Skip to content

Epic: move inference/ out of the host into tinyinference, tinyagents and tinyjuice #8092

Closed
43 of 43 issues completed
#8198
Closed
Epic: move inference/ out of the host into tinyinference, tinyagents and tinyjuice#8092
43 of 43 issues completed
#8198
Assignees
senamakel
Labels
refactorStructural code change without a primary behavior change.rust-coreCore Rust runtime in src/: CLI, core_server, shared infrastructure.

Description

@senamakel
senamakel
opened on Oct 18, 2026

Summary

crates/openhuman-core/src/inference/ is 13.8k non-test lines (plus 11.7k test lines). Most of it is generic model and provider plumbing that already half-exists in tinyinference (vendored at vendor/tinyagents/vendor/tinyinference) and tinyagents. This issue gives a verdict for every non-test file under inference/, plus three files the goal pulls in:

  • crates/openhuman-tinyhumans/src/inference/ (managed model, 1.3k);
  • agent/tinyagents/routes.rs;
  • config/schema/cloud_providers.rs.

That is 85 files and 15,636 lines in total, audited on main at c1ef2a11d0, after #7929 merged. Format follows #7929.

It supersedes the older #7393, #7394 and #7403 (parent #7390). Those were scoped when inference/ was about 25k lines, and their remaining items are folded into the leaves below.

Ownership rule.

  • tinyinference owns everything about models and providers:
    • building a provider from a spec, and the provider-string grammar and route kinds;
    • the builtin preset catalog, local runtimes, /models catalogs, and context-window resolution;
    • reasoning and temperature, failure classification and demotion, product-neutral failure copy;
    • usage, embeddings (catalog, factory, verification), the OpenAI-compatible server, and STT/TTS transport;
    • the TinyHumans managed model, as a provider like any other.
  • tinyagents owns the harness: workload routing, per-call model resolution and the turn model set, one-shot completions, stop/halt copy, and route enrichment.
  • tinyjuice owns compression and its host-side tool adapters.
  • OpenHuman keeps only:
    • the config schema and the config→spec mapping;
    • the seam implementations: CredentialResolver, ErrorReporter, ModelFactory/ModelBuildGate, CopyProvider, RouteSource, CostEstimator;
    • Privacy Mode, access gates and the egress ledger;
    • the inference.*, embeddings.*, tokenjuice.* and local_ai.* RPC surfaces, and mounting the /v1 router;
    • OpenHuman-branded copy, i18n and bus events.
  • openhuman-tinyhumans keeps only registration, bearer resolution and session-expiry handling.

Invariants.

  • Dependency direction: tinyagents → tinyinference; tinyinference never depends on tinyagents or OpenHuman.
  • RPC namespace/method strings, model-facing tool names, chat_error.* i18n keys and embedding signatures do not change.
  • Tests move with the code. No shims are left behind.

Verdicts

  • DELETE: a shim, re-export, dead code or duplicate. Callers import the real thing directly.
  • UPSTREAM: generic logic with no upstream equivalent yet. It moves to the named crate and module, and the host keeps the named seam.
  • REIMPLEMENT: the host rebuilt something upstream already has. Replace it with the upstream symbol; the listed gaps land upstream first.
  • RELOCATE: product or host code in the wrong host directory or crate.
  • KEEP: genuine wiring or a port implementation.

New modules proposed

  • New crate tinyjuice-host (tinyjuice repo, J1).
    • Contents: the retrieve tool (updated to the shipped juice_retrieve token+range contract), the REPL tool wrappers with QuerySource and the artifact reader, MlHost.Generate tickets, the summary-focus helpers, and the recovery-name tables and pass-through helper.
    • Dependencies: tinyjuice-bus, tinytools, tokio, uuid. OpenHuman links it next to tinyjuice-bus.
    • It replaces the unused vendor/tinyjuice/src/host/ copies, and keeps the bus contract free of runtime dependencies.
  • New module tinyinference-llm/src/providers/tinyhumans/ (U4): TinyHumansModel with a dynamic bearer, the reasoning wire, the usage-envelope projection, discovery, and an on_auth_failure callback.
  • New modules in existing crates:
    • tinyinference-llm: catalog/fetch.rs (U2), failure/{demotion,copy}.rs (U3), providers/overrides.rs (U8, owner decision).
    • tinyinference-embeddings: verify.rs (U9).
    • tinyagents-harness: oneshot.rs (T1), host::{ModelFactory, ModelBuildGate, TurnModelSet} (T2).
    • tinyagents-registry: router::RouteSource (T3).
  • No new host domain. agent_chat moves into the existing src/agent_runs/ (R1), and the cloud-provider config code into src/config/ (R2).

Totals

Verdict Lines Files Share
KEEP 5,347 34 34%
REIMPLEMENT 4,728 20 30%
UPSTREAM 3,328 18 21%
RELOCATE 2,085 9 13%
DELETE 148 4 1%

Projected size. If every row lands:

  • inference/ drops from about 13.8k to about 5.5k lines: RPC controllers, config mapping, seam impls, branded copy, and about 1k of thin residue left by REIMPLEMENT rows.
  • openhuman-tinyhumans/src/inference/ drops from 1.3k to about 0.3k.
  • About 7k lines leave the host. Some move upstream as new code; the rest are deleted in favour of upstream code that already exists.
  • About 2k lines move to agent_runs, config, security and openhuman-tinyhumans.

KEEP files that stay oversized and get split on touch: ops.rs 494, tokenjuice/mod.rs 466, tokenjuice/schemas.rs 438, schemas/catalog.rs 389 (data), host_runtime/schemas.rs 368.

Sub-issues (tracker)

Work order:

  1. Bugs and deletes (OpenHuman only).
  2. Relocations: one destination per PR, OpenHuman only, wire names unchanged.
  3. Upstream work in tinyinference, tinyagents and tinyjuice.
  4. Host adoption: each leaf is blocked on its upstream leaves, bumps the gitlink and deletes the host copy.

Every upstream item follows the same chain: upstream PR → tag/release → OpenHuman gitlink bump (and module digest re-pin for tinyjuice) plus deletion of the host copy, per .claude/skills/migrate-to-submodule.

1. Bugs found in the audit

1. Deletes (OpenHuman only)

2. Relocations (OpenHuman only; wire names unchanged)

3. Upstream: tinyinference

3. Upstream: tinyagents

3. Upstream: tinyjuice

4. Host adoption (each blocked on its upstream leaves)

Bugs found in the audit

Each bug has its own leaf, except one:

Open decisions

  1. f64 vs micro-unit charges (U5). Upstream ChargedAmount is integer micros, while the host persists exact f64 USD. Either add f64 upstream, or migrate stored cost rows.
  2. Provider overrides upstream (U8/H15). Moving inference/registry.rs's id-keyed tables gives tinyinference a host-seam type, which is what ProviderRegistry's "managed gateway should bypass this factory" doc anticipates. KEEP is the alternative.
  3. TinyHumans model in tinyinference (U4). providers/registry.rs:142 says the managed gateway bypasses the factory. Putting TinyHumansModel upstream reads that as "a separate provider module, not a ProviderSpec kind". Confirm that reading. The wire keys openhuman_usage_meta and the openhuman envelope keep their spelling.
  4. tinyjuice-host as a new crate (J1), versus feature-gating tokio/uuid/tinytools into tinyjuice-bus. The contract rules require bus crates to stay sync and free of runtime dependencies, so a new crate is recommended.
  5. Crate-chain seams for R3. Core cannot depend on openhuman-tinyhumans, so moving error_code.rs and tiers.rs out of core needs a registered BackendErrorCodeClassifier slot. The managed-default lookup goes through the existing inference::registry entry.
  6. /v1 role. http/backend.rs resolves every /v1 request as the agentic role, so Privacy Mode and fallback follow that workload. Confirm this is intended before H1 touches it.
  7. Subprocess providers. Claude Code and the Claude Agent SDK stay in tinyagents-harness, because tinyinference cannot depend on tinyagents. tinyinference only parses their route kind (U1), and the harness builds them (T2).

Evaluation

Evaluation A: provider/factory/ (2958 non-test lines)

UPSTREAM 255 (1 files) · KEEP 679 (4 files) · RELOCATE 497 (3 files) · REIMPLEMENT 1527 (6 files)

File Lines Verdict Target Reason
crates/openhuman-core/src/inference/provider/factory.rs 155 KEEP host Module root. Prefix consts and grammar doc duplicate tinyinference_providers::route / tinyinference_local::profile; after U1 only the test-override seam and re-exports remain. Replace the use super::* web with explicit imports.
crates/openhuman-core/src/inference/provider/factory/access_gates.rs 229 KEEP host (split) Privacy LocalOnly, egress emission, session checks are host policy. external_provider_label / "is external" string classification becomes ProviderRoute::kind()/is_external() upstream (U1).
crates/openhuman-core/src/inference/provider/factory/chat_model.rs 439 REIMPLEMENT tinyagents-harness ModelResolver; tinyinference-llm ModelRequestDefaults DefaultTemperatureChatModel re-implements ModelRequestDefaults/ProfileOverrideModel; the managed→SDK→claude-code→local→cloud cascade is duplicated with turn_model.rs. Host keeps probe_inference_readiness and a build-gate callback. Gap: role-keyed model factory (T2).
crates/openhuman-core/src/inference/provider/factory/cloud_slug.rs 419 REIMPLEMENT tinyinference-llm ProviderRegistry::build_with, providers::openai::codex Slug→model fallback, hint:* remap, Codex header/user-agent assembly and custom-CA client are generic. Host keeps the cloud_providers lookup, ephemeral headers, egress, gates. Gaps in U7.
crates/openhuman-core/src/inference/provider/factory/cloud_spec.rs 121 REIMPLEMENT tinyinference-llm ProviderBuildOptions, CredentialResolver CloudBuild mirrors ProviderBuildOptions; per-call closure resolver should be a direct CredentialResolver impl; registry_auth_style → From<AuthStyle>.
crates/openhuman-core/src/inference/provider/factory/credentials.rs 177 KEEP host CredentialResolver impl Auth profiles, OAuth fallback, ephemeral key, legacy api_key. redact_endpoint is a weaker copy of tinyinference_core::sanitize::redact_url (delete in H7).
crates/openhuman-core/src/inference/provider/factory/discovery.rs 98 REIMPLEMENT tinyinference-llm ProviderRegistry::discovery_request Re-derives "can I discover limits" per route and builds a dummy CloudBuild. Gap: discovery_request(ProviderRoute, Option<ProviderSpec>) (U1/U7). Host keeps the managed catalog limits hook.
crates/openhuman-core/src/inference/provider/factory/local_runtime.rs 118 KEEP host Thin over resolve_local_runtime + build_local_runtime_chat_model; gates, egress and LocalRuntimeInputs from config.local_ai.
crates/openhuman-core/src/inference/provider/factory/managed_backend.rs 161 RELOCATE crates/openhuman-tinyhumans/src/inference/ Pinned managed default, alias handling, managed_unavailable_error are TinyHumans-specific. Privacy gate + egress stay as a core wrapper. make_openhuman_backend_model_for_options goes with tools/delegate.rs (senamakel#200).
crates/openhuman-core/src/inference/provider/factory/primary_cloud.rs 187 RELOCATE src/config/ Config-compat (primary_cloud, legacy inference_url, BYOK-incomplete detection); no model logic. redact_inference_url duplicates redact_url (delete). Deprecation timer "keep until v0.66 (review after 2027-01-11)" is preserved.
crates/openhuman-core/src/inference/provider/factory/routing.rs 255 UPSTREAM tinyagents-registry router/ (WorkloadRouter + RouteSource) provider_for_role, default-model route honouring (#6938) and cloud-fallback logging are generic workload routing. Host keeps configured_route_for_role (config field per role) as the RouteSource impl.
crates/openhuman-core/src/inference/provider/factory/subprocess_providers.rs 152 REIMPLEMENT tinyagents-harness providers::{claude_code, claude_agent_sdk} Glue that parses claude-code:/claude_agent_sdk: prefixes. Gap: route kinds (U1) + harness builds them from the route (T2). Host keeps OpenHumanMcpEndpoint, workspace path, gates.
crates/openhuman-core/src/inference/provider/factory/tiers.rs 149 RELOCATE crates/openhuman-tinyhumans/src/inference/tiers.rs Managed tier vocabulary (legacy_tier_role, managed_default_model, oh_tier_supports_vision). hint_role should use upstream ROUTE_*; resolve_model_for_hint/role_for_model_tier go to WorkloadRouter (T3).
crates/openhuman-core/src/inference/provider/factory/turn_model.rs 298 REIMPLEMENT tinyagents-harness host::model_resolver (ModelResolver, ModelResolveRequest) Eight create_turn_chat_model* variants repeat chat_model.rs's cascade with native-tools/egress flags. Collapse into one resolver impl fed by a host model factory (T2).

Evaluation B: provider/ (rest), provider/ops/, config/schema/cloud_providers.rs (2101 non-test lines)

UPSTREAM 923 (6 files) · KEEP 37 (1 files) · DELETE 116 (3 files) · RELOCATE 284 (2 files) · REIMPLEMENT 741 (4 files)

File Lines Verdict Target Reason
crates/openhuman-core/src/inference/provider/mod.rs 37 KEEP host Module list; shrinks as children move.
crates/openhuman-core/src/inference/provider/types.rs 170 UPSTREAM tinyinference-llm usage/, model/ BilledUsage (f64 USD, cost_is_estimate, charge_reported), ChatResponse, ProviderDelta are generic. AGENT_TURN_MAX_OUTPUT_TOKENS → tinyagents-harness turn_policy. Owner note: f64 vs ChargedAmount micros.
crates/openhuman-core/src/inference/provider/model.rs 202 UPSTREAM tinyinference-llm usage/; tinyagents-harness response_to_model_response (incl. p-format text tool-call fallback), billed_usage_report, usage_info_from_response have no host policy. TurnChatModel/TierRoutes/BuiltTurnModels aliases die with T2.
crates/openhuman-core/src/inference/provider/error_classify.rs 41 REIMPLEMENT tinyinference-llm classify_provider_failure; tinyagents-harness retry::FailoverReason Wraps the upstream classifier plus three NonRetryable predicates. Gap: host is_session_expired predicate seam (U3). Only caller triage/evaluator/arm.rs.
crates/openhuman-core/src/inference/provider/error_code.rs 217 RELOCATE crates/openhuman-tinyhumans (backend_error_code.rs) behind a core classifier slot Managed-backend errorCode contract and Sentry-ownership rules. Core callers (core/observability.rs, web_chat/web_errors) cannot depend on tinyhumans, so core keeps a registered BackendErrorCodeClassifier slot (R3).
crates/openhuman-core/src/inference/provider/fallback_diagnostics.rs 169 UPSTREAM tinyagents-registry router/ (role fallback set + diagnostics); copy stays host CLOUD_FALLBACK_ROLES/role_falls_back_to_cloud/override_knob_for_role duplicate routing.rs's role knowledge and belong on WorkloadRoute. The branded "Connections → LLM" strings stay as host copy.
crates/openhuman-core/src/inference/provider/openai_codex.rs 45 REIMPLEMENT tinyinference-llm OpenAiCodexRouting::oauth/standard Hand-builds the routing struct. Gap: resolve_codex_routing(slug, endpoint, bearer_is_oauth, creds) (U7).
crates/openhuman-core/src/inference/provider/openhuman_backend_model.rs 67 RELOCATE crates/openhuman-tinyhumans PROVIDER_LABEL, REASONING_OFF_METADATA_KEY, without_reasoning go with the managed model; is_safe_endpoint_for_managed_bearer duplicates the tinyhumans copy (delete in D1).
crates/openhuman-core/src/inference/provider/ops/mod.rs 34 DELETE — Re-export hub with a stale doc naming a non-existent sanitize module and a test-only re-export shim. Callers import the real modules.
crates/openhuman-core/src/inference/provider/ops/http_error.rs 259 UPSTREAM tinyinference-llm failure/demotion.rs The Sentry demotion ladder (budget, no-model-loaded, custom upstream, access policy, config rejection, context window, quota → demote kind) is generic. Host keeps the ErrorReporter impl (Sentry + SessionExpired/ProviderApiKeyRejected bus publishes); backend polarity guards → tinyhumans.
crates/openhuman-core/src/inference/provider/ops/provider_factory.rs 43 DELETE — list_providers returns a hard-coded one-entry legacy list ("remote providers were removed"); callers channels/routes.rs:13, platform/doctor/core/run.rs:68 switch to configured cloud_providers. ProviderRuntimeOptions only feeds tools/delegate.rs (deleted by senamakel#200).
crates/openhuman-core/src/inference/provider/ops/models.rs 39 DELETE — use super::* glue; doc lists sub-modules that no longer exist; test-only re-exports.
crates/openhuman-core/src/inference/provider/ops/models/catalog_listing.rs 338 REIMPLEMENT tinyinference-llm catalog/ + ProviderRegistry::discovery_request Full /models pipeline (codex query param, OAuth headers, auth-style headers, 404 hint, 200-with-error check). Gap: async catalog::fetch::list_models(client, DiscoveryRequest) (U2). Host keeps CA client + Outcome wrapping. Builds the CA/proxy client twice.
crates/openhuman-core/src/inference/provider/ops/models/local_runtime.rs 58 UPSTREAM tinyinference-local provider.rs Endpoint synthesis for ollama/lmstudio listing over upstream base-url helpers → local_runtime_models_endpoint(slug, base_url).
crates/openhuman-core/src/inference/provider/ops/models/openrouter.rs 65 UPSTREAM tinyinference-llm catalog/ validate_openrouter_api_key ({base}/key + error payload) is generic; host passes the client.
crates/openhuman-core/src/config/schema/cloud_providers.rs 317 REIMPLEMENT tinyinference-providers builtin.rs endpoint_host, host_is_builtin_cloud_provider, endpoint_host_is_chat_completions_only, builtin_cloud_supports_responses_api, is_builtin_cloud_slug, builtin_cloud_provider re-implemented over host list = upstream + OPENHUMAN_PRESET. Gap: builtin registry accepting host presets (U1). Keep AuthStyle, CloudProviderCreds, is_slug_reserved.

Evaluation C: host_runtime/, turn models, reasoning, context window, auth registry, agent/tinyagents/routes.rs (3347 non-test lines)

UPSTREAM 549 (3 files) · KEEP 537 (6 files) · RELOCATE 834 (2 files) · REIMPLEMENT 1427 (5 files)

File Lines Verdict Target Reason
crates/openhuman-core/src/inference/host_runtime/mod.rs 35 KEEP host Module wiring; drop the rpc alias/globs (D1).
crates/openhuman-core/src/inference/host_runtime/core.rs 15 KEEP host Process singleton LocalAiService.
crates/openhuman-core/src/inference/host_runtime/ops.rs 26 KEEP host Re-export hub for ops/; shrinks after R1.
crates/openhuman-core/src/inference/host_runtime/ops/agent_chat.rs 696 RELOCATE src/agent_runs/ Agent-turn entry (OpenHumanSessionHost, attachment staging, seed/resume, RPC envelope) is not local-runtime code, contrary to the host_runtime README. agent_chat_simple (~50 lines) adopts the harness one-shot helper (T1/H10). Oversized: split on move.
crates/openhuman-core/src/inference/host_runtime/ops/complete_once.rs 133 UPSTREAM tinyagents-harness oneshot.rs Stateless call: refuse tools, reject reserved provider_options, require model, build an OpenAI-compatible model, optional budget, invoke. No upstream equivalent. Model build should use ProviderRegistry. Host keeps ModelBudget wiring.
crates/openhuman-core/src/inference/host_runtime/ops/runtime_ops.rs 191 REIMPLEMENT tinyinference-local LocalAiService Repeated bootstrap-if-not-ready and status provider overlay are service concerns (ensure_ready()); local_ai_transcribe_bytes (temp file, extension sanitising) → tinyinference-voice. Host keeps prompt guard + llm_gate.
crates/openhuman-core/src/inference/host_runtime/ops/turn_guards.rs 138 RELOCATE src/agent_runs/ Prompt-injection enforcement, grant_turn_cwd, effective_agent_chat_origin serve agent_chat only; move with it.
crates/openhuman-core/src/inference/host_runtime/schemas.rs 368 KEEP host (split) inference.*/local_ai.* controllers stay; the two agent_chat* controllers move to agent_runs (R1) and AgentChatParams merges with the near-copy there. Wire names unchanged.
crates/openhuman-core/src/inference/host_runtime/service/mod.rs 5 KEEP host Re-export.
crates/openhuman-core/src/inference/host_runtime/service/speech.rs 155 REIMPLEMENT tinyinference-voice piper::synthesize, ExternalSttClient tts hand-copies piper::synthesize; transcribe is read+mime+base64 → provider. Gaps: synthesize_to_file, configurable length scale, transcribe_file (U10). Host keeps effective_stt_provider, mark_*_ready.
crates/openhuman-core/src/inference/turn_models.rs 693 REIMPLEMENT tinyagents-harness ModelResolver/PrebuiltModelResolver, model_registry/ Same managed-vs-BYOK construction three times (primary, per-tier routes, summarizer) plus override injection and attachments wrap. Upstream owns the tier loop over WORKLOAD_ROUTE_TIERS (skip-failed, RouteRecordingModel, ErrorSlotModel) behind a host ModelFactory (T2).
crates/openhuman-core/src/inference/reasoning.rs 222 UPSTREAM tinyinference-llm model/types.rs; tinyagents-harness RunPolicy parse_reasoning_effort (aliases), with_output_room (55% budget, 1024 floor), "disabled means none" are generic; per-thread effort map sits next to RunPolicy.default_reasoning. Host keeps the runtime.reasoning_effort* mapping.
crates/openhuman-core/src/inference/context_window.rs 244 REIMPLEMENT tinyinference-llm ContextWindowResolver, model::discover Thin caller plus generic Ollama limits URL, warn-once static guess, fetcher client builder. Gap: resolve_for_route(provider, model, local_kind) with a fetcher factory (U6). Host keeps config_override and the CA/proxy client callback.
crates/openhuman-core/src/inference/model_context.rs 144 REIMPLEMENT tinyinference-llm context_window_for_model_id; tinyagents-registry pricing::lookup Fallback chain already upstream; tier aliases belong to workload.rs; MODEL_MANAGED_DEFAULT → tinyhumans. model_vision_enabled (config lookup) stays.
crates/openhuman-core/src/inference/auth_error_registry.rs 88 KEEP host Registry/latch already upstream; what remains is the singleton, clock and branded "Connections → API keys → LLM" copy.
crates/openhuman-core/src/agent/tinyagents/routes.rs 194 UPSTREAM tinyagents-harness middleware/library/workload.rs ResolvedRouteMiddleware (emit enrichment, record route, :free pricing, pricing::lookup) is generic apart from a cost-estimator callback and the run-context slot. turn_required_capabilities/route_fallback_policy are one-line wrappers to inline.

Evaluation D: embedding_host/, http/, registry.rs, mod.rs, ops.rs, schemas/ (3592 non-test lines)

UPSTREAM 679 (2 files) · KEEP 2443 (15 files) · RELOCATE 470 (2 files)

File Lines Verdict Target Reason
crates/openhuman-core/src/inference/embedding_host/mod.rs 37 KEEP host Module wiring; drop unused DEFAULT_CLOUD_EMBEDDING_* re-exports (D1).
crates/openhuman-core/src/inference/embedding_host/cloud_adapter.rs 148 RELOCATE src/security/credentials/ (state-dir); tinyinference-embeddings (UnavailableEmbeddingModel) ~90 lines are credential state-dir resolution mirroring config::load; ~50 lines are a generic erroring EmbeddingModel (U9). File name is stale.
crates/openhuman-core/src/inference/embedding_host/factory.rs 182 KEEP host Config → EmbeddingFactorySettings/CredentialResolver/ManagedModelFactory. The managed-factory closure is written twice; custom:<url> rewrite uses the U1 helper.
crates/openhuman-core/src/inference/embedding_host/rpc.rs 82 KEEP host provider_from_config, build_embedder, resolve_api_key (host credential impl).
crates/openhuman-core/src/inference/embedding_host/rpc/api_keys.rs 62 KEEP host Credential-store RPC.
crates/openhuman-core/src/inference/embedding_host/rpc/embed.rs 147 KEEP host embed/test_connection RPC. Custom-vs-factory branch collapses onto test_embedding_connection (U9).
crates/openhuman-core/src/inference/embedding_host/rpc/settings.rs 432 UPSTREAM tinyinference-embeddings verify.rs (new) ~230 lines orchestrate custom-endpoint verification (timeout, classify_embed_probe, served-model guidance, final_probe_dims, signature change) over upstream pieces; effective_embedder_slug is a pure fn. Host keeps config load/save and AuthService lookups.
crates/openhuman-core/src/inference/embedding_host/schemas.rs 293 KEEP host embeddings.* controllers. optional_bool is dead (D1); schema helpers duplicated with inference/schemas.
crates/openhuman-core/src/inference/http/mod.rs 37 KEEP host Mounts the upstream server_http::router.
crates/openhuman-core/src/inference/http/backend.rs 158 KEEP host ChatBackend impl over config. Bare-name→ollama: and route→(id, owned_by) mapping use U1 helpers. Hard-codes the agentic role for every /v1 call (note).
crates/openhuman-core/src/inference/registry.rs 247 UPSTREAM tinyinference-llm providers/overrides.rs (new) Id-keyed chat/embedding/catalog override tables — the "managed gateway bypass" ProviderRegistry's doc expects. Make request types generic over Ctx; host keeps the HookSlot storage. Owner decision (note).
crates/openhuman-core/src/inference/mod.rs 132 KEEP host local_runtime_config and impl LocalModelConfig for Config. pub use ops as rpc alias goes (D1).
crates/openhuman-core/src/inference/ops.rs 494 KEEP host (split) Mostly logged passthroughs. inference_analyze_sentiment prompt → tinyinference_llm::sentiment (U5); hand-rolled local-prefix check re-implements is_local_provider_string (H5); Sentry demotion policy stays. Split into ops/{status,oauth,models,sentiment}.rs.
crates/openhuman-core/src/inference/schemas.rs 159 KEEP host Controller registry (inference.*).
crates/openhuman-core/src/inference/schemas/catalog.rs 389 KEEP host Schema data.
crates/openhuman-core/src/inference/schemas/claude_code_handlers.rs 84 KEEP host Thin RPC over tinyagents_harness::providers::claude_code.
crates/openhuman-core/src/inference/schemas/oauth_handlers.rs 72 KEEP host Thin RPC; flow already on tinyinference_providers::oauth::OAuthFlow.
crates/openhuman-core/src/inference/schemas/prompt_handlers.rs 115 KEEP host Thin RPC handlers.
crates/openhuman-core/src/inference/schemas/settings_handlers.rs 322 RELOCATE src/config/ops/ ~110 lines turn InferenceCloudProviderUpdate into CloudProviderCreds (reserved slugs, legacy type, auth_style parse, defaults): config validation, not RPC. AuthStyle::parse → config::schema::cloud_providers.

Evaluation E: tokenjuice/, failure_copy/ (2319 non-test lines)

UPSTREAM 604 (3 files) · KEEP 1444 (6 files) · DELETE 32 (1 files) · REIMPLEMENT 239 (2 files)

File Lines Verdict Target Reason
crates/openhuman-core/src/inference/tokenjuice/mod.rs 466 KEEP host (split) Host adapter (config → InstallRequest, ModuleClient calls, pass-through, install fingerprint, gates). Recovery name tables and the compaction pass-through helper move to tinyjuice-host (J1).
crates/openhuman-core/src/inference/tokenjuice/config_patch.rs 74 KEEP host tokenjuice.settings_update config patch.
crates/openhuman-core/src/inference/tokenjuice/focus.rs 32 DELETE tinyjuice-host (copy exists at vendor/tinyjuice/src/host/focus.rs) Byte-near-identical to the root-crate copy; OH links only tinyjuice-bus, so both copies live. Consolidate in tinyjuice-host.
crates/openhuman-core/src/inference/tokenjuice/generate.rs 96 REIMPLEMENT vendor/tinyjuice/src/host/generate.rs → tinyjuice-host Identical ticket registry for MlHost.Generate. Host keeps the per-turn PreparedGenerate closure and ML_HOST wiring.
crates/openhuman-core/src/inference/tokenjuice/repl_tools.rs 394 UPSTREAM tinyjuice-host (new crate) ~280 lines generic (ModuleReplTool, QuerySource, error mapping, dispatch) + ~60 lines artifact containment reader. Host keeps repl_tools_for*, dispatch ctx, artifacts dir. Not tinytools-std (deleted by #8088).
crates/openhuman-core/src/inference/tokenjuice/savings.rs 116 KEEP host Attribution model + price lookup is host policy (should read tinyagents-registry pricing).
crates/openhuman-core/src/inference/tokenjuice/schemas.rs 438 KEEP host tokenjuice.* controllers; generate ContentKind lists from the enum.
crates/openhuman-core/src/inference/tokenjuice/tools.rs 143 REIMPLEMENT vendor/tinyjuice/src/host/retrieve_tool.rs (RetrieveToolOutputTool) Upstream retrieve tool + bus declarations are the stale hash/retrieve_tool_output form; OH ships juice_retrieve with token+range. Gap: update bus declarations and the tool to the OH contract (J1).
crates/openhuman-core/src/inference/failure_copy/mod.rs 23 KEEP host Re-export module.
crates/openhuman-core/src/inference/failure_copy/halt.rs 173 UPSTREAM tinyagents-hooks stop/ (copy seam) Retry/no-progress/budget/provider-config halt summaries and terminal_inference_failure_kind are product-neutral bar one settings path. user_actionable_escalation (Composio/Connections) moves to its only caller middleware/repeated_failure.rs.
crates/openhuman-core/src/inference/failure_copy/stop_summary.rs 37 UPSTREAM tinyagents-harness turn_stop.rs English fallback per turn_stop.<class>; upstream owns the class + keys but only status_message(). Add TurnStopKind::user_copy() with a CopyProvider override.
crates/openhuman-core/src/inference/failure_copy/table.rs 327 KEEP host (split) 29 rows class→error_type/source/retryable/copy/i18n key. ~15 neutral provider-failure rows → tinyinference_llm::failure::copy defaults behind FailureCopyProvider (U3); ~14 OpenHuman-branded rows stay.

Evaluation F: crates/openhuman-tinyhumans/src/inference/ (1319 non-test lines)

UPSTREAM 318 (3 files) · KEEP 207 (2 files) · REIMPLEMENT 794 (3 files)

File Lines Verdict Target Reason
crates/openhuman-tinyhumans/src/inference/mod.rs 111 KEEP openhuman-tinyhumans Registration into core's inference::registry; maps Config → ManagedOptions.
crates/openhuman-tinyhumans/src/inference/managed_calls.rs 160 UPSTREAM tinyinference-llm providers/tinyhumans/ ChatModel decoration around OpenAiModel (reasoning/thread_id injection, usage projection, profile, cache identity, failure logging). Only maybe_publish_session_expired is host → on_auth_failure callback.
crates/openhuman-tinyhumans/src/inference/managed_catalog.rs 259 REIMPLEMENT tinyinference-llm catalog/ (+ U2 fetch); tinyhumans-sdk InferenceApi::list_models ~150 lines are a hand-rolled GET with bearer/status/error-payload handling. Host keeps listing_auth, managed_401_means_signed_out, entry().
crates/openhuman-tinyhumans/src/inference/managed_embedding.rs 149 REIMPLEMENT tinyinference-embeddings CloudEmbeddingModel::with_egress_guard Wrapper only forwards and calls the guard. Gap: empty-batch guard behaviour (U9). Bearer resolution must share the chat model's checks (bug).
crates/openhuman-tinyhumans/src/inference/managed_model.rs 386 REIMPLEMENT tinyinference-llm ProviderKind::TinyHumans → providers/tinyhumans/ Core is OpenAiModel::compatible_provider over /openai/v1, rebuilt per call for a fresh JWT. Gaps: dynamic BearerResolver, pooled client builder, readiness probe, empty-model fallback (U4). Host keeps resolve_bearer.
crates/openhuman-tinyhumans/src/inference/managed_reasoning.rs 83 UPSTREAM tinyinference-llm providers/tinyhumans/reasoning.rs Pure ModelRequest shaping (reasoning object, thread_id); the reasoning-off metadata key moves upstream with it.
crates/openhuman-tinyhumans/src/inference/managed_session.rs 96 KEEP openhuman-tinyhumans Session-token policy and SessionExpired publishing: the on_auth_failure impl. Match on a typed flag, not the "OpenHuman" label literal.
crates/openhuman-tinyhumans/src/inference/managed_usage.rs 75 UPSTREAM tinyinference-llm usage/meta.rs Projects the openhuman.{billing,usage} envelope into merge_usage_meta; no host dependency.

Activity

  1. senamakel
    added
    refactorStructural code change without a primary behavior change.
    rust-coreCore Rust runtime in src/: CLI, core_server, shared infrastructure.
    on Oct 18, 2026
  2. senamakel
    added sub-issues on Oct 18, 2026

  3. 23 remaining items

  4. senamakel
    added sub-issues on Oct 18, 2026
  5. senamakel
    self-assigned this
    on Oct 18, 2026

  6. senamakel commented on Oct 18, 2026

    @senamakel
    senamakel
    on Oct 18, 2026
    MemberAuthor

    core-shrink audit: inference rows handed to this epic

    The reworked #7390 audit found these config files that belong to this epic. #7390 files no leaves for them.

    File Lines Verdict Target
    config/schema/cloud_providers.rs 319 REIMPLEMENT `tinyinference_providers::builtin::{endpoint_host, host_is_builtin_cloud_provider, endpoint_host_is_chat_completions_...
    config/schema/types/model_ids.rs 101 RELOCATE crates/openhuman-tinyhumans/src/inference/ (managed model catalog)
    config/schema/ephemeral_route.rs 166 KEEP same
  7. Al629176
    added this to Team Openhumanon Oct 19, 2026
  8. github-project-automation
    moved this to Todo in Team Openhumanon Oct 19, 2026

  9. senamakel commented on Oct 21, 2026

    @senamakel
    senamakel
    on Oct 21, 2026
    MemberAuthor

    Paused handoff — 2026-10-21

    Stopped at the user's explicit request. The native goal is paused, all workers are finished, and no build/test process remains running. 35 of 43 leaves are merged into the fork epic branches. The epic is not complete and no final upstream PR has been opened. No PR was opened or merged after the stop request; the completed, unmerged #264 feature branch was pushed solely to preserve the handoff.

    Resume here: OpenHuman #264

    The change adopts host ModelFactory/ModelBuildGate and native TurnModelSet; removes the eight turn-model constructors, obsolete aliases, generic subprocess assembly, and discovery module; preserves host privacy/session/egress, credentials, explicit context/injection, attachments, and readiness policy. Native model defaults preserve optional profiles and explicit request values. Attachments use the actual resolved route. Budget decoration updates the native resolver/error slots with the wrapped models.

    Owner prerequisites already merged with all nonzero exact-head hosted checks green:

    • TinyInference #39, optional-profile request defaults: merge adc2d0967aaa22d881790af66b94c22b3eda051f; 829 local tests; 6 hosted checks.
    • TinyAgents #157, adopts that pin: merge c29ab0ee8724e1a500f9ad0bb648546b34fd3c34; 17 focused model-factory tests; 10 hosted checks.

    Verified #264 gates:

    Gate Result / qualification
    Default core + CLI all-target Clippy, -D warnings Passed on final 5561a55; receipt 8092-264-oh-default-strict-r7.json
    Explicit no-default core product all-target Clippy Passed on final source; 8092-264-oh-core-product-strict-r3.json
    CLI full product all-target Clippy, including os-keyring Passed on final source; 8092-264-oh-cli-product-strict-r3.json
    Mandatory full workspace product mock inference 613 passed, 0 failed, 0 ignored, including 500 core inference tests; 8092-264-oh-product-inference-r3.json
    Runtime source qualification Full 613 ran at 446821beb418d8fa585b67f08aa3fbd01a837af6. Its only delta to final 5561a55 moves one existing test unchanged into an external sibling module. The moved test passed again on final source (8092-264-cloud-slug-test-r1.json). Do not claim the full 613 executed exactly at final HEAD.
    Profile-less injection regression Genuine failing assertion at 6174b0d in inference R2, then passed after the host attachment wrapper preserved Option<ModelProfile>; focused receipt 8092-264-profileless-attachment-test-r2.json; same assertion retained.
    Workspace cargo check --workspace --all-targets Passed on final source; 8092-264-oh-workspace-alltargets-check-r1.json
    App cargo check --manifest-path crates/openhuman-app/Cargo.toml Passed on final source; 8092-264-oh-app-manifest-check-r1.json
    Focused core product attachment unit tests 67 passed, 0 failed, 0 ignored, final source; 8092-264-oh-attachments-product-r1.json. This is a focused -p openhuman --lib attachments:: lane, separate from the mandatory workspace lane. Global workspace env was unset; fixtures use explicit temporary roots and local Probe/CaptureModel instances, with remote rejection before inference.
    Full pnpm rust:layout Recorded final-source exit 0: layout, 8-manifest crate chain, hosted-free, storage-bypass, storage-owner, schema dump (27 owners); 8092-264-oh-rust-layout-r1.json. Earlier unrecorded layout output had no durable final receipt and was not accepted; this recorded rerun closes that gap.
    Formatting / diff / source scan Passed; deleted constructors/aliases absent throughout owned Rust crates/tests/examples; all recursive pins exact.

    Production privacy → explicit-host session → egress ordering is positively tested without a cfg-test gate bypass. Constructor-only fake-provider fixtures now run under actual Library host scopes; the explicit CLI policy test remains unchanged. Both Some-profile and None-profile injection, independent workloads, DeepSeek route remapping, explicit provider selection, and native budget error slots passed.

    Durable local state and evidence

    Admin root: /home/enamakel/work/openhuman/.git/implementing-an-epic/.

    • Authoritative ledger: tinyhumansai-openhuman-8092.json (43 ordered leaves, merges, prerequisite PRs, gate evidence, paused handoff).
    • Requirement audit: 8092-completion-audit.json (87 rows; final completion remains unproven).
    • Shared worker contract: agent-contract.md; #264 brief: 8092-264-worker-brief.txt.
    • Prepared #264 PR description: 8092-264-oh-pr-body-prepared.md (all required gate results updated; no PR created).
    • Root verification files under evidence/: 8092-264-parent-final-strict-verification.json, 8092-264-parent-runtime-source-and-profileless-regression-verification.json, 8092-264-parent-workspace-check-verification.json, 8092-264-parent-app-check-verification.json, 8092-264-parent-final-remaining-gates-verification.json, 8092-264-parent-final-fmt-diff-pins-verification.json, 8092-264-parent-row-ownership-source-scan.json.
    • Every recorded gate has a sibling raw .log, actual terminal exit, source identity, controls, and SHA-256. Earlier failing compiler/runtime attempts remain preserved, not relabeled as passes. Full test-log audits are stored as *-parent-log-audit.json.

    All repositories use fork epic branch epic/8092-inference-out. Last verified tips:

    Repo Fork epic tip
    OpenHuman 7b174a921b094b91f5f331026c6080b4d4a51db0
    TinyInference adc2d0967aaa22d881790af66b94c22b3eda051f
    TinyAgents c29ab0ee8724e1a500f9ad0bb648546b34fd3c34
    TinyJuice 7cdfee4d84e1536fc9980886e9a9973e263562b1 — not yet adopted by OpenHuman

    OpenHuman still pins TinyJuice 7c43eedd7d7917fa41ea1bd488649fe233d00531 until #255. Do not confuse the fork owner tip with the host gitlink.

    Remaining leaves, in execution order

    Finish #264 PR/review/merge first, then dispatch fresh workers for:

    1. senamakel/openhuman#265 — native WorkloadRouter, branded fallback diagnostics, native route middleware / host CostEstimator. Preserve full billed usage, reported/submicro charge precision, free pricing, non-response outcomes, and the routing deprecation timer through the v0.66 review after 2027-01-11. Delete the old routes hub instead of relocating it. Prepared brief: 8092-265-worker-brief-prepared.txt; API research: evidence/8092-265-parent-prepared-api-scope.json. Refresh actual merged source before dispatch.
    2. senamakel/openhuman#251 — native one-shot helpers, remove complete_once body. Native simple_chat does not set model/temperature itself; preserve selected defaults mechanically and avoid double budgeting.
    3. senamakel/openhuman#252 — native stop/halt copy, move only host-specific actionable escalation; retain permission/scope exclusions from missing-connection detection.
    4. senamakel/openhuman#253 — native embedding verification/connection helpers and unavailable placeholder; consolidate managed factory; persisted embedding signatures must remain byte-identical.
    5. senamakel/openhuman#254 — native LocalAiService readiness/status and voice file/byte helpers; retain host prompt/capacity gates, credentials/STT selection, and B2 output-path security checks.
    6. senamakel/openhuman#255 — TinyJuice adoption. Prepared release-readiness evidence: evidence/8092-255-parent-release-readiness-refresh.json. Canonical release/artifact compatibility still needs proof; inspect the existing release before deciding tags/releases are required. No tag/release/workflow dispatch was authorized by the intermediate-fork exception.
    7. senamakel/openhuman#256 — provider registry ownership. Preserve all current ProviderModelHooks metadata and explicit-context-then-process fallback; native ProviderOverrides currently covers build/probe only, so missing owner APIs must land in the owner first. Prepared evidence: evidence/8092-256-parent-prepared-api-scope.json.

    Prepared issue snapshots for #251–254 and shared API notes are in evidence/8092-251-254-parent-prepared-api-scope.json. These are preparation only, not implementation or completion credit.

    Final-union work that must not be waived

    All 43 leaf merges are necessary but insufficient. Before completing the epic:

    • Verify all 87 requirement rows, deletion/residue ownership, actual consumers, and the stated RPC/tool-name/chat_error/embedding-signature invariants against the merged union. A moved/re-exported shim is not completion.
    • Re-run and resolve the broader embed aggregate. Original run: 336 passed / 24 failed; one fixture was fixed, 23 aggregate failures remain unverified. Isolated/focused passes do not close the aggregate. Full names and receipts are in ledger final_union_consumer_gate_obligations.OH268_embed_broader_audit.
    • Resolve the previously failing SaaS/runtime scanners (recorded failures 1/6 and 1/37); do not label them green without current proof.
    • Ownership boundary scanner is still failing: 2 unbaselined findings + 13 stale baseline entries, reproduced on #264 source. Evidence: evidence/8092-264-parent-boundary-audit-review.json. The TinyHumans native .with_thread_id builder is caught by a broad accessor-name rule: investigate a narrow regression-tested distinction from actual task-local functions, not a blanket allowance. TinyAgents hooks reexports native truncate_for_halt; remove forwarding residue through the owner repo and update its tests/imports with an owner PR. Remove obsolete baseline entries only after proving source absence.
    • Broad JSON-RPC run remains unresolved: 101 passed / 1 failed / 7 ignored, receipt evidence/8092-262-oh-consumer-json-usage.json. Failure: json_rpc_cron_manual_worker_survives_caller_cancel_on_published_tinyflows, tests/json_rpc_e2e.rs:3010. caller.abort(); caller.await.unwrap_err() can encounter a request that already returned a queued acknowledgement before callback start. Source supports a fixture timing/semantics investigation, but no baseline runtime or final resolution has been proved. Narrow usage/context passes do not erase this failure.
    • Prove changed-line coverage ≥80% against the latest canonical base for the whole final union. Tools are installed but no final coverage percentage/artifacts are proved. Use the authoritative product coverage lane, scripts/ci/self-hosted/diff-cover.sh, and scripts/ci/assert-coverage-presence.sh; zero measured executable lines and weakened exclusions are not proof. Preflight: evidence/8092-final-union-coverage-toolchain-preflight.json.
    • Fetch/merge latest canonical branches normally (no rebase/squash/reset), components first, then host gitlinks; run final union gates on the resulting tips.
    • Open final canonical upstream PRs for every changed component, then OpenHuman. If component PRs are unmerged, the final OpenHuman PR must remain draft with explicit must-land-first links. Do not merge final upstream PRs under this skill.

    Workflow constraints retained for the next agent

    Use $implementing-an-epic. The user explicitly approved intermediate fork PRs, followed by final canonical upstream PRs, overriding the local canonical-only intermediate convention; do not ask again. Use fresh Luna workers per leaf; root orchestrates/reviews/merges rather than implements leaf code. Preserve all checkpoint commits, never squash/rebase/reset/amend.

    Create each next worktree at the superproject level using worktree <slug> after checking/fetching latest canonical and preserving primary checkout commits/edits. No nested submodule worktrees. Every linked-worktree command starts cd -- "<absolute-worktree>" &&. Primary checkout has preserved tree-identical checkpoint commits and submodule dirt; inspect rather than reset them.

    Run long gates through run-recorded-gate.py (already wraps ci-cancel-aware.sh). Use jobs=1, private SCCACHE_SERVER_UDS=/home/enamakel/work/openhuman/.git/implementing-an-epic/sccache.sock, fsync-probed temp/profile directories outside Git, CARGO_PROFILE_TEST_DEBUG=0, RUST_MIN_STACK=16777216, and no CARGO_TARGET_DIR. Preserve healthy running handles. Shell heredocs on /tmp have failed EDQUOT despite free disk; use apply_patch or properly quoted Python, never delete other goals' preserved evidence/build output.

    Mandatory inference command is pnpm debug rust inference --verbose -- --no-default-features --no-fail-fast, with an actual Cargo --workspace product graph. Do not substitute --lib for that gate. For focused mock cases use positional filters with flags after --; the runner's special --test branch does not forward feature flags as expected. Install frozen pnpm deps and verify ws before new mock worktrees. Record actual test matches and exits, not outer-command success.

    Owner PRs require nonzero all-green exact-head hosted checks and normal pinned merges. OpenHuman fork PRs use verified local gates, normal --merge --admin --match-head-commit, without awaiting hosted CI. Read all channels with pr-comments; persist before each nonempty review round (maximum two), reply to every resolved thread, and never dismiss valid feedback. After merge verify metadata/tree, close the fork sub-issue, update ledger/audit, then dispatch the next leaf.

    Completed leaves and rolled-up PRs

    The table below is the ledger's 35 merged leaves. Additional owner prerequisites (including TinyInference #39 / TinyAgents #157 above) do not add leaf credit. The local ledger retains merge SHAs, per-leaf gate evidence, review rounds, and any unresolved review notes.

    Merged leaf PRs
    senamakel/openhuman#244 senamakel/openhuman/pull/393
    senamakel/openhuman#245 senamakel/openhuman/pull/406
    senamakel/openhuman#246 senamakel/tinyinference/pull/21, senamakel/tinyagents/pull/97
    senamakel/openhuman#247 senamakel/openhuman/pull/426
    senamakel/openhuman#248 senamakel/openhuman/pull/432
    senamakel/openhuman#249 senamakel/openhuman/pull/439
    senamakel/openhuman#250 senamakel/openhuman/pull/454
    senamakel/openhuman#266 senamakel/openhuman/pull/462
    senamakel/openhuman#267 senamakel/openhuman/pull/472
    senamakel/openhuman#268 senamakel/openhuman/pull/482
    senamakel/openhuman#269 senamakel/openhuman/pull/512
    senamakel/tinyinference#5 senamakel/tinyinference/pull/15
    senamakel/tinyinference#6 senamakel/tinyinference/pull/16
    senamakel/tinyinference#7 senamakel/tinyinference/pull/17
    senamakel/tinyinference#8 senamakel/tinyinference/pull/18, senamakel/tinyinference/pull/19
    senamakel/tinyinference#9 senamakel/tinyinference/pull/20
    senamakel/tinyinference#10 senamakel/tinyinference/pull/22
    senamakel/tinyinference#11 senamakel/tinyinference/pull/23
    senamakel/tinyinference#12 senamakel/tinyinference/pull/24
    senamakel/tinyinference#13 senamakel/tinyinference/pull/25
    senamakel/tinyinference#4 senamakel/tinyinference/pull/26
    senamakel/tinyagents#72 senamakel/tinyagents/pull/99
    senamakel/tinyagents#73 senamakel/tinyagents/pull/100
    senamakel/tinyagents#74 senamakel/tinyagents/pull/101
    senamakel/tinyagents#75 senamakel/tinyagents/pull/102
    senamakel/tinyagents#76 senamakel/tinyagents/pull/103
    senamakel/tinyagents#77 senamakel/tinyagents/pull/104
    senamakel/tinyjuice#6 senamakel/tinyjuice/pull/10
    senamakel/openhuman#257 senamakel/openhuman/pull/525
    senamakel/openhuman#258 senamakel/tinyinference/pull/28, senamakel/openhuman/pull/531
    senamakel/openhuman#259 senamakel/openhuman/pull/537
    senamakel/openhuman#260 senamakel/openhuman/pull/549
    senamakel/openhuman#261 senamakel/openhuman/pull/551
    senamakel/openhuman#262 senamakel/openhuman/pull/555
    senamakel/openhuman#263 senamakel/openhuman/pull/557
    23 unresolved broader embed aggregate cases
    • a_harness_runs_a_turn_against_the_provider_it_was_given
    • a_profiles_cron_run_waits_for_its_chat_turn_but_other_profiles_do_not
    • a_telegram_message_is_answered_by_the_bound_runtime_agent
    • a_tool_loop_ends_in_a_structured_answer
    • attaching_before_first_turn_preserves_and_executes_native_tools
    • background_services_drive_the_scheduler_and_stop_with_the_runtime
    • chat_and_relay_use_only_selected_tools_inside_their_profile
    • cwd_and_identity_follow_the_turn_and_then_return_to_the_agent_defaults
    • deferred_services_wait_for_the_host_so_a_due_job_finds_its_handler
    • due_jobs_run_once_as_their_own_profile_even_when_the_profile_is_closed
    • empty_truncated_terminal_answers_use_only_the_explicit_repair_allowance
    • host_only_forces_read_only_access_and_sandbox
    • host_only_refuses_builtin_tools_the_model_names
    • one_runtime_hosts_independently_configured_agents
    • operator_named_cron_and_time_tools_run_on_a_direct_provider_within_the_iteration_cap
    • operator_system_prompt_is_the_only_system_prompt_a_profile_sees
    • profile_uses_operator_direct_chat_and_search_with_no_managed_traffic
    • profiles_are_isolated_held_and_relayed
    • refused_tool_call_does_not_mark_an_earlier_turns_result_with_the_same_call_id_as_failed
    • shared_budget_stops_the_tool_loop_before_its_next_provider_call
    • successful_turn_preserves_reported_charges_and_invalid_buyer_cost_is_unknown
    • terminal_schema_validation_retries_without_accepting_a_wrong_type
    • untrusted_input_passes_the_prompt_guard_only_on_a_host_only_agent
  10. github-project-automation
    moved this from Todo to In progress in Team Openhumanon Oct 22, 2026
  11. senamakel
    closed this as completedin #8198on Oct 22, 2026
  12. github-project-automation
    moved this from In progress to Done in Team Openhumanon Oct 22, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

refactorStructural code change without a primary behavior change.rust-coreCore Rust runtime in src/: CLI, core_server, shared infrastructure.

Type

No type

Projects

Milestone

No milestone

Relationships

None yet

Development

No branches or pull requests

Issue actions