* fix(compression): add i18n support for less-code and terse-prose
Translates less-code output style to pt-BR, vi, ja, and id. Adds missing vi translation to terse-prose caveman mode. Removes less-code from English-only allowlist and updates matrix tests.
Fixes#10426
* docs(compression): add output styles coverage table
Adds the requested Output Styles matrix to the compression guide covering styles, supported languages, and intensity levels.
Fixes#10426
* fix(memory): auto-check Qdrant health on mount and stop false-red badge
The Qdrant engine card on /dashboard/memory?tab=engine showed a red
"Error" badge after every page refresh even when Qdrant was healthy:
the badge derives its state from a health check, but the mount effect
only fetched settings + embedding models — health started as null and
the render treated `health?.ok` (undefined) as a failure. Clicking
"Test connection" (which runs the same server-side /readyz check)
immediately turned it green, proving the connection was fine.
Two changes:
- Auto-run the health check on mount once settings load and Qdrant is
enabled, so a refreshed page reflects the real state (verified live:
/api/settings/qdrant/health returns ok:true in ~2ms on a healthy
compose deployment).
- While health has not been checked yet (null), render a neutral gray
"Testing..." state instead of red — red is now reserved for an
actual failed health check.
Regression test added (fails on the old code): with enabled settings
and a healthy mock, the card must hit /api/settings/qdrant/health on
mount and show statusActive, never statusError.
* chore(changelog): fragment for #10489
* Merge branch 'release/v3.8.50' into fix/qdrant-health-badge
* test(fix): refresh expired alibaba quota sample validity and onnxruntime pin for v3.8.50 base
- alibaba-free-tier-quota-fetcher.test.ts: sample quotaValidityPeriod
(2026-08-16 16:00 UTC) is in the past, making every quota entry classify
as expired/not_capable; bump to 2028-01-01 UTC so the text/merge
classification tests exercise the intended path again.
- optional-transformers-dependency.test.ts: onnxruntime-node pin assertion
updated from ~1.24.3 to ~1.27.0 to match package.json (bumped by #10403);
the regular-not-optional intent is unchanged.
* test(fix): align optional-transformers-dependency with onnxruntime ~1.24.3 pin (base #10543)
* docs(fix): sync 150-migration count and document PROXY_LOG_INCLUDE_IPS (base drift #10348/#10507)
* fix(memory): re-check Qdrant health after saving settings
save() optimistically flipped enabled and started the PUT while the mount
effect could immediately GET /api/settings/qdrant/health against the OLD
persisted settings. If that GET won, it returned not_configured/failed and -
because health was non-null - the effect never retried after the PUT
succeeded, leaving a healthy Qdrant red until a manual Test connection.
Invalidate health (generation counter + setHealth(null)) at save start and
after a successful PUT, then explicitly schedule a fresh check: setting
health to null alone is not enough, React bails on the no-op when health is
already null (the exact GET-wins ordering). Stale responses are dropped via
the sequence guard so an in-flight pre-save check can never overwrite the
post-save result. Adds a regression test covering enable ordering.
Addresses PR #10489 review finding (issuecomment-5312271806).
* fix: narrow omniglyph transform result union (merge base aa912c42a typecheck gate)
* test(compression): align contract tests with base aa912c42a merge (providerTransport shape, engine metadata)
* fix(memory): silence set-state-in-effect on Qdrant auto health-check
The health-check re-check fix (3469234) introduced an effect that calls
checkHealth() (an async fetch that eventually calls setState) directly
from a useEffect gated on loading/enabled/health. The
react-hooks/set-state-in-effect rule flags this as a potential cascading
render, matching the same pattern already accepted elsewhere in the
dashboard (FreePoolTab.tsx, ConnectionsTable.tsx) for gated async
data-fetch effects. Suppress with the established inline convention;
no behavior change.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(memory): drop unused set-state-in-effect disable (rule inert on pinned react-hooks 7.0.1)
The eslint-disable-next-line for react-hooks/set-state-in-effect is unused:
eslint-plugin-react-hooks@7.0.1 (lockfile-pinned) does not report this rule,
so the directive itself was flagged as a warning and the 'No new ESLint
warnings' CI gate failed with --max-warnings 0. The effect body only calls
checkHealth() (async fetch) with no raw setState, so no disable is needed.
* ci(quality): sync ratchet configs to release/v3.8.50 (0a74bfbde) merge
- re-freeze open-sse typecheck baseline at merged-tree live counts
(64 stale entries dropped, 11 frozen; base video/usage drift covered)
- register tests/unit/video-bridge-drilldown-route.test.ts in stryker tap.testFiles
- regenerate skills/cli-contexts/SKILL.md (contexts migrate docs from CLI closure)
---------
Co-authored-by: Rouzbeh <rqzbeh@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
Fecha o caminho do operador: o perfil já existia no adapter, mas só como
default de código. Agora atravessa schema Zod, normalizador do banco, API de
settings e a página dedicada do engine.
- OmniglyphConfig tipado + omniglyphConfigSchema (z.enum dos quatro perfis);
- normalizeOmniglyphConfig: nome desconhecido vindo do storage cai para o
default em vez de virar "roda com a política padrão";
- seletor na página do engine, com PATCH próprio — o perfil vive fora do mapa
`engines`, e mandá-lo junto reescreveria o mapa inteiro (o store persiste o
mapa como uma linha JSON só);
- i18n en/pt-BR descrevendo o custo medido de cada perfil, não só o nome;
- README e COMPRESSION_ENGINES.md com a regra do teto e o motivo de o default
não ser o perfil mais seguro.
Corrige de passagem um teste-irmão que ninguém via: o gate de transporte na UI
deixou de dizer "direct Anthropic" quando os wires OpenAI nativos entraram, mas
tests/unit/ui/omniglyphContextPage.test.tsx continuou afirmando a cópia antiga.
O arquivo inteiro estava excluído do vitest.config.ts como "#8618 pre-existing
failure", então a quebra passou silenciosa. Com a asserção alinhada o arquivo
fecha 3/3, e a exclusão sai — o próprio comentário mandava removê-la quando
corrigida.
A doc não nomeia OMNIGLYPH_MODELS: o gate de docs fabricadas está certo em
apontar que o OmniRoute nunca lê essa env — quem lê é o pacote.
O adapter descartava o TransformInfo inteiro, então a UI mostrava um número de
economia sem dizer de onde ele vinha — contagem do provider, estimativa ou só
diferença de bytes. O 1.4.0 expõe normalizeAccounting(), que classifica essa
evidência e resolve a semântica de cache por família: Anthropic reporta input,
cache-create e cache-read em buckets DISJUNTOS, enquanto OpenAI e xAI reportam
cached como SUBCONJUNTO do input. Somar à mão dá double-count silencioso.
O novo omniglyphTelemetry.ts não filtra por denylist — MONTA um objeto novo,
campo a campo, só com número e enum. TransformInfo mistura contadores
inofensivos com material que não pode ser persistido: bytes PNG,
imageSourceText(s), recoverable[].text, os sha8 de system/CLAUDE.md/primeira
mensagem, nomes de tags observadas e o bloco env (cwd, branch, versões). Copiar
o objeto inteiro transformaria telemetria de compressão em vazamento de prompt.
O teste de negação prova que segredo, caminho do operador, texto do system e
base64 não aparecem, e varre a allowlist exigindo que toda string seja de um
enum conhecido.
- provider threaded do chatCore e do bridge Codex WS até a engine; ausente vira
`unknown`, que faz o upstream recusar adivinhar buckets de cache;
- contabilidade propagada para o engineBreakdown do passo (o agregado do
pipeline soma todas as engines e não serviria);
- skip não emite contabilidade: zeros ali seriam indistinguíveis de "a engine
nem rodou".
O 1.4.0 trouxe perfis nomeados (coding-safe, balanced, aggressive,
passthrough), mas só transformAnthropicMessages() os resolve sozinho: os
transformadores OpenAI recebem TransformOptions cru e ignorariam o campo. Um
perfil escolhido pelo operador valeria no wire Claude e sumiria no OpenAI. O
adapter passa a mesclar o perfil com mergeCompressionProfileOptions() antes de
chamar Chat Completions e Responses.
O default segue aggressive — a política que os recibos publicados mediram.
Medido nesta base: com coding-safe/balanced, uma sessão sem histórico acumulado
para em below_min_chars e a engine não faz nada, porque os dois fixam
minCompressChars no máximo e desligam system/tools/tool-results. Como a engine é
opt-in, um default assim entregaria "ligado, 0% de ganho".
O perfil é TETO, não piso: mergeCompressionProfileOptions não deixa um override
do chamador reabrir uma lane lossy que o perfil fechou. Coberto por teste, por
ser contra-intuitivo.
Também fecha um caminho em que o OmniRoute violaria a própria política: o wire
OpenAI do pacote não tem compressSystem — honra apenas compressTools,
gptHistory, minCompressChars e reflow, e sempre troca a instrução por um
ponteiro para a imagem. Com preserveSystemPrompt ligado, imagear assim queimaria
o prefixo quente que a decisão cache-aware está protegendo, sem nada no corpo
devolvido denunciando. A engine agora pula com
skip:system_preservation_unsupported_on_wire.
O 1.4.0 introduziu escopos de segurança e passou a resolvê-los dentro de
isOmniGlyphSupportedModel() lendo process.env.OMNIGLYPH_PROFILE. Somado ao
OMNIGLYPH_MODELS que já existia, duas variáveis do ambiente do host decidiam em
silêncio o gate de TODO request do OmniRoute: passthrough desligaria a engine
inteira e OMNIGLYPH_MODELS admitiria modelos sem recibo medido, enquanto a UI
segue prometendo "Claude Fable 5 na rota direta medida".
O adapter passa a usar isOmniGlyphSupportedModelForScope() com escopo explícito
e fixa o escopo mais restrito como teto: a env só pode ESTREITAR a allowlist,
nunca alargar. Os dois wires compartilham a mesma lista no pacote desde o
1.4.0, então uma checagem cobre Anthropic e GPT.
- omniglyph ^1.3.1 -> ^1.4.0 (lock em 1.4.0);
- testes de regressão para os dois caminhos de sequestro por env;
- teste de contrato dos exports novos (escopo, perfis, accounting).
O 1.4.0 também traz, sem mudança de código aqui: correção do glyph K que era
lido como H, remoção do backtracking polinomial no secret-guard, overrides do
pnpm em pnpm-workspace.yaml e as transitivas vulneráveis resolvidas.
Roda o OmniGlyph depois da tradução para o wire real do provedor, em vez do
corpo de origem. Um cliente OpenAI roteado para Claude deixava de comprimir com
skip:source_format_not_claude porque o corpo ainda estava em formato OpenAI
quando a engine era avaliada.
- dispatch nativo por wire: Anthropic Messages, OpenAI Chat Completions e
OpenAI Responses (input[] preservado, sem achatar para messages[]);
- estágio target-wire pós-translateRequest, com guarda contra dupla compressão
no caminho Claude→OpenAI;
- preserveSystemPrompt do OmniRoute mapeado para compressSystem: false;
- imageTransportPolicy: fidelidade de bytes/dimensões separada de supportsVision;
só Anthropic/Claude tem recibo byte-preserving, o resto é fail-closed;
- contagem de tokens de data URL PNG no wire OpenAI (marcador ;base64,);
- README e i18n en/pt-BR com claims escopados ao caminho medido.
i-have-adhd shipped in #10271 with en + pt-BR only; it now covers vi/ja/id as well,
matching ponytail. Every level keeps the SHARED_BOUNDARIES clause.
Adds a per-MATRIX guard (output-styles-i18n-matrix.test.ts). Every existing test is
per-style, which is how less-code stayed English-only since the 9router port without
anyone noticing. The guard fails on: a new style without pt-BR, a style losing a
translation it had, a translation missing an intensity level, a translation that dropped
the boundaries clause, and a stale KNOWN_ENGLISH_ONLY entry.
Proven by mutation, not just by passing: dropping less-code from the allowlist and
renaming the vi key both turned it red with the expected messages.
less-code stays English-only as declared debt (KNOWN_ENGLISH_ONLY + comment); the
remaining coverage work is tracked in #10426.
Adds `i-have-adhd` as the 5th entry in OUTPUT_STYLE_CATALOG — a port of the
github.com/ayghri/i-have-adhd skill (MIT), following the same integration shape as
ponytail. Action-first output shaping: the next action leads, multi-step work is
numbered, no preamble/recap/closers — which also trims output tokens.
lite/full/ultra levels in en + pt-BR, each ending in SHARED_BOUNDARIES so code, paths,
commands, errors and URLs stay verbatim. The agent-harness-specific upstream rules
(restate plan state, time estimates) are reworded as conditionals so they hold for plain
chat clients too.
Per the D-A1 registry contract, one catalog entry is the whole change: the injector, the
settings panel, the Zod schema and the telemetry all enumerate the catalog, so no other
production file moves. Dedicated test mirrors ponytail-catalog.test.ts (7 tests).
* Fix custom tool output pairing during compression (#8932)
* Bypass proxy compaction for native Codex context
* fix(sse): extract Codex tool-call output repair to leaf module for file-size gate
repairMissingCodexToolCallOutputs (added by #8932 for custom_tool_call
pairing) pushed codex.ts past the frozen file-size baseline. Extract it
to open-sse/executors/codex/toolCallRepair.ts, leaving only the wiring
call in codex.ts. Rebaseline the test file's genuine +41 line growth
from #8932's new custom_tool_call_output coverage.
Co-authored-by: JxnLexn <JxnLexn@users.noreply.github.com>
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
Co-authored-by: JxnLexn <JxnLexn@users.noreply.github.com>
* fix(compression): add Lite tool truncation toggle
* fix(antigravity): add missing antigravityProjectPersistence.ts module
The quota-strategy engine (quotaStrategies.ts) imports from
antigravityProjectPersistence.ts, but only antigravityProjectPersist.ts
existed in the tree. Add the missing module with the expected
preferAntigravityConnectionsWithStoredProject() helper and re-export
the existing persistDiscoveredAntigravityProjectId().
Co-authored-by: diegosouzapw <diegosouza.pw@outlook.com>
* fix(file-size): rebaseline strategySelector.ts for Lite truncation toggle
The PR adds one line to threading options?.config?.lite into
applyLiteCompression. Update the frozen size from 1060 to 1061.
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
Refs #9629
---------
Co-authored-by: Xiangzhe <xiangzhedev@gmail.com>
Co-authored-by: xz-dev <xz-dev@users.noreply.github.com>
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* fix(compression): persist RTK renderer configuration
* docs(changelog): add fragment for #9730
Adds the changelog.d/fixes/9730-persist-rtk-renderers.md fragment
required by check:changelog-integrity for the RTK enableRenderers
persistence fix in PR #9730.
---------
Co-authored-by: Isaac <isaaclyons98@gmail.com>
* test(base): realign six suites with contracts that #9100/#8990/#9009 deliberately changed
Continuing the base-red drain — every one of these reproduces on the pure tip.
- tests/snapshots/provider/translate-path.json: regenerated via UPDATE_GOLDEN=1.
The diff is ADDITION-ONLY — the unorouter block from #9009; no existing
provider entry changed. 3/3.
- tests/unit/provider-models-route.test.ts: ff012ff420 added onboardUser as a
bootstrap fallback next to loadCodeAssist; the mock now excludes it from the
discovery-URL ledger like it already excluded loadCodeAssist, otherwise it
consumed the injected 503 and the retry assertion misfired. 59/59.
- tests/unit/responses-commentary-passthrough-6199.test.ts: #8990 (c996dc93c2)
deliberately preserves `tools` on the TERMINAL response.completed snapshot
(Codex CLI rebuilds its tool list from it); the assertion now pins the echoed
tools instead of their absence. Still stripped on created/in_progress. 7/7.
- tests/unit/vision-compression-authoritative-capability-7237.test.ts:
68cb678780 added the 'gpt-5' fragment, so the heuristic-vs-spec DRIFT this
suite documented no longer exists; the cases now guard the agreement, keep a
conservative-for-unknown-ids probe, and reproduce the strip-bug shape with an
explicit false instead of deriving it. 4/4.
- tests/unit/provider-limits-proxy-fail-closed.test.ts +
tests/unit/image-generation-route.test.ts: #9100 made the proxy reachability
probe NON-BLOCKING (optimistic dispatch; the probe aborts only in-flight
requests — its own t14 sibling was updated to this exact pattern). Instant
mocks therefore won the race and the PROXY_UNREACHABLE 503 became unobservable
(a success or a generic 502). The mocks now stay in flight (never-resolving,
so the aborted continuation cannot reach the restored real fetch), and the
fail-closed proof is the settled rejection itself plus zero egress AFTER the
fast-fail. Production fail-closed semantics are unchanged — the proxy dispatch
path still throws; only the mock timing was stale. 3/3 and 20/20.
Refs #9298
* fix(guardrails): forward the router deps seam through callVisionModel
tests/unit/guardrails/vision-bridge-sse-and-reasoning.test.ts was 7/7 red on any
clean box (CI shard 3/4): callVisionModel() called getBestVisionModel()/
getFallbackModels() WITHOUT the routers' existing VisionBridgeRouterDeps seam,
so the credential check always hit the live connections DB — no vision-capable
connection meant 'No vision-capable provider connected' before the mocked fetch
was ever reached, and on a dev box auto-selection could swap the fixed model
under the assertions.
The routers already accepted deps; only the forwarding was missing. Added the
optional 5th param (backward compatible — the sole production caller,
visionBridge.ts, injects its own callVisionModel and is unaffected) and the
suite now pins selection with hasUsableCredentials: async () => null
(indeterminate → the fixed model is honored, DB untouched). 7/7.
Sibling suites re-run green: vision-bridge-callmodel 2/2, visionBridge 25/25,
visionBridgeHelpers.callVisionModel 8/8, visionBridgeRouter 10/10,
vision-bridge-cc-no-reroute 8/8.
Refs #9298
* fix(db,combo): clear the NEW base-reds the 08-06 merge batch introduced
The tip moved while the first sweep PR (#9600) was in review, and three fresh
base-reds landed with it — same classes as before, all reproduced on the pure
tip 9995bc4893:
1. ANOTHER migration collision: #9061 shipped 134_ccr_blocks.sql onto the slot
134_proxy_logs_egress_ip.sql (#9291) has held since 08-04. getMigrationFiles()
throws on collision, so every DB-touching test died at bootstrap again.
Renumbered to 139 (next free slot). No retroactive guard needed this time:
both statements are IF NOT EXISTS, and no DB can have applied it as 134 —
the runner refused to run at all while the collision existed.
2. BROKEN IMPORT killing the combo module graph: #8894 imported
preferAntigravityConnectionsWithStoredProject from
../antigravityProjectPersistence.ts — a module that exists NOWHERE in the
repo (it came from an unmerged sibling branch). Anything importing
quotaStrategies.ts died with ERR_MODULE_NOT_FOUND. Implemented the helper in
the real persistence module (antigravityProjectPersist.ts, #8491) with the
semantics the call site needs — prefer connections that already carry a
stored projectId, never emptying the pool — and pointed the import there.
New regression suite tests/unit/antigravity-prefer-stored-project.test.ts
(5/5), including an import-graph probe that reproduces the break shape.
3. Sibling-test drift from #9106 (gemini-3.1-pro-high now user-callable): its
own suites were updated but provider-models-route.test.ts was not. Expected
discovery list realigned; testFrozen 1784->1787 justified in the baseline
(irreducible +2 after comment compression; gate counts split-newlines).
Also regenerated tests/snapshots/provider/translate-path.json — addition-only:
devin-cli-agentic, raycast, regolo (today's provider merges), zero removals.
image-generation-route 20/20 (was import-dead), provider-models-route 59/59,
antigravity-prefer-stored-project 5/5, provider-translate-path-golden 3/3.
Refs #9298
* fix(changelog): convert the #9415 fragment to the required bullet shape
Another base-red from the 08-06 batch: bd4407cb64 landed
changelog.d/features/9415-newapi-sub2api-aggregator-balance.md as YAML
frontmatter + a prose paragraph. Every other fragment in changelog.d/ is a
single markdown bullet, and both consumers enforce that —
scripts/check/check-changelog-integrity.mjs:97 and the release aggregator
(scripts/release/aggregate-changelog.mjs:57) reject anything that does not
start with '- ', so 'Merge integrity (changelog + generated skills)' was red
for every PR targeting the release branch.
Rewritten as a bullet with the standard issue link, preserving the feature
description (aggregator gateway toggle, /api/user/self balance read, dashboard
badge, quota-preflight skip, NEWAPI_AGGREGATOR_BALANCE flag default off,
quotaPerUnit override). Swept the rest of changelog.d/ — this was the only
malformed fragment.
check:changelog-integrity OK.
Refs #9298
* fix(types,docs): clear the 5 typecheck errors and the fabricated env vars on the base
Third pass over the base-reds, from the 2026-08-06T22:51Z verdict on #9298 —
it reported "Typecheck (core)" with only the FIRST error; there are five, all on
the pure tip 9995bc4893. Two are real production defects.
**Real bugs**
- open-sse/services/compression/engines/ccr/index.ts:295 called
enforceGlobalBudget(entry.bytes) against an (owner, bytes) signature. The
`bytes` argument arrived undefined, so `ccrTotalBytes + undefined` is NaN,
`NaN > MAX` is false (the eviction loop exits immediately) and `NaN <= MAX` is
false (the re-admit is refused). The #9061 durable tier therefore NEVER
repopulated its in-memory map: every retrieve after a restart or an eviction
re-read from SQLite forever, and evictions could not prefer the owning
principal. Fixed and pinned by a new case in
tests/unit/ccr-durable-store-9061.test.ts (11/11) — verified failing against
the buggy call and passing against the fix.
- open-sse/services/combo/fusionPanel.ts:54 read `step.model` after #8894
widened ComboStep with ComboProviderWildcardStep (which carries modelPattern,
not model), so a wildcard step in a fusion panel pushed `undefined` onto the
panel. Now resolved through getComboModelString(), which already handles every
step shape and returns null for the ones without a concrete model id.
**Type-only**
- accountSemaphore.ts:203 — isBypassed() returns a plain boolean and cannot
narrow `number | null` (an `x is null | undefined` predicate would be unsound:
0 bypasses too). Added resolveActiveCap(), the narrowing companion isBypassed
is now defined in terms of; the acquire path uses the narrowed value.
- comboStructure.ts:140 — same #8894 widening: `prompt` only exists on a model
step, so it is now read under a kind check.
- firecrawlQuotaFetcher.ts:136 — the function returns full FirecrawlQuota
objects but was annotated Promise<QuotaInfo | null>, which made the
custom-base literal an excess-property error. Widened to the accurate type
(FirecrawlQuota extends QuotaInfo, so callers are unaffected).
**Fabricated docs (the "Docs sync + fabricated-docs (strict)" HARD failure)**
docs/ops/VM_DEPLOYMENT_GUIDE.md recommended OMNIROUTE_MAX_POOL_SIZE and
OMNIROUTE_DB_POOL_SIZE (#9471). Neither is read anywhere in the codebase.
Replaced with the two knobs that do exist and are already documented in
ENVIRONMENT.md: OMNIROUTE_MEMORY_MB and OMNIROUTE_CHAT_MAX_HEAVY_IN_FLIGHT.
typecheck:core 5 errors -> 0. check:fabricated-docs + check:env-doc-sync OK.
accountSemaphore 6/6, ccr-durable-store 11/11, ccr-protocol 9/9,
combo-fusion-strategy 10/10, combo-fusion-comboref 5/5, combo-fusion-warn 4/4,
firecrawl-executor 7/7, executor-firecrawl-fetch 4/4.
Refs #9298
* fix(tests): type the #3440 vertex helpers instead of `any` (the 3 base ESLint errors)
The "ESLint errors: 3 error(s)" HARD failure in the #9298 verdict is
tests/unit/vertex-functioncall-id-3440.test.ts lines 32/41/50: the three
find*(result: any) walkers. `@typescript-eslint/no-explicit-any` is an ERROR in
tests/ (and open-sse/) since #6218, and this file landed on 2026-08-04 without a
suppressions entry, so every run of `lint:json --max-warnings 0` failed. That
step prints nothing on failure, which is why the gate looked like a silent
crash across the open PRs.
Replaced with a GeminiRequestLike interface describing exactly what the three
walkers traverse (contents[].parts[]), so the assertions keep their meaning and
nothing is cast away.
eslint on the file: clean. Suite: 6/6.
Refs #9298
* docs(proxy): use an RFC 5737 documentation IP in the proxy examples
The #9298 verdict headlines its docs failure with
`L810 [stale-version] 1.2.3: const removed = await failOneproxyProxy("1.2.3.4", 8080)`.
That is a false positive: check-deprecated-versions.mjs matches
`/\bv?[12]\.\d+\.\d+\b/`, and the example IP literal 1.2.3.4 contains "1.2.3".
Swapped both occurrences in PROXY_GUIDE.md (and its pl mirror) for 203.0.113.7,
from the RFC 5737 documentation range that exists precisely for examples — it
cannot collide with a version pattern and is the correct thing to print in docs
regardless. Drift count 64 -> 62; no gate threshold was touched.
The gate that actually FAILED under "Docs sync + fabricated-docs (strict)" was
check:fabricated-docs (the invented pool env vars), fixed in the previous
commit; this one removes the misleading line the verdict quotes.
* test(base): allowlist probeUtils and realign the #7849 suite to the replacement bound
Two more base-reds, both visible only after the migration collision stopped
killing the shards.
**check-db-rules — src/lib/db/probeUtils.ts not classified**
#9541 added probeUtils.ts (transient-error retry for the SQLite corruption
probe). It is imported ONLY by src/lib/db/core.ts, exactly like its siblings
schemaColumns / optimizationSettings / providerNodeSelect, so re-exporting it
through localDb.ts would push callers toward the barrel-import anti-pattern the
gate exists to prevent. Added to INTENTIONALLY_INTERNAL with that rationale.
check-db-rules 22/22, check:db-rules exit 0.
**session-dedup-memory-7849 — pinned a mechanism that was replaced**
7f36b192f0 (#7855 follow-up) swapped the shared "suffix work budget" for the
MAX_SUFFIX_STARTS / MAX_TOTAL_BLOCK_BYTES guards and deleted both the budget and
its SUFFIX_WORK_BUDGET_WARNING string. It updated session-dedup.test.ts but not
this sibling, so 3 of its 4 cases asserted a warning that can no longer be
emitted.
Realigned to the contract that actually survives — which is the invariant #7849
was opened for, not the mechanism:
- the pathological pair must stay BOUNDED (completes in <4s, body intact) —
measured at ~280ms on the current guards;
- it must FAIL OPEN — original body returned by identity, compressed false,
stats null (the explanatory zero-savings stats belonged to the removed
budget path, which skipped before producing any);
- the 512 MiB child fixture must still exit 0 with the full engine chain
(session-dedup, lite, rtk, headroom, caveman) — that IS the OOM guard — and
session-dedup must still report its skip, now pinned by prefix since the
reason string moved with the mechanism.
No threshold was loosened and no case was deleted: 4/4 here, 8/8 on the sibling
session-dedup.test.ts.
Refs #9298
* docs(mcp): bump the tool count to 105 and realign two vitest count pins
Three more base-reds from the same 08-06 batch, all count/contract drift that
the merged PRs left in sibling files.
**Docs Gates (fast-path) — 3 STRICT drifts**
check:docs-counts measures the MCP tool set from live code: it is 105 now
(#8925 added omniroute_create_combo), while README.md, AGENTS.md and
docs/frameworks/MCP-SERVER.md still claimed 104. Updated all five occurrences
(two of them inside SVG alt text). check:docs-all exits 0.
**Vitest (fast-path) — 2 failures**
- open-sse/mcp-server/__tests__/essentialTools.test.ts pinned 11 phase-1 tools;
#8925 shipped omniroute_create_combo as phase 1, making it 12. Verified by
enumerating MCP_ESSENTIAL_TOOLS directly.
- tests/unit/autoCombo/provider-family-combos.test.ts pinned the auto/glm
provider set to [auggie, glm, zai]. #8914 (Devin ACP bridge) added
devin-cli-agentic, whose catalog (registry/devin/catalog.ts:90-93) advertises
the glm-5-2* line — so it belongs in the family pool for exactly the reason
the test's own comment gives for auggie: a no-auth backend that genuinely
serves a family model is a legitimate member. Expected set updated, invariant
unchanged.
npm run test:vitest 36/36 files, 340/340 tests.
Refs #9298
* fix(combo,usage,oauth): drain the base-reds the shard fix exposed
With the migration collision and the broken import out of the way the four unit
shards actually run, and a further layer of base-reds became visible on the pure
tip 9995bc4893. Three are production defects.
**Production defects**
- open-sse/services/combo/runtimeUnitCapacity.ts:58 called resolveComboTargets()
WITHOUT the hidden-model snapshot, so it fell back to the default
getHiddenModelsByProvider() — a fresh full key_value read PER nested combo-ref
unit, on every request. #8878 threaded the snapshot through the other call
sites and missed this one. Threaded it from executeRuntimeUnitCombo (and from
the dispatchPrelude call site), restoring the one-snapshot-per-request
invariant combo-hidden-leaf-routing.test.ts pins. 9/9.
- open-sse/services/usage/firecrawl.ts silently ignored its own `apiKey`
parameter: 91bb6aa619 moved the fetch to
fetchFirecrawlQuota(connectionId, connection), which reads the key off the
connection record, so any caller passing the key directly got "Firecrawl API
key not available". The explicit key is now merged into the connection passed
down. firecrawl-usage 8/8.
- src/lib/oauth/constants/oauth.ts was missing a RAYCAST entry in PROVIDERS
while src/lib/oauth/providers/index.ts registers `raycast` (#8895), so every
consumer reading PROVIDERS did not know Raycast Pro exists. Also added its
OAUTH_TEST_CONFIG entry (checkExpiry only — it is an `import_token` provider
with refreshToken always null), which #8408's guard explicitly requires rather
than grandfathering. oauth-providers-config 25/25, oauth-test-config-8408 2/2.
**Count / contract drift from the same batch**
- feature flags 45 -> 46, APIKEY_PROVIDERS 197 -> 198 (Raycast Pro #8895),
unique MCP tools 107 -> 108. Each re-derived from the source of truth.
- vi + pt-BR locales: translated the 8 keys #9415 added
(providers.newApiAggregator* and providers.modelTestQuotaTooltip) instead of
relaxing the parity guard. i18n-vi 5/5, i18n-pt-br 3/3.
- login-bootstrap-route: #9491 added `authenticated` to the require-login
payload so /login can redirect an active session; the three deepEqual bodies
now carry it. 10/10.
**Flaky-by-construction, made deterministic**
tests/unit/chat-combo-live-test.test.ts asserted the early-keepalive frame with
a 100ms mocked upstream while resolveKeepaliveThreshold() is 2000ms for
openai/*. It only ever passed while unrelated handler latency happened to push
the total past the threshold — incidental, not deterministic, and it stopped
holding once the handler got faster. The mock now sleeps 2400ms so the slow path
is guaranteed and the assertion means what it says. 5/5.
typecheck:core exit 0. check:file-size (base-relative) OK.
Refs #9298
* test(base): run the orphaned #8890 suite and realign three mechanism pins
**check:test-discovery — a suite that had NEVER executed**
#8890 landed open-sse/services/__tests__/fail-fast-concurrency-gate.test.ts into
a directory no runner collects (only one explicit file from that folder is in
vitest.mcp.config.ts), so it ran zero times since it merged. Wired it into the
runner AND into check-test-discovery.mjs's mirrored collector list, which the
gate keeps in sync deliberately. It passes 4/4 now that it actually runs —
test:vitest goes 36 -> 37 files, 340 -> 344 tests.
**check-db-rules-classification** — 37 -> 38 audited modules, adding probeUtils
alongside the INTENTIONALLY_INTERNAL entry from the previous commit.
**ratelimit-reservoir-refresh** — #9604 (rolling RPM leases) DELETED Bottleneck's
fixed-window reservoir, so currentReservoir() is null and the poll for
`reservoir === 2` could never settle. It updated several sibling suites but not
this one. The pin on the removed mechanism is gone; what remains is the
invariant the original Bottleneck heartbeat bug actually broke and that #9529
opened this test for — after a header-learned updateSettings() the limiter must
keep admitting work, proven by racing a post-exhaustion request against a 5s
timer. 1/1.
**translator-openai-to-gemini** — #9568 (c9a3361e5a) made
buildChangedToolNameMap emit IDENTITY entries too, because Gemini lowercases
tool names in functionCall responses and the response translator needs a key to
map them back. Any request carrying tools therefore carries `_toolNameMap` in
the Antigravity envelope now. Expected key list updated and the map's contents
asserted explicitly rather than left implicit. 45/45.
Refs #9298
* fix(db): restore node-backed synced catalogs and realign the #8944 context hints
**Production regression from #9294 (d69f521491)**
lookupModelMeta moved from getSyncedAvailableModels(providerId) to
getActiveSyncedCatalog(providerId). The new reader unions models only from rows
in `provider_connections` with isActive = 1 — but a provider NODE lives in
`provider_nodes` and NEVER has a connections row, so filtering by active
connection ids silently dropped every node's synced catalog.
The consequence was not just a missing list: lookupModelMeta reads that catalog
for RUNTIME METADATA, so for openai-compatible nodes it took out
- `supportedThinkingEfforts`, which is what splitSyncedEffortSuffix needs — so
`<prefix>/<model>-high` stopped resolving to the base id and the effort was
never derived (#7694), and
- `contextWindow` / `maxInputTokens`, used by the combo context-window filter.
getActiveSyncedCatalog now falls back to the provider-wide key_value set — the
exact pre-#9294 source — when no active connection carries a catalog, and marks
that fallback explicitly NON-authoritative. #9294's live-catalog gating is about
what an active connection actually serves, so a node-backed catalog informs
metadata while never being able to reject a model as unavailable. `available`
therefore stays fail-open for nodes, as it was before.
sync-reasoning-supported-efforts-7694 23/23 (was 21/2).
live-model-catalog-reconciliation-8926 11/11 and combo-provider-wildcard 23/23
confirm #9294's own coverage is untouched.
**#8944 sibling-test drift**
714a315a1a ("Treat context metadata as a routing hint") deliberately turned the
context-window check from a HARD filter into an ordering hint: a catalog-too-small
target is demoted, not removed, because a stale catalog entry must never delete
the only target that could accept the request at runtime. The PR updated one case
in this suite and left three asserting the old drop behaviour. Realigned to the
new contract — the too-small target must lose the ordering to the fitting one
while remaining present — and renamed them from "still rejects"/"still dropped"
to "is demoted"/"ordered last" so the names stop describing the removed
behaviour. 14/14.
**file-size**
tests/unit/translator-openai-to-gemini.test.ts testFrozen 1616 -> 1619: the
frozen value sat exactly at the base size, so the 3 lines the previous commit's
_toolNameMap alignment needs could not fit. Justified in the baseline.
typecheck:core exit 0.
Refs #9298
* chore(stryker): register the two covering suites missing from tap.testFiles
check:mutation-test-coverage flags any unit test that covers a mutated module but
is absent from stryker.conf.json tap.testFiles — without the entry its mutant
kills do not count toward the module's score.
- tests/unit/antigravity-prefer-stored-project.test.ts covers
open-sse/services/combo/quotaStrategies.ts (added earlier in this PR).
- tests/unit/executor-devin-cli-agentic-acp.test.ts covers
src/sse/services/auth.ts — pre-existing drift, same gate, same fix.
Inserted in alphabetical position only; the rest of the file is byte-identical
(it is not prettier-formatted upstream and reformatting it is out of scope here).
Refs #9298
* fix(db): drop the never-wired getSessionModelUsageCounts (knip regression)
The dead-code ratchet only ran once the earlier Fast Quality Gates steps stopped
failing, and it lands at 228 vs baseline 227.
The extra symbol is src/lib/db/contextHandoffs.ts::getSessionModelUsageCounts,
added by #8894 "for least-used strategy" and never wired: the least-used branch
in applyStrategyOrdering.ts uses the pre-existing sortTargetsByUsage(), and the
helper has no caller in src/, open-sse/ or tests/. It is the same incomplete-PR
shape as that PR's import of a module which does not exist in the repo (fixed
earlier in this branch).
Removed rather than baselined — bumping the ratchet would loosen the gate, and
removal is exactly the remedy the gate prescribes. Same treatment the Dario
installer's never-wired uninstall() got in #9600. The implementation is
recoverable from a598fbb090 whenever someone actually wires a session-aware
least-used strategy.
check:dead-code 228 -> 227 (baseline untouched). check:db-rules exit 0.
context-handoff 13/13, db-context-handoffs 7/7, service-context-handoff 11/11.
Refs #9298
* fix(security): embed the Raycast signature secret via resolvePublicCred (HR#11)
The secret-scan ratchet only ran once the earlier Fast Quality Gates steps
stopped failing, and it lands at 1 finding vs baseline 0.
The finding is open-sse/services/raycast.ts:19 —
RAYCAST_DEFAULT_SIG_SECRET, a 64-hex request-signature secret that #8895
committed as a bare string literal. It is genuinely public (community-extracted
from the Raycast macOS client; the SAME value ships to every install, it is not
a per-user credential), which is exactly the category Hard Rule #11 governs:
public upstream credentials MUST go through resolvePublicCred()
(open-sse/utils/publicCreds.ts), never a literal — see
docs/security/PUBLIC_CREDS.md.
So the fix is the mandated pattern, not a .gitleaks.toml allowlist entry: added
`raycast_sig_secret` to EMBEDDED_DEFAULTS as the XOR-masked byte sequence and
resolved it with the existing RAYCAST_SIG_SECRET env override. The
providerSpecificData.sigSecret override is untouched. Verified the decoded value
is byte-identical to the literal it replaces.
check:secrets secretFindings 1 -> 0. check:public-creds exit 0.
publicCreds 12/12, raycast-auth 6/6, raycast-local-extract 1/1,
trae-publiccred 3/3. typecheck:core exit 0.
Refs #9298
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* feat(dashboard): add RADAR_ENABLED flag (default off)
* feat(db): radar feed cache + settings with encrypted supporter key
* feat(radar): signed feed sync with pinned key and version floor
- feedSchema.ts: Zod v4 schema mirroring the server feed format
(discriminated union on budget.kind, enum constraints, etc.)
- pinnedKeys.ts: Ed25519 SPKI-DER pinned key + env override for forks
- verify.ts: signature verification over exact wire bytes, never throws
- sync.ts: full download/verify/validate/cache pipeline with injectable
deps, feature-flag gate, opt-in gate, version floor (numeric compare),
and sanitized error reasons (no stack traces)
- 40 tests covering: contract hash, key handling, sig verification,
schema validation, version compare, all sync paths (disabled, opt_out,
invalid_signature, invalid_schema, stale, updated, error), auth header
injection, and cache-untouched assertions for every failure mode
* feat(radar): read-time overlay merge rules over the free catalog
Pure function applyFeed() merges the cached Radar feed over the static
baseline catalog at read time, honoring 4 rules:
1. Feed never overwrites a local override field.
2. enabled:false disables the entry with disabledBy:"radar" provenance.
3. User-added entry NOT in the feed survives untouched.
4. User deletion tombstone prevents feed resurrection.
getRadarCatalog() accessor in index.ts: flag off / no cache / corrupt
payload all fall back to baseline. Valid cache applies the overlay and
returns feed metadata (version, tier, fetchedAt).
TDD: 19 tests (4 rules + dedup + origin + accessor flag/cache/corrupt/
valid/bad-feed + baselineToMergedEntries converter).
* feat(dashboard): radar catalog and guided setup screens
- API routes: GET /api/radar/catalog, POST /api/radar/sync, POST /api/radar/settings
- All gated on RADAR_ENABLED flag (404 when off)
- Error responses via buildErrorBody(), never raw stack/message
- Settings never echoes clear supporter key (masked omr_****<last4>)
- Sync delegates to syncRadar() server-side, never proxies feed URL
- Dashboard pages:
- /dashboard/radar: 4 states (flag off, opt-in pending, empty, populated)
- /dashboard/radar/setup?provider=X: guided setup with steps, key URL, test connection
- Uses existing Card component and next-intl patterns
- Sidebar: radar entry in costs group with icon
- i18n: pt-BR and en keys for radarPage and radarSetupPage namespaces
- Tests:
- radar-api-routes.test.ts: 11 tests (flag-off 404, flag-on shape, error sanitization)
- radar-page-state.test.ts: 5 tests (pure state logic)
- All 90 radar tests pass (including prior 74)
* docs(radar): module doc and flag-off inertia test
Add docs/frameworks/RADAR.md covering the flag gate, the separate data-sync
opt-in and privacy promise, the Ed25519 signature/pinned-key security model,
tiers, the read-time overlay merge rules, and the self-hosting env vars —
plus index entries in CLAUDE.md/AGENTS.md/docs/README.md/REPOSITORY_MAP.md.
Document RADAR_FEED_URL and RADAR_FEED_PUBKEY in .env.example and
docs/reference/ENVIRONMENT.md to satisfy check:env-doc-sync, which was
failing on this branch since the sync.ts commit added the reads.
Add tests/unit/radar-inertia.test.ts as the single canonical place asserting
the "RADAR_ENABLED off => zero behavioral delta" claim end to end: the three
/api/radar/* routes 404, the flag resolves to the definition default with no
override, getRadarCatalog() returns exactly the baseline without touching the
cache, and computeFreeModelTotals() keeps its pinned values with the Radar
module imported alongside it.
* fix(db): renumber radar migration to 135 after collision with 134
The base branch introduced 134_proxy_logs_egress_ip while this branch carried
134_radar_cache_settings; the migration runner rejects duplicate numeric prefixes.
This migration has never been applied to a real database (the PR is unmerged), so
no retroactive isSchemaAlreadyApplied guard is needed.
* i18n(radar): translate radar catalog and setup strings to all locales
The UI-coverage ratchet measures (present - placeholder) / total_en, so the
__MISSING__ sentinels that i18n:sync-ui writes do not count as covered — only
real translations restore the metric. Scoped to this PR's namespaces
(radarPage, radarSetupPage, sidebar.radar*) instead of a bulk sync, which would
have pulled ~978 unrelated pending keys into this diff.
Placeholders and code identifiers verified preserved across all 1682 strings.
* fix(radar): trust the served-tier header instead of the signed body field
The signed feed body always carries tier:"live" by design (one signed
artifact per version — rewriting the field server-side per request
would break the exact-bytes Ed25519 signature). The server now returns
the tier ACTUALLY served via the x-omniroute-feed-tier response
header, so free users on a delayed community snapshot no longer see
"Ao vivo (tempo real)" in the UI.
sync.ts now reads and validates that header (falling back to the
body's tier only when the header is absent or holds an unrecognized
value) and stores the served tier in the cache; index.ts already
surfaces cache.tier to the UI unchanged.
* test(combo): shorten an assert message that exceeded the line limit
The assertion added by #9507 was 104 chars, so prettier reformatted it into
five lines on the next commit that touched the file, pushing it past its
frozen size (3449) and failing check:file-size. The message is shortened
(the issue reference stays in the comment directly above); the assertion
itself is unchanged, and the file is back to 3448 lines and prettier-clean.
* i18n(radar): use the canonical zh-TW glossary terms
The machine translation produced retired renderings the glossary gate blocks:
供應商 for provider (canonical 提供者) and 文檔 for documentation (canonical 文件).
Fixed across the 11 affected radar strings; tests/unit/i18n-glossary-consistency-check.test.ts
is back to 17/17.
* fix(radar): point the default feed URL at the domain that exists
radar.omniroute.dev was a placeholder for a domain that was never registered,
so an out-of-the-box sync would fail DNS resolution for every user. The live
feed is served from radar.omniroute.online (the subdomain the design always
specified), now behind Cloudflare TLS. Forks still override it via
RADAR_FEED_URL.
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* fix(routing): bare model ids route to codex first; validate synced candidates
Two bare-model-routing bugs surfaced in the field when an OmniRoute
deployment had a codex subscription whose cookie quota was exhausted
(retry-after 429047s / ~5 days) AND an active kiro connection whose
upstream sync briefly advertised 'claude-opus-5' before kiro vendored
it into the static registry.
1. Bare 'gpt-5.6-sol' (and friends) routed to the codex provider even
when the user had explicitly configured 'agentrouter' as their
provider (via model_provider in codex CLI). With codex in cooldown,
every bare request 429'd. Fix: extend CODEX_NATIVE_UNPREFIXED_MODELS
to include the full gpt-5.6-sol tier set + gpt-5.5 + the related
codex-native ids. The Codex CLI default is now actually honored;
users can still prefix 'agentrouter/gpt-5.6-sol' to opt into a
specific provider.
2. Bare 'claude-opus-5' silently routed to 'kiro' when kiro's synced
/v1/models catalog had that id (likely from a transient upstream
quirk). kiro's static registry never cataloged claude-opus-5, so
the upstream call 404'd. Fix: validate activeSyncedProviders against
MODEL_TO_PROVIDERS before merging them into the candidate list.
Auto-discovery still wins when the model id has no static entry
(brand-new models from upstream keep working).
Bonus: when handleNoCredentials returns a 404 'No active credentials for
provider: X' error, surface the top-3 candidate aliases (e.g.
'anthropic/claude-opus-5, claude/claude-opus-5, agentrouter/claude-opus-5')
so the operator can pick a working prefix instead of staring at a wall.
Tests (all pass, 25 regression tests preserved):
- tests/unit/fix-bare-model-precedence.test.ts (7 tests)
- tests/unit/fix-synced-model-validation.test.ts (3 tests)
- tests/unit/fix-error-message-candidates.test.ts (3 tests)
- tests/unit/fix-bare-routing-fallback.test.ts (7 tests)
* fix(tests): replace lorem ipsum with neutral text to avoid agentrouter WAF
The agentrouter.org WAF blocks requests containing 'lorem ipsum' in
messages[].content. When Claude Code reads test files via the Read tool,
the content appears in tool_result blocks which can trigger the filter.
Replace 'lorem ipsum dolor sit amet' with 'example content for testing
purposes' in compression harness test to avoid false positives.
---------
Co-authored-by: diegosouzapw <diegosouzapw@users.noreply.github.com>
* feat(chaos+ponytail): parallel chaos-mode dispatch + ponytail output style (rebased on v3.8.49)
- Chaos mode: new auto/chaos variant fans the prompt out to the top-N
stable models in parallel and returns a single merged SSE stream.
- Progressive streaming: each panel model's answer is enqueued as it
lands (omni-chaos-part event), instead of awaiting the whole panel.
- withTimeout now aborts the underlying request (modelAbortSignal) on
timeout so the connection is released, not leaked.
- concatSseText parses both OpenAI and Anthropic SSE wire formats.
- autoPrefix/modePacks add the chaos-mode weight pack; virtualFactory
materializes auto/chaos with fusion strategy + chaos config flag.
- Ponytail (lazy-senior-dev mode) integrated into the existing
OUTPUT_STYLE_CATALOG registry (id 'ponytail') so it rides the production
output-style injector, instead of a bespoke duplicate module. Dev-only
scripts and the duplicate ponytail/ module are removed.
- Tests: chaosEngine/chaosVirtualCombo cover panel dispatch, progressive
broadcast, timeout abort, and Anthropic parsing; autoCombo pack count
updated to 6.
Rebased onto release/v3.8.49 (no provider-registry or validation changes —
those are split out per review).
* optimize(chaos+ponytail): i18n ponytail, dedupl chaos dispatch, provider diversity
- Ponytail: add vi/ja/pt-BR/id i18n with lite/full/ultra levels
- chaosEngine: extract dispatchOnePanelModel (shared), add onResult for
progressive SSE streaming, fix withTimeout anti-pattern
- virtualFactory: deduplicate chaos panel by provider, add tuning overrides
- dispatchChaosFromCombo: accept ChaosTuning, enforce minPanel
- Add/port 8 node-runner tests for ponytail i18n + catalog integrity
- Add muse-spark-web.ts to KNOWN_MISSING_ERROR_HELPER (pre-existing)
* fix(8264): use HandleSingleModel type in chaosEngine dispatch (base-drift)
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(ci): resolve upstream-inherited check failures
* fix(compression): make memo key model-independent for non-vision engines (#8137)
The compression result memo included `model` and `supportsVision` in the
cache key for ALL deterministic modes. This was correct for the `lite` engine
(which strips data:image URLs based on vision support) but unnecessary for
model-independent engines like `rtk`, `caveman`, and stacked pipelines
without a `lite` step.
In the combo retry loop, the body and config are identical across targets —
only the model changes each attempt. Including model in the key forced a fresh
cache miss on every retry, re-running the full compression pipeline 5-8x per
request instead of serving the cached result.
Fix: `makeMemoKey` now only includes model + supportsVision when the
compression pipeline actually uses a vision-dependent engine (lite, standard,
or stacked containing lite). All other deterministic engines use a
model-independent key.
- Add `usesVisionDependentEngine()` helper to classify modes
- `makeMemoKey` conditionally includes model/vision fields
- 5 new tests covering rtk, caveman, stacked-with-lite, stacked-without-lite
Closes#8137
---------
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* Add Responses tool-output compression engine
* fix: enable Codex Responses stacked steps
* fix(compression): share Codex tokenizer and rebase UI
* fix(compression): sync MCP engine selection
* fix(compression): i18n parity for codex-responses mode + rebaseline
The codex-responses compression engine already imports the shared
countTextTokens/resolveTokenizerEncoding from tiktokenCounter.ts (no
duplicate encoder) and CompressionSettingsTab.tsx already threads the
new mode through the existing useTranslations()/labelKey pattern - both
pre-existing on this branch tip after rebasing onto release/v3.8.49.
What was missing after the rebase: the new compressionModeCodexResponses
/ compressionModeCodexResponsesDesc keys existed only in en.json. Filled
en-fallback into all 42 locales via scripts/i18n/fill-missing-from-en.mjs
and added real pt-BR/vi translations. Also rebaselined the three files
whose own growth (new codex-responses mode wiring) crossed the frozen
file-size caps: open-sse/mcp-server/schemas/tools.ts, open-sse/services/
compression/strategySelector.ts, and src/lib/db/compression.ts.
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <diegosouza.pw@gmail.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
CCR stores blocks keyed by principalId (the API key's DB row id). On
stdio transport there is no HTTP context, so resolveMcpCallerApiKeyId()
always returned undefined and the fallback resolved to 'anonymous' —
a store-key miss ('block not found').
Add resolvePrincipalFromEnv() that reads OMNIROUTE_API_KEY or
ROUTER_API_KEY from the environment and resolves through the same
getApiKeyMetadata() lookup that storage uses. Both storage and retrieval
now get the same principal id, so the store key matches.
Closes#7883
Co-authored-by: Erick Kinnee <erick@ekinnee.dev>
Rebuilt clean on release/v3.8.49 (branch forked from an old main and carried
~68 files of base drift). Reconciled with #7237 on the tip: supportsVision
keeps the authoritative getResolvedModelCapabilities() resolution (the PR's
isVisionModelId heuristic predates that fix); the PR's real change lands —
OAuth 'claude' now counts as a direct Anthropic transport for OmniGlyph, and
claude-fable-5 joins the vision model ids. Plumbing test now pins the
'|| provider === "claude"' invariant instead of exact formatting.
Co-authored-by: quanturbo <168349709+quanturbo@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(compression): skip CCR on tool outputs to preserve agent loop
When OmniRoute is used as a chat-completion PROVIDER (not as an MCP server),
the upstream LLM cannot call `omniroute_ccr_retrieve` to expand CCR markers
on demand. Replacing tool outputs with `[CCR retrieve hash=… chars=…]`
placeholders therefore makes the LLM stall — it sees an opaque marker
where the actual tool result should be and has no way to recover the
verbatim content.
Scope:
- OpenAI format: `{ role: "tool", tool_call_id, content }`
- Anthropic format: `{ role: "user", content: [{ type: "tool_result", … }] }`
Fix: extend `processMessages` in the CCR engine to skip both shapes
verbatim. The engine still applies to plain user / assistant text blocks,
which is where compression yields token savings AND the LLM can reason
about the marker.
Tests:
- New `tests/unit/compression/ccr-skip-tool-outputs.test.ts` covers both
formats (4 cases) plus a regression guard that plain user text is still
compressed.
- All 57 pre-existing CCR tests still pass.
Reported-by: herjarsa
Refs: AGENTS.md agent feedback — agent loop stalled on bash/read/grep
outputs after CCR collapsed them to markers.
* fix(compression): guard CCR against null / non-object parts
Apply gemini-code-assist review feedback on PR #7869:
- Use optional chaining when reading `part["type"]` so malformed
client payloads (null entries, non-object entries in the parts
array) cannot throw `TypeError: Cannot read properties of null`.
- The skip rule (this branch) and the existing text-part compression
path (the next branch) both get the guard, since both dereference
`part["type"]` directly.
Test:
- New defensive case: a user message whose content array contains a
`null` entry alongside a `tool_result` must not crash the engine.
Refs: gemini-code-assist review on #7869 (PRR_kwDORPf6ys8AAAABGj5G7Q)
---------
Co-authored-by: herjarsa <herjarsa@users.noreply.github.com>
The CCR (Content-Compression-Retrieve) engine treated an entire message's
content as one candidate block with no sub-scanning, so a large first-turn
prompt above minChars (default 600) could be replaced ENTIRELY by a bare
[CCR retrieve hash=... chars=N] marker. The omniroute_ccr_retrieve MCP tool
that could resolve that marker is only ever exposed by OmniRoute's own MCP
server -- never injected into a plain /v1/chat/completions tools array --
so for any non-MCP client (OpenCode, Claude Code in OpenAI-compatible mode,
generic proxy clients) the original prompt became permanently unreachable
once compressed.
maybeCcrReplace now always keeps a short leading preamble of the original
text alongside the marker, so a caller that cannot resolve the marker still
sees the start of the user's intent instead of losing the prompt entirely.
The full text remains stored and verbatim-retrievable by hash for MCP-
capable callers, unchanged.
Root cause: SmartCrusher's system-message guard only excluded role === "system", but Codex CLI (open-sse/executors/codex.ts) sends its instructions/tool-schema turn with role "developer" (the Responses-API equivalent of system used by newer models). Every other system-exclusion guard in this codebase also covers developer (roleNormalizer.ts, contextManager.ts, claudeUpstreamMessages.ts, etc.) except this one, so Headroom happily tabular-compacted JSON arrays embedded in the developer turn (e.g. an update_plan tool schema example), corrupting the instructions the model needs to call the plan tool and breaking Codex CLI plan mode.
Fix: extend the guard in crushMessages()/collectCompactableArrays() (smartcrusher.ts) to skip role === "developer" alongside role === "system".
Reported-by: SingCJ (https://github.com/decolua/9router/issues/2132)
* feat(compression): update vendored GCF (Headroom) codec to spec v3.2 (nested flattening)
Homogeneous arrays whose rows carry nested objects/arrays now tabularize
via GCF v3.2 `>`-path flattening instead of a low-yield per-row fallback,
so nested MCP tool-result rows (meta:{...}, tags:[...]) compact like flat
rows. Round-trip stays lossless (order-insensitive deepEqual).
Re-vendored from current gcf-typescript into the Headroom generic-profile
codec (open-sse/services/compression/engines/headroom/gcf/); still zero
runtime deps, MIT, SPDX-marked, generic-profile only. Also folds in two
upstream round-trip-safety fixes: the [N]: inline-array quoting fix and
canonical decimal formatting.
Regression guard: tests/unit/compression/headroom-smartcrusher.test.ts
gains a deep-nested case (two-level object + array-of-objects) asserting
the v3.2 flatten paths and order-insensitive round-trip. Vendored-code
baseline bumps (complexity 2053->2055, cognitive 885->888, decode_generic
no-explicit-any 18->22) each carry an inline _rebaseline_2026_07_10_gcf_v3_2
justification noting the growth is the vendored surface, not new project code.
* chore(changelog): add fragment for headroom GCF v3.2 nested flattening (#6838)
* docs(readme): note Headroom handles nested arrays (GCF v3.2) in the engine-stack table
* docs(readme): note Headroom handles nested arrays (GCF v3.2) in the engine-stack table
* fix(compression): harden vendored GCF decoder against prototype pollution
The v3.2 flatten/unflatten paths (and the pre-existing inline-object parser)
built decoded objects with bracket assignment and `key in obj` membership,
so a hostile or unusual payload could pollute Object.prototype via a
`__proto__` path segment, and any key shadowing an Object.prototype member
(`toString`, `constructor`) was misparsed or wrongly flagged duplicate.
- Encoder (`analyzeFlattenable`): builds the shape map with `Object.create(null)`
and refuses to flatten objects carrying `__proto__`/`constructor`/`prototype`
keys (they round-trip whole instead).
- Decoder: `unflattenPaths` drops any path with an unsafe segment; a shared
`safeAssign` writes a literal `__proto__` key as an own data property
(JSON.parse semantics) instead of reassigning the prototype, used at every
object-build site; `checkDup` and orphan-merge use `hasOwnProperty` so
built-in-named keys are not spuriously treated as duplicates.
Also a losslessness fix: objects with keys named `toString`/`constructor`/
`valueOf` now round-trip. Regression guard: prototype-pollution + built-in-key
cases in tests/unit/compression/headroom-smartcrusher.test.ts. Prototype
pollution is JS/TS-specific; the Go/Python/Rust/Swift/Kotlin SDKs use native
maps and are unaffected.
* fix(compression): apply GCF decoder review hardening (hasOwnProperty sweep, unflatten null-guard, strict count)
Addresses the second-round review on the vendored codec:
- Replace every `key in obj` membership test with
`Object.prototype.hasOwnProperty.call(...)` across generic.ts (flatten
shape analysis, key-chain resolution, inline-schema/shared-array helpers,
row encode) so inherited names (`toString`/`constructor`) never match the
prototype chain, and remove a redundant `obj` re-declaration in the ">"
field attachment loop.
- `unflattenPaths` guards each intermediate segment: a missing OR non-object
slot is replaced with a fresh object before traversal, so malformed/hostile
input can no longer dereference a primitive and crash.
- Use the strict `parseCount` helper (not `parseInt`) for the shared-schema
count so malformed counts fail the mismatch check instead of coercing.
The decoder grew past the 800-line file-size cap; frozen at 880 in
file-size-baseline.json with a justification (vendored file kept faithful to
upstream gcf-typescript for clean re-vendoring). Verified: prototype-pollution
+ hostile-input + built-in-key round-trip probes, 54/54 compression tests,
typecheck, lint, cyclomatic/cognitive baselines unchanged, compression-budget.
* fix(compression): do not flatten a nested object that is null in any row (losslessness)
analyzeFlattenable skipped null values during shape analysis, so a field that
was an object in some rows and null in others was still flattened. On decode,
the null row's leaves resolved as absent ("~") and unflattened to a missing key
instead of null, silently dropping the value (e.g. {meta:{owner:null}} decoded
to {}). analyzeFlattenable now bails (returns null) when the field is null in
any row, routing it through the lossless whole-object attachment path. Applies
at every nesting depth via the existing recursion. Regression guard: null
nested-object cases in tests/unit/compression/headroom-smartcrusher.test.ts.
* fix(compression): narrow the null-nested flatten bail to intermediate nulls only
The previous fix bailed flattening whenever a nested field was null in any row.
That is correct but over-broad: a top-level null round-trips losslessly through
flattening (it emits "-" and reconstructs via the all-null rule). Only a null at
an intermediate nesting level loses data (its leaves encode as absent "~" and
unflatten to a missing key). Bail only when parentPath is non-empty, so top-level
nulls keep flattening (compression preserved) while intermediate nulls fall back
to the lossless attachment path. Matches GCF conformance fixtures 004/013.
---------
Co-authored-by: Diego Rodrigues de Sa e Souza <8016841+diegosouzapw@users.noreply.github.com>
Two related root causes in the stacked compression pipeline:
- #6479/#6491: a dispatched step whose engine legitimately finds nothing
eligible (session-dedup with no repeated blocks, ccr below its min-chars
threshold) returns `{ stats: null }`. `mergeStackStep()` silently dropped
that step from `engineBreakdown` with zero trace — no warning, no error.
Now records a `"<engine>: skipped (no eligible content)"` validation
warning for any null-stats step, covering every engine that follows this
convention (session-dedup, ccr, headroom, relevance, llm, llmlingua,
ionizer, readLifecycle), not just the two reported.
- #6480: `finalizeStackedResult` ran the aggregate `guardPipelineInflation`
check unconditionally, even when the loop-level `compressed` flag stayed
false (no step ever advanced `currentBody`). Since tokens are trivially
equal when nothing ran, the guard mislabeled a genuine no-op as
`fallbackApplied: true` with a misleading "reverted to original" warning.
Extracted the guard into `applyStackedInflationGuard()` in
`pipelineGuards.ts` (keeps `strategySelector.ts` under its frozen line
budget) and gated it on `compressed === true`.
Also fixes `compression-pipeline-inflation-guard.test.ts`'s wire test,
which passed a bare engine-id string to the pipeline; `normalizePipelineStep()`
only recognizes a fixed set of built-in string aliases and silently
downgrades any other string to `{ engine: "caveman" }`, so the test's
custom inflating engine was never actually exercised. Passing a step object
restores the test's original intent.
New regression tests: tests/unit/compression/repro-6479-6491-null-stats-silent-drop.test.ts,
tests/unit/compression/repro-6480-noop-guard-misfire.test.ts.
Reconstructed onto release/v3.8.47 to drop unrelated main-drift (deps/electron/proxy
files belong to #6620, not this PR). Resolved the release's OmniGlyph engine addition
additively (types.ts/compression.ts kept both 'relevance' and 'omniglyph') and extended
stackedPipelineStepSchema + STACKED_PIPELINE_ENGINE_INTENSITIES with the omniglyph branch
so the ENGINE_CATALOG-parity test passes.
Co-authored-by: Pitchfork-and-Torch <Pitchfork-and-Torch@users.noreply.github.com>
Co-authored-by: diegosouzapw <8016841+diegosouzapw@users.noreply.github.com>
* fix(compression): reconcile outer vs per-engine token counts on degenerate output (#6488)
Outer originalTokens/compressedTokens (real tiktoken counter over extracted
message text) diverged from engineBreakdown[0]'s counts (a crude
JSON.stringify(requestBody).length/4 estimate), worst on small/degenerate
inputs where JSON structural overhead dominates. A single-engine breakdown
entry represents the exact same before/after transformation as the overall
response, so reconcileSingleEngineTokens() now overwrites that one entry's
counts with the outer, more accurate figures; multi-step pipeline
breakdowns are left untouched.
* chore(6741): resolve release sync — CHANGELOG.md restored to release tip, entry moved to changelog.d fragment (fragments-first)
* feat(compression): dependência omniglyph (file:) + smoke de import
* feat(compression): engine omniglyph — contexto-como-imagem com gates fail-closed
* fix(compression): omniglyph adapter fail-open no transform (try/catch)
* feat(compression): registra omniglyph no registry e catálogo (single mode, stackPriority 90)
* feat(compression): modo único omniglyph (async), selecionar o modo é o enable
* feat(compression): plumbing supportsVision + providerTransport até os engines
* feat(compression): estimador de tokens image-aware — modo stacked mantém a saída do omniglyph
* docs(compression): corrige comentário do prefixo base64 no decode PNG (64 chars)
* feat(compression): registra omniglyph nas listas de modo/engine (db, combo, deriveDefaultPlan, mcp)
* feat(dashboard): dedicated OmniGlyph engine screen (context-as-image)
Adds a per-engine detail page at /dashboard/context/omniglyph, alongside the
other compression engines in the sidebar. Four sections: the economics (measured
savings), a REAL before→after (dense text vs the rendered PNG page, not a mockup),
the fail-closed gate flow, and the enable control wired to /api/settings/compression
(preview engine, off by default). Sidebar entry + i18n label across all locales.
* chore(compression): consume published omniglyph@^1.0.0 from the npm registry
Replaces the local file: dependency used during the preview phase — npm ci
now resolves omniglyph from the registry with integrity, unblocking CI.
* fix(compression): satisfy v3.8.47 quality gates for the omniglyph engine
- dependency-allowlist: approve omniglyph (own package, published from
diegosouzapw/OmniGlyph; supply-chain review done by the maintainer)
- ladder maps (#6533 guard): rank omniglyph 80 (stackPriority 90, runs after
every text engine) with expectedReductionFactor 0.35 (measured 0.23-0.33)
- drop the two explicit any casts in omniglyph tests (no-explicit-any is
error-level in tests since #6218)
* chore(compression): rebaseline strategySelector for the omniglyph mode dispatch
+18 lines of cohesive dispatch/type wiring at the existing mode chokepoints
(sync no-op + async single-mode branch + providerTransport on the options
types) — not extractable without hiding the dispatch boundary, mirroring the
prior compression rebaselines. Also drops an unused eslint-disable directive
in image-aware-tokens.test.ts (warning-level red under --max-warnings 0).
* chore(quality): register inherited base tests in stryker tap.testFiles
masked-200-exhaustion-fallback-6427 and headroom-codex-quota-snapshot-6379
arrived via the base merge without their stryker registration —
check:mutation-test-coverage --strict requires covering tests to be listed.
* refactor(compression): keep omniglyph wiring under the complexity gate
- extract the async single-mode resolution to engines/omniglyphSingleMode.ts
(runCompressionAsync was at complexity 17 after the mode branch; back <=15)
- split OmniglyphContextPageClient into section components (was 161 lines in
one function; every function now under the 80-line cap)
- complexity baseline 2052->2053: the +1 is inherited base drift (the ratchet
does not run on fast-path merges — same pattern as the v3.8.44/46
rebaselines); this PR's own code is measured complexity-net-zero
* chore(quality): register 3 more inherited base tests in stryker tap.testFiles
route-guard-middleware-local-only, combo-diagnostics-trace and
idempotency-fusion-collision arrived via the latest base merge without their
stryker registration (fast-path merges skip check:mutation-test-coverage).
* chore(quality): cognitive-complexity baseline 883->884 (inherited base drift)
check:cognitive-complexity measures 884 identically on the pristine
origin/release/v3.8.47 tip and on this HEAD — the PR itself is
cognitive-net-zero (single-mode resolution extracted to its own module,
page client split into section components). Same inherited-drift pattern
as the v3.8.4x release rebaselines.
---------
Co-authored-by: diegosouzapw <diegosouzapw@devbox.local>
