposal only: this assessment does not dispatch repair, suppress jobs, mutate sibling items, close, or merge anything.

Merge-risk options

Maintainer options:

Decide the mitigation before merge
Land the existing narrow precedence correction after the active checks finish, keeping the same positive-safe-integer fallback contract used by OpenRouter and model scanning.
Pause or close
Do not merge this PR until maintainers decide whether the risk is worth taking.

Technical review

Best possible solution:

Land the existing narrow precedence correction after the active checks finish, keeping the same positive-safe-integer fallback contract used by OpenRouter and model scanning.

Do we have a high-confidence way to reproduce the issue?

Yes, by source: a catalog row with a valid smaller top_provider.context_length currently produces the larger catalog-wide context window at extensions/kilocode/provider-models.ts:114. The submitted before/after public-catalog output also exercises that exact discovery path, although this runner could not independently resolve the API host.

Is this the best way to solve the issue?

Yes. Reusing the already-shipped OpenRouter and model-scan precedence is the narrowest maintainable fix; it corrects the producer of the metadata without adding a configuration or compatibility path.

AGENTS.md: found and applied where relevant.

Codex review notes: model internal, reasoning high; reviewed against c37cfb0f260f.

Labels

Label changes:

add P2: This is a bounded provider-metadata bug that can overstate usable context for affected Kilocode models but does not indicate data loss, security exposure, or an availability outage.
add proof: sufficient: Contributor real behavior proof is sufficient. The PR body includes before/after live public-catalog output fed through the actual discovery implementation, showing the registered context window changes to the primary-provider value; no sensitive data is present in the supplied transcript.
add rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🐚 platinum hermit and patch quality is 🐚 platinum hermit.
add status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (live_output): The PR body includes before/after live public-catalog output fed through the actual discovery implementation, showing the registered context window changes to the primary-provider value; no sensitive data is present in the supplied transcript.

Label justifications:

P2: This is a bounded provider-metadata bug that can overstate usable context for affected Kilocode models but does not indicate data loss, security exposure, or an availability outage.
rating: 🐚 platinum hermit: Overall readiness is 🐚 platinum hermit; proof is 🐚 platinum hermit and patch quality is 🐚 platinum hermit.
status: 👀 ready for maintainer look: ClawSweeper has no concrete contributor-facing blocker left for this PR. Sufficient (live_output): The PR body includes before/after live public-catalog output fed through the actual discovery implementation, showing the registered context window changes to the primary-provider value; no sensitive data is present in the supplied transcript.
proof: sufficient: Contributor real behavior proof is sufficient. The PR body includes before/after live public-catalog output fed through the actual discovery implementation, showing the registered context window changes to the primary-provider value; no sensitive data is present in the supplied transcript.

Evidence

What I checked:

Current behavior: The current Kilocode converter uses only the catalog-wide context_length for contextWindow, so a smaller valid value under top_provider is ignored. (extensions/kilocode/provider-models.ts:114, 5a65c96fc708)
Existing canonical sibling policy: The OpenRouter provider catalog already prefers a positive safe-integer top_provider.context_length, then falls back to the catalog-wide value and default; the proposed precedence matches that implementation. (extensions/openrouter/provider-catalog.ts:197, 5a65c96fc708)
Related merged implementation: Current src/agents/model-scan.ts uses the same independent primary-provider context fallback, and commit 8fa1b82e42a35447482f2722df75128916d0c059 is contained by local main. (src/agents/model-scan.ts:240, 8fa1b82e42a3)
Validation contract: The shared coercion helper accepts only positive safe integer numbers, so the proposed fallback correctly rejects zero, negatives, fractions, non-finite values, strings, and null. (packages/normalization-core/src/number-coercion.ts:99, 5a65c96fc708)
Current-main relevance check: The supplied current-main commit does not change either touched Kilocode file relative to the checked-out base, so the reported behavior is not already implemented on current main. (extensions/kilocode/provider-models.ts:114, c37cfb0f260f)
Feature-history provenance: Blame attributes the current conversion block to a recent provider-area change by Vincent Koc; recent history also includes a merged Kilocode catalog fix by mushuiyu886 and the parallel OpenRouter-scan repair. (extensions/kilocode/provider-models.ts:102, 8987e93c65e8)

Likely related people:

Vincent Koc: Local blame assigns the current Kilocode conversion block to Vincent's recent provider-area commit, and the extension history shows repeated provider and plugin maintenance work. (role: recent area contributor; confidence: high; commits: 8987e93c65e8, b88f80940add, 21728777dfa0; files: extensions/kilocode/provider-models.ts)
mushuiyu886: Authored the merged OpenRouter scan normalization that establishes the same provider-limit precedence and previously contributed a focused Kilocode catalog repair. (role: adjacent implementation contributor; confidence: high; commits: 8fa1b82e42a3, e3a1c64d7b08; files: src/agents/model-scan.ts, extensions/kilocode/provider-models.test.ts)

Rating scale

Score	Internal tier	Crab rank	Meaning
6/6	S	🦀 challenger crab	Exceptional readiness
5/6	A	🦞 diamond lobster	Very strong readiness
4/6	B	🐚 platinum hermit	Good normal PR; ordinary maintainer review
3/6	C	🦐 gold shrimp	Useful, but confidence is limited
2/6	D	🦪 silver shellfish	Proof or implementation needs work
1/6	F	🧂 unranked krab	Not merge-ready
N/A	NA	🌊 off-meta tidepool	Rating does not apply

Overall follows the weaker of proof and patch quality.
Shiny media proof means a screenshot, video, or linked artifact directly shows the changed behavior. Runtime, network, CSP, and security claims still need visible diagnostics.

Workflow

ClawSweeper keeps one durable marker-backed review comment per issue or PR.
Re-runs edit this comment so the latest verdict, findings, and automation markers stay together instead of adding duplicate bot comments.
A fresh review can be triggered by eligible @clawsweeper re-review comments, exact-item GitHub events, scheduled/background review runs, or manual workflow dispatch.
PR/issue authors and users with repository write access can comment @clawsweeper re-review or @clawsweeper re-run on an open PR or issue to request a fresh review only.
Maintainers can also comment @clawsweeper review to request a fresh review only.
Fresh-review commands do not start repair, autofix, rebase, CI repair, or automerge.
Maintainer-only repair and merge flows require explicit commands such as @clawsweeper autofix, @clawsweeper automerge, @clawsweeper fix ci, or @clawsweeper address review.
Maintainers can comment @clawsweeper explain to ask for more context, or @clawsweeper stop to stop active automation.

Takhoffman · 2026-08-03T20:44:01Z

Land-ready for PR #118868.

Reviewed and prepared head: 1919aab2e8310047c75282eafa7a97cae5cb2044.
ClawSweeper exact-head review: no actionable findings; current-head review comment 5170572558 is clear, and no separate Rank-up moves: list was present.
Focused proof: scripts/pr review-tests 118868 extensions/kilocode/provider-models.test.ts — 13 tests passed; git diff --check origin/main...HEAD passed.
Hosted exact-head CI: openclaw/ci-gate, build-artifacts, check-lint, and preflight are successful in CI run 30842897375.
Canonical preparation: OPENCLAW_TESTBOX=1 scripts/pr prepare-run 118868 passed with hosted_exact_or_recent_parent; the prepared head equals the live PR head and no patch was pushed.
Review artifacts were validated with scripts/pr review-validate-artifacts 118868.

Known proof gap: no separate Testbox lease was needed because the exact-head hosted CI gate was already green. The standard canonical squash path will attribute the merge to @Takhoffman; no additional graph co-author trailer for the merger was requested.

Takhoffman · 2026-08-03T20:44:51Z

Merged via squash.

Prepared head SHA: 1919aab2e8310047c75282eafa7a97cae5cb2044
Landed commit: 38aa704a313ae2254f1a9b640877c22b9ef19fa4

* commit 'efcced7207baa0b1b1db5e8e1836673a24212657': test(agents): prove process control roundtrip (#118843) test(qa): add memory redaction boundary coverage (#118916) fix(kilocode): model discovery overstates the usable context window (#118868) fix(media): isolate generated-provider text from delivery directives (#118780) test(gateway): add node control-plane QA coverage (#118784) fix(config): reject unsupported diagnostics otel grpc (#93087) fix(providers): recognize configured media credentials across provider boundaries (#118761) refactor: delete dead post-landing plugin/hooks scaffolding (#118920) fix(cron): deliver child results after heartbeat-only parent responses (#118743) fix(ci): keep Matrix chunk tests aligned with inventory (#118917) fix(tts): preserve hidden speech answers when synthesis fails (#118690)

…-worker-update-20260803 * origin/main: (48 commits) test(ui): add shell and PWA primary QA proof (#118788) test(qa): cover CLI distribution workflows (#118948) chore(i18n): refresh native locales (#118883) test(qa): prove plugin tool ownership in gateway flow (#118958) docs: correct onboarding risk flags and Google Chat credential shapes (#118936) fix(imessage): honor explicitly enabled accounts during startup (#118944) test(gateway): cover client transport defaults (#118798) fix(signal): normalize accounts during noninteractive setup (#118932) fix(msteams): enable channels for federated environment credentials (#118931) fix(openai): enforce final transcript bounds before callback delivery (#118928) test(tasks): cover plugin-managed flows (#118919) test(agents): prove process control roundtrip (#118843) test(qa): add memory redaction boundary coverage (#118916) fix(kilocode): model discovery overstates the usable context window (#118868) fix(media): isolate generated-provider text from delivery directives (#118780) test(gateway): add node control-plane QA coverage (#118784) fix(config): reject unsupported diagnostics otel grpc (#93087) fix(providers): recognize configured media credentials across provider boundaries (#118761) refactor: delete dead post-landing plugin/hooks scaffolding (#118920) fix(cron): deliver child results after heartbeat-only parent responses (#118743) ...

…penclaw#118868) The Kilocode gateway publishes an OpenRouter-shaped catalog in which `context_length` is the catalog-wide ceiling while `top_provider.context_length` describes the primary provider that actually serves the request. Discovery read only the catalog-wide value, so models whose primary provider offers a smaller window were registered with a context window larger than requests can use. Prefer the primary-provider window and fall back to the catalog-wide value, matching the precedence already owned by `extensions/openrouter/provider-catalog.ts`. Completion-token handling is left alone on purpose: the live Kilocode catalog exposes no top-level `max_completion_tokens` or `max_output_tokens`, so the existing `top_provider` read is already the only available source.

openclaw-barnacle Bot added extensions: kilocode size: S labels Aug 3, 2026

Takhoffman self-assigned this Aug 3, 2026

Takhoffman merged commit 38aa704 into openclaw:main Aug 3, 2026
94 of 101 checks passed

github-actions Bot mentioned this pull request Aug 3, 2026

📡 Upstream Digest — 2026-08-03 20:56 UTC curtismercier/openclaw-mods#1488

Open

mushuiyu886 mentioned this pull request Aug 4, 2026

fix(chutes): report accurate context windows for sparse catalog rows #119196

Merged

Navigation Menu

Uh oh!

Uh oh!

fix(kilocode): model discovery overstates the usable context window - #118868

fix(kilocode): model discovery overstates the usable context window#118868
Takhoffman merged 1 commit into
openclaw:mainfrom
Yigtwxx:fix/kilocode-primary-provider-context-window

Yigtwxx commented Aug 3, 2026

Uh oh!

clawsweeper Bot commented Aug 3, 2026

Security

PR surface

Review metrics

Root-cause cluster

Merge-risk options

Technical review

Labels

Evidence

Rating scale

Workflow

Uh oh!

Takhoffman commented Aug 3, 2026

Uh oh!

Uh oh!

Takhoffman commented Aug 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

Uh oh!

Uh oh!

Conversation

Yigtwxx commented Aug 3, 2026

What Problem This Solves

Why This Change Was Made

User Impact

Evidence

Scope

Uh oh!

clawsweeper Bot commented Aug 3, 2026

ClawSweeper review

What this changes

Merge readiness

Review scores

Verification

How this fits together

Before merge

Security

PR surface

Review metrics

Root-cause cluster

Merge-risk options

Technical review

Labels

Evidence

Rating scale

Workflow

Uh oh!

Takhoffman commented Aug 3, 2026

Uh oh!

Uh oh!

Takhoffman commented Aug 3, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants