omnius 1.0.849 → 1.0.851

package/docs/getting-started/model-providers.md

@@ -6,6 +6,7 @@ Omnius routes model requests through a provider abstraction instead of treating
 
 Supported endpoint patterns include:
 
+- built-in free inference, available without user credentials
 - local Ollama
 - Omnius-managed Ollama pool
 - vLLM
@@ -20,6 +21,26 @@ Supported endpoint patterns include:
 - sponsor endpoints discovered through Nexus
 - COHERE distributed inference peers
 
+## Built-in Free Models
+
+An installation without a selected profile defaults to **Built-in free**
+inference. `/model` lists discovered free models by name in that category,
+alongside the configured endpoint's models. Current catalog eligibility and
+model capabilities determine the choices; known bundled choices remain visible
+when discovery is offline and are marked as unverified. Existing local and
+keyed-provider profiles are preserved.
+
+Choosing a built-in free model switches its whole transport profile and clears
+unrelated endpoint keys. Model-specific adapters handle Chat or Responses
+framing while retaining the original host validation. Inference diagnostics
+continue to identify the actual provider and model used.
+
+Free services have availability and quota limits. Automatic fallback is enabled
+by default and tries compatible alternatives within bounded deadlines and
+rate-limit cooldowns. It does not guarantee that every request succeeds. Use
+`/fallback manual` to require a choice before moving to an alternative after
+the selected route fails; `/fallback auto` restores automatic selection.
+
 ## Endpoint Selection
 
 Use:
@@ -28,10 +49,16 @@ Use:
 /endpoint
 /endpoint sponsor
 /model
+/setup
+/setup local
 ```
 
 Endpoint history records recently used URLs and auth metadata so users can return to external routers without retyping them.
 
+Setup is optional. Inside `/setup`, use `/free` for credentialless inference or
+`/key <provider>` only after choosing a provider that requires a key. Local
+model setup is explicit and retains its hardware preflight requirements.
+
 For the complete protocol, authentication, and fail-closed behavior, read
 [Bring Your Own Inference](../guides/bring-your-own-inference.md). The important
 distinction is that provider names are not cosmetic:

package/docs/work-orders/WO-built-in-free-model-startup-20261012.md

@@ -0,0 +1,137 @@
+# Built-in free inference and clean Telegram settlement
+
+Status: implemented, tested and committed. Live service limits are recorded
+below rather than claimed as model availability. Publication and bot restart
+remain operator-owned.
+
+## Ordered work
+
+1. Repair the public independent-review request. Real installed-bundle probes
+   reproduced sequential HTTP 400 rejections for `uniqueItems`, then top-level
+   `anyOf`. Keep the full host schema and public semantic-review gate. Permit a
+   bounded repair for each newly observed supported limitation on the same route.
+2. Exercise other eligible free models with real synthetic generations through
+   the actual production request builders/parsers. Cover public review as well
+   as simple replies. Record actual failures/quotas; never claim availability
+   from mocked inference or promise that external services never fail.
+3. Include the free models in the default model-selection list without requiring
+   credentials or endpoint configuration. Use model names and neutral free-model
+   presentation in the picker. Retain truthful transport identity in internal
+   diagnostics/provenance. Preserve existing configured model choices.
+4. Update first startup and lightweight setup to offer a working built-in free
+   route before asking for credentials or downloading/loading local models.
+   Account-based providers and local inference remain selectable options.
+5. Repair the live TUI's premature/duplicate completed states and later failed
+   state shown after a healthy `pong`; inspect the separate public conversation
+   failure and stalled continuation path. Separate inference-phase completion
+   from task settlement and preserve actual Telegram acknowledgement authority.
+6. Rebuild, run meaningful regression tests and live remote acceptance, verify
+   the packaged artifact, then make scoped commits and push the tracked branch.
+   Publication/restart remains operator-owned.
+
+## Current live anchors
+
+The running bot uses 1.0.850, bundle SHA256
+`d8636120a09b01e5510cb048c5cafcbb0598474653831174196e6e708de3c33e`.
+Read-only Telegram history confirms private `pong` (message 3368) and public
+`pong` (message 5805). Public generation succeeded; the independent
+`chat-fast-path-review` request then exercised an additional schema shape that
+the earlier attention/simple-reply live harness did not cover.
+
+Private admin input can use its explicitly admitted single-pass path and avoid
+that independent public review. Do not remove the public review to copy the
+private behavior. The earlier schema-only `review` probe used the producer’s nominal 1,300-token
+limit. The complete Telegram method applies its observable-request budget
+(16,384 tokens in these remote probes), so the new all-model harness invokes
+the actual attention, generation, independent-review methods and parsers.
+
+
+## Implementation and regression verification, 2026-10-13
+
+- Public and private generation share the same provider adapters. Public review
+  retains its independent original host schema and semantic acceptance gate.
+- Every remote JSON phase receives its full producer contract in the prompt,
+  including on models with unknown or supported native structured output.
+  Explicitly unsupported native formats use prompt-only generation with original
+  host validation, complete-response buffering and unchanged tool authority.
+- Big Pickle’s exact Chat identity has an observed native-format override:
+  native requests returned HTTP 400 `{"model":"big-pickle"}`; omitting only that
+  format gave HTTP 200 generation and review with host acceptance. Production
+  revalidation then passed attention, generation and review without the probe flag.
+- Cat Jam’s bounded diagnostics learn `uniqueItems` and a root conditional-union
+  restriction. Only the native grammar is lowered. Nested closed objects, identity,
+  authorized tools and coupled fulfillment/continuation checks remain unchanged.
+- Exact pinned-gateway HTTP 429 `Endpoint is unavailable` errors retain their
+  HTTP status but are classified as upstream outages, not quota exhaustion.
+  They cool that model without scheduling a quota retry. Genuine rate limits keep
+  their existing cooldown and retry policy.
+- Remote streaming keeps the backend’s 45-second raw-byte idle watchdog and
+  the 120-second aggregate absolute ceiling. Hidden reasoning and buffered JSON
+  no longer hit a second 45-second absolute timer. Shorter caller limits and Stop
+  still apply; unary requests retain their original absolute attempt bound.
+- Built-in selected models enter the same catalog, transport, schema adapters,
+  validation and lifecycle reporting used by fallback. Explicit selection is first;
+  automatic alternatives use the capability ordering. Manual mode asks before
+  alternatives; disabled fallback still permits the selected built-in model.
+- New installations use built-in free inference with no key or local download.
+  `/model`, Telegram’s isolated model picker and completion include the neutral
+  `Built-in free` category. Bundled choices remain visible offline with truthful
+  availability labels; live pricing/retirement still control admission.
+  Existing local/keyed profiles and explicit transport overrides remain intact.
+- Independent regression checks: 504 orchestrator, 254 CLI and 15 provider registry
+  tests passed. These use controlled network fixtures for invariants and are
+  separate from the live model evidence. Full workspace rebuild passed.
+- The compiler-free fresh npm install passed for the packaged candidate, including
+  SQLite and AIWG checks; artifact audit passed and the 566-entry tarball has no
+  sourcemaps. No version was published and the running bot was not restarted.
+
+The existing README generated REST inventory check fails at the unchanged REST
+block (the task’s README diff only updates setup/prerequisites). The independent
+REST source check passes; this pre-existing inventory drift is not included in
+this repair.
+
+Live coverage deliberately sends only invented messages through real Telegram
+methods and providers. It never calls the Telegram Bot API, dispatches a native
+host tool, reads private conversation history, or invokes local inference. Thus
+these receipts establish inference-path behavior, not a new delivery ACK or
+native cross-application action test. Delivery/settlement regressions are in
+`87cecd19`; the installed 1.0.850 bot remains the operator’s running version.
+
+
+## Live packaged results and delivery
+
+Implementation: `b94a6179` public review, `87cecd19` Telegram settlement,
+`273a7bdb` built-in free integration and common adapters, `65a15404` Step’s
+opaque complex-schema compatibility. Commits are delivered to `origin/main`.
+
+The [durable live receipt](evidence/telegram-all-free-live-20261013.json) records
+all nine eligible models, exact bundle/catalog hashes, phase outcomes, safe
+provider diagnostics, original host acceptance, and physical request counts.
+No inference reply or catalog response in these live runs was mocked.
+
+| Model | Live result through Telegram methods | Adapter or remaining limit |
+| --- | --- | --- |
+| Big Pickle | Attention, generation and independent review passed in installed sweep | Exact Chat native-format opt-out; original host contract retained |
+| Cat Jam | All three phases passed in installed sweep | Learned `uniqueItems` and root conditional-union restrictions |
+| Ling 3.1 Flash | Earlier attention/generation passed; final isolated probes returned endpoint-unavailable HTTP 429 | Declared native-format opt-out; complete installed Ling→Muse cascade passed |
+| LongCat 2.5 Preview | All three phases passed in installed sweep | Full original JSON guidance, native Chat transport |
+| MiMo V2.6 Flash | All three phases passed in installed sweep | Full original JSON guidance, native Chat transport |
+| Muse Spark 1.3 | All three phases passed in installed sweep | Responses auto-only tool choice and learned native schema restrictions |
+| Nemotron 3.5 Lightning | All three phases passed in installed sweep; attention took 61.9 seconds under default deadlines | Raw-byte idle watchdog plus aggregate absolute bound; no invented thinking flag |
+| Space Bunny | All three phases passed in installed sweep | Full original JSON guidance fixed the prior JSON-syntax failure |
+| Step 5 Preview | Initial attention passed; patched final installed generation and independent review both passed (HTTP 200/200) | Simple native schema works, complex schema returned opaque HTTP 400; exact Chat prompt-only policy preserves full host validation |
+
+Step’s final attention probe and Ling’s final probes also recorded intermittent
+endpoint-unavailable HTTP 429s. These are external availability limits. The final
+installed Step→Muse public cascade completed with HTTP 429→200→400→400→200:
+Step’s outage, Muse generation, then two bounded review-schema repairs and an
+accepted independent review. No unavailable-model quota retry or fabricated
+answer was used. Both selected-model isolation and real automatic cascade modes
+are available in `scripts/live-telegram-free-models-harness.mjs`.
+
+The final unpublished candidate bundle SHA256 is
+`93c9f8bae03b447f603c3245a796b32b7ee51da12389a99929489dcb2335ca27`.
+Its compiler-free fresh install, SQLite/AIWG checks, 566-entry tarball audit,
+README wiring and sourcemap exclusion passed. The staging version remains
+1.0.850; the operator’s publish workflow must choose the next release version.
+The running production bot was not replaced by these harness installs.

package/docs/work-orders/WO-opencode-production-compatibility-20261012.md

@@ -0,0 +1,90 @@
+# OpenCode production request compatibility
+
+Status: implemented and verified. Telegram presentation is also completed.
+Publication and active-runtime restart remain with the operator. Step's native
+attention-format compatibility remains unverified after a provider limit; the
+working cascade can continue to other compatible models.
+
+## Evidence and cause
+
+The running 1.0.849 bundle matched the prepared artifact. The discrepancy was
+request compatibility and routing, not an old installed bundle. Earlier live
+Nemotron/Kilo probes did not establish compatibility of the newly prioritized
+Muse/Step models with the full Telegram attention schema.
+
+Real synthetic probes found:
+
+- Muse Responses rejects `tool_choice: none`, `required` and named choices;
+  it accepts only `auto`. The compatibility core definitions made the original
+  tool-free request include tools on the wire. Host tool authority must remain
+  unchanged when adapting that wire choice.
+- After the choice repair, Muse rejects `uniqueItems` with the diagnostic
+  “JSON schema keyword `uniqueItems` is not supported for structured outputs”.
+  The existing schema-recovery path did not recognize this HTTP 400 wording.
+- Step returned `pong` for a plain conversation, but the original attention
+  schema returned an opaque upstream HTTP 400. Its cause is still under live
+  investigation; do not attribute it to a particular keyword without evidence.
+- The public native fast path promoted a sealed `conversation` judgment into
+  an action runner. That produced the empty progress panel and false incomplete
+  suffix. Admission must require both action route and action execution class.
+
+## Required work and acceptance
+
+1. Condition the selected model's actual protocol and observed wire limitations.
+   Keep original tool choice, offered names, argument schemas and structured
+   response validation authoritative. Compatibility definitions grant no actions.
+2. Recognize only evidenced HTTP 400 schema limitations and retain bounded
+   same-route recovery. Preserve auth, rate, context, cancellation and ownership
+   behavior; never treat arbitrary upstream messages as capability evidence.
+3. Validate with real OpenCode generations through production builders, parsers,
+   fallback discovery, budgeting and original host-schema checks. No fabricated
+   model replies. The reusable `scripts/live-opencode-fallback-harness.mjs`
+   enters the exact production remote phase without inventing a local result or
+   local failure. This does not test local broker admission or Telegram transport.
+4. Fix public conversation admission and complete the paragraph/Activity
+   presentation work under its separate work order. Preserve final ACK authority.
+5. Rebuild all workspace packages, verify the packaged artifact, retain safe live
+   receipts and commit/push scoped changes.
+
+## Verification
+
+The real remote cascade returned original-schema-valid attention JSON from
+Muse after HTTP 400 -> same-model adapted HTTP 200, then returned `pong` on
+HTTP 200. The first selected-backend probes also returned `pong` from Step.
+Synthetic input only; zero local inference requests, Telegram messages or tool
+executions. Mocked regressions protect invariants and are not availability proof.
+
+The clean rebuild, tarball audit and compiler-free fresh npm installation passed.
+The [installed-bundle live receipt](evidence/opencode-installed-production-stream-live-20261012.json)
+exercises the exact shipped fallback stream phase, actual free/available model
+discovery, production request/parser paths and the original attention schema.
+Both attention and conversation passed with 56,000 bytes of synthetic background:
+Muse returned HTTP 400 -> adapted same-model HTTP 200 for attention, and HTTP 200
+with exactly `pong` for conversation. No model/metadata replies were mocked.
+The installed and prepared bundle SHA256 is
+`d8636120a09b01e5510cb048c5cafcbb0598474653831174196e6e708de3c33e`.
+The packed artifact SHA256 is
+`88e078ab0bfd38895f8d1691060526f5c7bba8b894b874be7b1c0b361a72545b`.
+The validation artifact retains version 1.0.849; the operator's publication
+workflow must still select a new release version.
+
+Step's [bounded diagnostic](evidence/opencode-step-rate-limited-diagnostic-20261012.json)
+stopped after one HTTP 429 (“Endpoint is unavailable”). Its second diagnostic
+was withheld. No evidence supports a static Step schema adaptation, so none was
+invented. This is not a promise that all remote providers remain available.
+
+404 focused inference regression tests and 237 Telegram regression tests passed.
+The latter use explicit mocked transport acknowledgements, not real Telegram
+sends. Inference acceptance uses real providers. Local admission and the active
+bot's restart were deliberately outside this live remote harness.
+
+Reproduce after building and installing the artifact:
+
+```sh
+node scripts/live-opencode-fallback-harness.mjs --run --cascade --stream \
+  --bundle /path/to/installed/omnius/dist/index.js --context-bytes 56000 \
+  --cases attention,conversation --output /tmp/omnius-opencode-live.json
+```
+
+Delivered implementation commits: `7a48a712` (inference compatibility) and
+`82fce9cf` (Telegram narration, activity controls and conversation admission).

package/docs/work-orders/WO-public-review-schema-compatibility-20261012.md

@@ -0,0 +1,31 @@
+# Public reply-review schema compatibility
+
+Status: repaired and verified; broader built-in-model/startup/lifecycle work is
+tracked in `WO-built-in-free-model-startup-20261012.md`.
+
+The live 1.0.850 bot delivered both private and public `pong`. Public generation
+succeeded on Muse, but its separate independent review used a larger schema
+that the private single-pass path did not require. The previous live acceptance
+covered attention and simple replies, leaving this phase untested.
+
+Real probes against the active installed bundle reproduced two sequential
+HTTP 400 errors: unsupported `uniqueItems`, then a forbidden top-level `anyOf`.
+The fallback allowed only one schema-repair retry and therefore moved to Step
+and Mimo. See the safe [before receipt](evidence/public-review-before-20261012.json).
+
+The repair recognizes the exact observed second diagnostic, lowers only the
+root `anyOf` of a complete closed response object, preserves nested alternatives
+and tool schemas, and permits one bounded retry per newly learned supported
+limitation. The original full host schema, identity/evidence checks and semantic
+review remain authoritative; structured output stays buffered until valid.
+
+The [live after receipt](evidence/public-review-after-20261012.json) records
+Muse HTTP 400 -> 400 -> 200 on the same model, with valid public-review output at
+the production 1,300-token limit. No inference or metadata response was mocked;
+synthetic context only, no Telegram sends or tools. The harness now includes
+the production review schema, policy and parser; the full Telegram-method
+multi-model harness is tracked under the broader work order.
+
+187 focused orchestrator regressions and 21 new public-review boundary tests
+passed. Those invariant tests use explicit fixture networking and are separate
+from the live acceptance above. The workspace build passed.

package/docs/work-orders/WO-telegram-progress-presentation-20261012.md

@@ -1,8 +1,8 @@
 # Telegram progress presentation
 
-Status: queued, explicitly deferred by the user until the inference endpoint
-patch is completed and verified. Do not implement this presentation work as part
-of the current inference repair.
+Status: implemented and verified, resumed by the user with instructions to
+complete, commit and push. The evidenced OpenCode compatibility failure is handled in
+[the current repair order](WO-opencode-production-compatibility-20261012.md).
 
 ## Requested behavior
 
@@ -16,13 +16,30 @@ of the current inference repair.
 - Deliver the final answer separately, with completion wording derived from
   actual delivery state rather than an internal handler's acceptance.
 
-## Future verification
+## Implementation and verification
 
-Trace the existing TUI/chat-mode intermediate narration emitter and Telegram
-progress delivery path. Check Telegram's current official API limits, formatting
-and callback behavior when implementing. Verify private/public scope, edit and
+The runner emits finalized main-agent prose as `assistant_text` with source
+`model_visible_text`; the terminal chat/voice path renders that visible text.
+Telegram action runs previously kept it only as a final reply candidate. The
+bridge now publishes prose accompanying a same-response nonterminal work tool
+through its existing bound progress writer. Standalone explicit
+`telegram_progress` remains available. Hidden reasoning, final-only drafts,
+transport contracts and internal guidance do not become progress.
+
+Keep the four latest distinct paragraphs in collapsed public/private summaries.
+Activity opens the chronological details; public Summary returns to prose.
+Admin calls/results use separate escaped HTML code blocks; public detail shows
+safe tool names and phases without leaking private arguments or outputs.
+Keep telemetry and adversarial diagnostics in private archives while filtering
+ordinary sub-agent display callbacks.
+
+Verify private/public scope, edit and
 send limits, interruption/Stop, callback ownership, long tool-output handling,
 and that diagnostics/adversarial guidance remain out of ordinary public chat.
-
-This work order records the request only; the current source changes target
-inference compatibility, safe diagnostics and tool guidance.
+Progress never settles work or substitutes for final provider acknowledgement.
+Verification: 126 focused presentation/routing tests passed; 51 public/native
+fallback/diagnostic tests and 60 private-native tests also passed. These delivery
+tests use explicit mocked Telegram acknowledgements and do not claim real bot
+messages. The actual listener is exercised in `telegram-runner-commentary.test.ts`.
+The full workspace rebuild and package audit passed. Real inference acceptance,
+including production streaming, is recorded under the inference repair order.

package/docs/work-orders/WO-telegram-reply-lifecycle-20261013.md

@@ -0,0 +1,28 @@
+# Telegram reply lifecycle
+
+Status: implemented and verified. The built-in-model/startup and multi-model
+live acceptance work continues under `WO-built-in-free-model-startup-20261012.md`.
+
+Read-only live history and delivery receipts confirm that public ping 5804
+received acknowledged reply 5805. ESP32 input 5806 failed later and has no
+delivery receipt. The display reused one chat view across inputs and reported
+later chat failures through an already-completed attention view, mixing these
+separate outcomes. Generic exhausted-fallback errors also became unclassified
+quick-chat failures.
+
+Reply views are now bound to their exact input. Attention/reply labels identify
+the phase; later errors preserve earlier successful replies and do not overwrite
+completed attention. Terminal callbacks are idempotent. Exhausted remote fallback
+has a typed diagnosis, preserving its nonretryable policy.
+
+Provider acknowledgement is retained immediately. Optional memory/display/notes
+or latency observers cannot turn an acknowledged delivery into failure or cause
+regeneration. Observation failures are recorded as hashes in a local diagnostic
+archive. Durable intake writes and ambiguous send handling retain their existing
+authority and no-replay behavior. Manual fallback wording now distinguishes an
+unavailable local endpoint from a selected remote provider.
+
+53 lifecycle, recovery, telemetry and public routing regression tests passed;
+CLI TypeScript and whitespace checks passed. Transport acknowledgements in these
+tests are fixtures, not real Telegram sends. Live inference coverage is recorded
+separately in the multi-model work order.

package/docs/work-orders/evidence/opencode-installed-production-stream-live-20261012.json

@@ -0,0 +1,218 @@
+{
+  "schema": "omnius.live-opencode-acceptance.v1",
+  "startedAt": "2026-10-13T05:54:44.013Z",
+  "syntheticOnly": true,
+  "mockedReplies": 0,
+  "localRequests": 0,
+  "telegramMessages": 0,
+  "toolsExecuted": 0,
+  "contextBytes": 56000,
+  "stream": true,
+  "mode": "production_remote_cascade",
+  "receipts": [
+    {
+      "model": "cascade",
+      "case": "attention",
+      "apiShape": "chat",
+      "physicalCalls": 2,
+      "wires": [
+        {
+          "model": "muse-spark-1.3-contributor-free",
+          "apiShape": "responses",
+          "stream": true,
+          "toolChoice": "auto",
+          "toolNames": [
+            "read",
+            "shell"
+          ],
+          "schemaSha256": "15d827c6ec46a34d5edfe219dda0c382b8d7a28480d4952b3e0df14c7d48507a",
+          "bodySha256": "9b164756b715a755719057ac9da5bfa256c5efbd7049f60691a200ae36a9585d",
+          "maxOutputTokens": 16384
+        },
+        {
+          "model": "muse-spark-1.3-contributor-free",
+          "apiShape": "responses",
+          "stream": true,
+          "toolChoice": "auto",
+          "toolNames": [
+            "read",
+            "shell"
+          ],
+          "schemaSha256": "bf3674cf984aad367a0684e7ac2ce288ee8f2e9bf3035e9837b13734a6894e90",
+          "bodySha256": "c88ebead202073dc3850f2a66c33b50b31f2a6367109f03baf914b4b7d48401a",
+          "maxOutputTokens": 16384
+        }
+      ],
+      "http": [
+        {
+          "status": 400,
+          "error": {
+            "bodySha256": "ff87146d8064d14b2548e804bb5560f63857ef4d36cf1da734369825bab20656",
+            "type": "invalid_request_error",
+            "code": null,
+            "param": "schema",
+            "message": "JSON schema keyword `uniqueItems` is not supported for structured outputs"
+          }
+        },
+        {
+          "status": 200
+        }
+      ],
+      "routes": [
+        {
+          "location": "remote",
+          "phase": "selected",
+          "model": "muse-spark-1.3-contributor-free",
+          "providerId": "opencode",
+          "backendType": "vllm"
+        },
+        {
+          "location": "remote",
+          "phase": "failed",
+          "model": "muse-spark-1.3-contributor-free",
+          "providerId": "opencode",
+          "backendType": "vllm",
+          "failure": {
+            "category": "request_rejected",
+            "status": 400,
+            "validationReason": "unsupported_schema",
+            "expectedKind": "json_schema",
+            "schemaDigest": "d4a92266cef1561638f557cd9c2566d957e4ce22cc0c2df0d4e1cd67a39a0c1e",
+            "apiShape": "responses"
+          }
+        },
+        {
+          "location": "remote",
+          "phase": "selected",
+          "model": "muse-spark-1.3-contributor-free",
+          "providerId": "opencode",
+          "backendType": "vllm",
+          "schemaAdapter": {
+            "omittedNativeKeywords": [
+              "uniqueItems"
+            ],
+            "hostValidation": "unchanged"
+          }
+        },
+        {
+          "location": "remote",
+          "phase": "succeeded",
+          "model": "muse-spark-1.3-contributor-free",
+          "providerId": "opencode",
+          "backendType": "vllm",
+          "schemaAdapter": {
+            "omittedNativeKeywords": [
+              "uniqueItems"
+            ],
+            "hostValidation": "unchanged"
+          }
+        }
+      ],
+      "finished": true,
+      "selectedRoute": {
+        "providerId": "opencode",
+        "model": "muse-spark-1.3-contributor-free",
+        "kind": "keyless"
+      },
+      "outcome": {
+        "state": "passed",
+        "schemaValid": true,
+        "visibleChars": 355,
+        "toolCalls": 0,
+        "outputSha256": "2b51f68922960d78fc0422ba1aff422d11ff42fe7501788a109dbe00627f0060"
+      },
+      "elapsedMs": 6692
+    },
+    {
+      "model": "cascade",
+      "case": "conversation",
+      "apiShape": "chat",
+      "physicalCalls": 1,
+      "wires": [
+        {
+          "model": "muse-spark-1.3-contributor-free",
+          "apiShape": "responses",
+          "stream": true,
+          "toolChoice": "auto",
+          "toolNames": [
+            "read",
+            "shell"
+          ],
+          "bodySha256": "b3273c4fbea88e3ccdf258ae8742adcb99e2e9834d33e1f9b40f93ba7d570eec",
+          "maxOutputTokens": 16384
+        }
+      ],
+      "http": [
+        {
+          "status": 200
+        }
+      ],
+      "routes": [
+        {
+          "location": "remote",
+          "phase": "selected",
+          "model": "muse-spark-1.3-contributor-free",
+          "providerId": "opencode",
+          "backendType": "vllm",
+          "schemaAdapter": {
+            "omittedNativeKeywords": [
+              "uniqueItems"
+            ],
+            "hostValidation": "unchanged"
+          }
+        },
+        {
+          "location": "remote",
+          "phase": "succeeded",
+          "model": "muse-spark-1.3-contributor-free",
+          "providerId": "opencode",
+          "backendType": "vllm",
+          "schemaAdapter": {
+            "omittedNativeKeywords": [
+              "uniqueItems"
+            ],
+            "hostValidation": "unchanged"
+          }
+        }
+      ],
+      "finished": true,
+      "selectedRoute": {
+        "providerId": "opencode",
+        "model": "muse-spark-1.3-contributor-free",
+        "kind": "keyless"
+      },
+      "outcome": {
+        "state": "passed",
+        "visibleChars": 4,
+        "toolCalls": 0,
+        "outputSha256": "9795c5ff8937f23526ccb207a5684c1fc94a7854e19c021b39d944e51f5baef2"
+      },
+      "elapsedMs": 3088
+    }
+  ],
+  "bundleSha256": "d8636120a09b01e5510cb048c5cafcbb0598474653831174196e6e708de3c33e",
+  "catalogSha256": "24be782386c5cb701cfc0f074ff314f7f50d08fa9b05dc7cafba745772099f3e",
+  "availableSha256": "dbf15b4be0144ca8ad229201e2dfc8c07c35066905877a030ed60389f070f3b3",
+  "selectedModels": [
+    "big-pickle",
+    "cat-jam-free",
+    "ling-3.1-flash-free",
+    "longcat-2.5-preview-free",
+    "mimo-v2.6-flash-free",
+    "muse-spark-1.3-contributor-free",
+    "nemotron-3.5-lightning-free",
+    "space-bunny-free",
+    "step-5-preview-free"
+  ],
+  "bundleVersion": "1.0.849",
+  "sourceModules": {
+    "agenticRunner": "b4a0291effd917bae0fb710b3d0369f3ef0213f678de5b0c3293c4ffc1f10629",
+    "opencode-responses-compatibility": "c99cb4c324a966936050266f205427da3a64b6e7665a2190e166e64906c15afa",
+    "opencode-chat-compatibility": "aeed3f7bc98dc2fa817b3364ef874b95684ddcd21db38b102e9d3e5e491bac94",
+    "opencode-free-models": "d8fbc9a3d30c0ee298647be7f8a3b55402d08e821c3d35a937b1f0ad793eb3c1"
+  },
+  "attentionSchemaSha256": "15d827c6ec46a34d5edfe219dda0c382b8d7a28480d4952b3e0df14c7d48507a",
+  "finishedAt": "2026-10-13T05:54:56.439Z",
+  "physicalCalls": 3,
+  "passed": true
+}