omnius 1.0.675 → 1.0.677

package/docs/ADJUDICATION.md

@@ -15,6 +15,46 @@ The host validates all identifiers, citations, panel results, and the verdict.
 The tool holds the case when it cannot validate the required quorum or final
 contract. It does not guess a result.
 
+## Clean-context purpose
+
+The main agent context can contain old plans, persuasive language, stale
+assumptions, and momentum from work that happened before the current decision.
+`adjudicate` does not pass that transcript into the panel. The parent must first
+freeze one explicit admissible record. Each constituent starts in a fresh
+context and receives only its assignment, the relevant record subset, and the
+decision rules. The judge receives the immutable record and validated public
+findings. It does not receive the parent's accumulated narrative.
+
+This boundary reduces anchoring and accidental context contamination. It does
+not guarantee an unbiased decision. Record selection, question framing,
+decision rules, and model behavior can still influence the result. The receipt
+makes those inputs, exclusions, validation failures, findings, and dissent
+auditable instead of hiding them in the parent context.
+
+## Availability
+
+`adjudicate` is an agent-bound tool in each top-level full-agent path:
+
+| Surface | Execution | Progress projection |
+| --- | --- | --- |
+| Interactive `omnius` TUI | Yes | Specialized streamed adjudication block |
+| One-shot `omnius "<task>"` | Yes | Ordinary tool events and final result |
+| `POST /v1/run` | Yes, subject to the selected tool profile | Run events and final result |
+| Default full-agent `POST /v1/chat` | Yes, subject to the selected tool profile | Chat/run events and final result |
+
+The tool is intentionally unavailable to child agents. This prevents nested
+panels, recursive terminal projections, and unbounded inference fan-out.
+Constituents are isolated tools-free inference contexts inside the
+adjudication engine. They are not `full_sub_agent` workers, so an empty
+sub-agent activity list does not indicate that panel fan-out was skipped.
+
+The daemon direct-tool registry, `/v1/chat/completions` bounded mini-loop,
+Telegram runners, and voice-call runners do not expose the native tool. Do not
+infer `GET /v1/tools/adjudicate` or
+`POST /v1/tools/adjudicate/call` from the tool name. A caller-supplied OpenAI
+tool with the same name is not the native Omnius implementation. Use
+`omnius show tool.adjudicate` for the static discovery contract.
+
 ## When to use it
 
 Use `adjudicate` when all of these conditions are true:
@@ -23,7 +63,7 @@ Use `adjudicate` when all of these conditions are true:
 - Two or more outcomes remain materially plausible.
 - The decision affects the next action.
 - The available evidence conflicts or supports different risk tradeoffs.
-- A fresh, impartial context can assess the decision more reliably than the
+- A fresh, evidence-scoped context can assess the decision more reliably than the
   main loop's large working context.
 
 Do not use it for an ordinary implementation choice, a factual lookup, or a
@@ -161,6 +201,11 @@ The host rejects unknown IDs and incomplete evidence coverage. It retries one
 time with validation codes only. The invalid model output is not added to the
 repair prompt.
 
+Planner, constituent, and judge requests use strict provider JSON Schemas.
+When a visible response still fails host validation, the receipt retains
+bounded field-level issue codes, a response hash, and a content-free top-level
+shape summary. It does not retain the rejected response body.
+
 ### 3. Constituent fan-out
 
 The host runs assignments with bounded concurrency. Each constituent gets a

package/docs/DISCOVERY.json

@@ -70,6 +70,11 @@
         "query": "web search tool exposure agent bound",
         "expand": "workflow.agent-bound-tools"
       },
+      {
+        "intent": "Resolve an evidence-based decision impasse",
+        "query": "adjudicate impartial evidence quorum verdict",
+        "expand": "workflow.evidence-bound-adjudication"
+      },
       {
         "intent": "Debug a stale or unhealthy daemon",
         "query": "daemon version port logs health debug",
@@ -36078,6 +36083,125 @@
         "packages/cli/src/tui/omnius-directory.ts"
       ]
     },
+    {
+      "id": "tool.adjudicate",
+      "kind": "tool",
+      "title": "Evidence-bound adjudication",
+      "summary": "Resolve one genuine decision impasse outside the accumulated parent narrative through a frozen evidence record, fresh scoped constituent review, host-validated citations and quorum, a final judge, and a durable verdict receipt.",
+      "aliases": [
+        "adjudicate",
+        "adjudication",
+        "decision impasse",
+        "impartial decision",
+        "evidence-bound decision"
+      ],
+      "keywords": [
+        "agent-bound",
+        "impasse",
+        "evidence",
+        "constituents",
+        "quorum",
+        "judge",
+        "verdict",
+        "dissent"
+      ],
+      "maturity": "stable",
+      "layer": "orchestration",
+      "audiences": [
+        "coding-agent",
+        "maintainer",
+        "interactive-user"
+      ],
+      "direct_callable": false,
+      "use_when": [
+        "A top-level full agent has one exact unresolved decision with at least two materially plausible outcomes",
+        "The admissible evidence supports conflicting conclusions or risk tradeoffs",
+        "The decision changes the next action and benefits from fresh evidence-scoped review that reduces parent-context anchoring",
+        "The caller can make record selection and framing explicit and auditable"
+      ],
+      "avoid_when": [
+        "The question is an ordinary implementation choice, factual lookup, or substitute for gathering missing evidence",
+        "The caller is a child agent, daemon direct-tool client, bounded /v1/chat/completions loop, Telegram runner, or voice-call runner"
+      ],
+      "interfaces": [
+        {
+          "type": "tui-agent-tool",
+          "target": "omnius",
+          "description": "Top-level interactive agent with the specialized streamed adjudication block"
+        },
+        {
+          "type": "cli-agent-tool",
+          "target": "omnius \"<task>\"",
+          "description": "Top-level one-shot agent with ordinary tool events"
+        },
+        {
+          "type": "agent-run",
+          "target": "POST /v1/run",
+          "description": "Spawned top-level full agent; subject to the active tool profile"
+        },
+        {
+          "type": "agent-chat",
+          "target": "POST /v1/chat",
+          "description": "Default full-agent chat mode; subject to the active tool profile"
+        }
+      ],
+      "references": [
+        {
+          "type": "guide",
+          "target": "docs/ADJUDICATION.md",
+          "relation": "canonical-guide"
+        },
+        {
+          "type": "source",
+          "target": "packages/orchestrator/src/adjudication.ts",
+          "relation": "implementation"
+        },
+        {
+          "type": "source",
+          "target": "packages/cli/src/tui/interactive.ts",
+          "relation": "top-level-registration"
+        },
+        {
+          "type": "source",
+          "target": "packages/cli/src/tui/adjudication-live-block.ts",
+          "relation": "live-projection"
+        }
+      ],
+      "related": [
+        "workflow.evidence-bound-adjudication",
+        "layer.orchestration",
+        "layer.observability",
+        "module.orchestrator",
+        "guide.adjudication-uppercase"
+      ],
+      "verification": [
+        {
+          "check": "Run pnpm harness:adjudication",
+          "expected": "Parallel constituent streams, judge synthesis, and a durable receipt complete without loading a real model"
+        },
+        {
+          "check": "Inspect the parent-runner and child-agent tool registration tests",
+          "expected": "Top-level full agents keep adjudicate inline and host scope enforcement excludes it from child runners"
+        }
+      ],
+      "failure_modes": [
+        {
+          "symptom": "A caller attempts /v1/tools/adjudicate/call",
+          "likely_cause": "Static tool discovery was mistaken for daemon direct-tool exposure",
+          "recovery": "Use a top-level TUI, one-shot CLI, /v1/run, or default full-agent /v1/chat path"
+        },
+        {
+          "symptom": "The case returns held instead of a verdict",
+          "likely_cause": "Admission, citation validation, quorum, or final verdict validation failed",
+          "recovery": "Inspect the receipt and provide a corrected admissible record; do not infer a decision"
+        }
+      ],
+      "source_of_truth": [
+        "docs/ADJUDICATION.md",
+        "packages/orchestrator/src/adjudication.ts",
+        "packages/cli/src/tui/interactive.ts"
+      ]
+    },
     {
       "id": "tool.agenda",
       "kind": "tool",
@@ -44015,6 +44139,124 @@
         "packages/execution/src/tools/tool-manifest.ts"
       ]
     },
+    {
+      "id": "workflow.evidence-bound-adjudication",
+      "kind": "workflow",
+      "title": "Resolve an evidence-bound decision impasse",
+      "summary": "Freeze one exact decision record outside the accumulated parent narrative, run fresh evidence-scoped constituent review, validate citations and quorum, synthesize a verdict through a final judge, and preserve an auditable durable receipt without granting the panel tools or mutation authority.",
+      "aliases": [
+        "adjudicate impasse",
+        "impartial review",
+        "decision court",
+        "constituent panel"
+      ],
+      "keywords": [
+        "evidence",
+        "arguments",
+        "quorum",
+        "judge",
+        "verdict",
+        "dissent",
+        "receipt",
+        "anchoring"
+      ],
+      "maturity": "stable",
+      "layer": "orchestration",
+      "audiences": [
+        "coding-agent",
+        "maintainer",
+        "interactive-user"
+      ],
+      "prerequisites": [
+        "Top-level full-agent TUI, one-shot CLI, /v1/run, or default /v1/chat path",
+        "One exact unresolved decision",
+        "At least two allowed outcomes",
+        "Caller-supplied admissible evidence"
+      ],
+      "inputs": [
+        "question",
+        "allowed outcomes",
+        "evidence",
+        "optional arguments, decision rules, burden, constituents, and quorum"
+      ],
+      "outputs": [
+        "validated findings",
+        "verdict or held status",
+        "durable receipt"
+      ],
+      "workflow": [
+        {
+          "step": "1",
+          "action": "Confirm that a genuine impasse exists and submit an immutable record with stable evidence and argument IDs.",
+          "expected": "Host admission succeeds and creates a record hash"
+        },
+        {
+          "step": "2",
+          "action": "Frame or validate distinct constituent assignments with explicit evidence subsets.",
+          "expected": "Every admitted evidence item is covered without unknown IDs"
+        },
+        {
+          "step": "3",
+          "action": "Run tools-free constituent assessments in fresh scoped contexts that do not receive the parent transcript.",
+          "expected": "Public findings cite only assigned evidence"
+        },
+        {
+          "step": "4",
+          "action": "Validate findings and require quorum before invoking the judge.",
+          "expected": "Invalid findings cannot create facts or satisfy quorum"
+        },
+        {
+          "step": "5",
+          "action": "Have the judge apply the burden and rules to the immutable record and validated findings, preserve material dissent, and select an allowed outcome or insufficient_evidence.",
+          "expected": "A host-validated verdict or held case"
+        },
+        {
+          "step": "6",
+          "action": "Inspect the available run events and durable receipt before acting on the decision.",
+          "expected": "The decision and remaining framing risks are traceable to the admitted record, validated findings, and receipt hashes"
+        }
+      ],
+      "avoid_when": [
+        "Using adjudication to replace normal judgment, gather evidence, or let a child agent spawn nested panels",
+        "Treating clean contexts as a guarantee against biased record selection, framing, rules, or model behavior",
+        "Inferring a direct registry, /v1/chat/completions, Telegram, voice, or child-agent exposure that is not registered"
+      ],
+      "verification": [
+        {
+          "check": "Run pnpm harness:adjudication",
+          "expected": "The deterministic panel overlaps constituent work, streams attributed assessments, synthesizes a verdict, and persists a receipt"
+        },
+        {
+          "check": "Inspect .omnius/adjudications/<case-id>/<run>/receipt.json",
+          "expected": "Hashes, quorum, validation failures, elapsed time, and final status are present"
+        }
+      ],
+      "failure_modes": [
+        {
+          "symptom": "The case is held before judgment",
+          "likely_cause": "The record, framing, findings, or quorum failed host validation",
+          "recovery": "Use receipt validation codes to correct the admissible record; do not guess an outcome"
+        },
+        {
+          "symptom": "A direct REST call is attempted",
+          "likely_cause": "Agent-bound full-agent exposure was confused with direct registry exposure",
+          "recovery": "Use top-level TUI, one-shot CLI, /v1/run, or default full-agent /v1/chat"
+        }
+      ],
+      "source_of_truth": [
+        "docs/ADJUDICATION.md",
+        "packages/orchestrator/src/adjudication.ts",
+        "packages/cli/src/tui/interactive.ts",
+        "scripts/adjudication-impasse-harness.mjs"
+      ],
+      "related": [
+        "tool.adjudicate",
+        "layer.orchestration",
+        "layer.observability",
+        "module.orchestrator",
+        "guide.adjudication-uppercase"
+      ]
+    },
     {
       "id": "workflow.extend-omnius",
       "kind": "workflow",

package/docs/DISCOVERY.md

@@ -648,6 +648,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 
 | ID | Title | Summary |
 | --- | --- | --- |
+| `tool.adjudicate` | Evidence-bound adjudication | Resolve one genuine decision impasse outside the accumulated parent narrative through a frozen evidence record, fresh scoped constituent review, host-validated citations and quorum, a final judge, and a durable verdict receipt. |
 | `tool.agenda` | Agenda | agenda is a directly callable Omnius tool. |
 | `tool.agent` | Agent | agent is a directly callable Omnius tool. |
 | `tool.aiwg-health` | Aiwg Health | aiwg_health is a directly callable Omnius tool. |
@@ -793,6 +794,7 @@ Daemon equivalents are `GET /v1/discovery/bootstrap`, `GET /v1/discovery?q=<inte
 | `workflow.daemon-tray-update` | Operate daemon, tray, and updates | Ensure one current daemon owns the service port, start the tray against it, install updates through the real global npm flow, stream progress, restart components, and verify the target runtime. |
 | `workflow.debug-runtime` | Debug an Omnius runtime failure | Diagnose from identity and ownership outward: version, health, port/process, live contract, status/events, state scope, logs/evidence, then the owning module. |
 | `workflow.direct-tool-call` | Call a directly exposed tool | Inspect live metadata, confirm direct-call exposure and safety, submit the exact schema, and verify the tool result. |
+| `workflow.evidence-bound-adjudication` | Resolve an evidence-bound decision impasse | Freeze one exact decision record outside the accumulated parent narrative, run fresh evidence-scoped constituent review, validate citations and quorum, synthesize a verdict through a final judge, and preserve an auditable durable receipt without granting the panel tools or mutation authority. |
 | `workflow.extend-omnius` | Extend or modify Omnius safely | Locate the owning layer/module and canonical registry, change the smallest source boundary, update discovery/docs/contracts, and run targeted plus freshness tests. |
 | `workflow.provider-selection` | Select an inference provider and model | Resolve an explicit provider protocol, credentials, endpoint, and model; verify live reachability and hardware placement for local inference. |
 | `workflow.publish-package` | Build and publish the Omnius package | Follow the repository Minimal Publish SOP: clean all workspaces, rebuild, bundle publish/, inspect a local-cache tarball, patch-bump, publish only from publish/, and verify npm metadata. |

package/docs/agent-memory/INDEX.md

@@ -9,6 +9,10 @@ Omnius documentation is exposed to agents through project-local AIWG-style bundl
 | `omnius-docs` | General Omnius docs entrypoint and feature guides |
 | `omnius-rest-docs` | REST API docs entrypoint and endpoint-family map |
 
+Evidence-bound decision questions route through `omnius-tools-docs` to
+`docs/ADJUDICATION.md`. The catalog entry is `tool.adjudicate`; the operational
+workflow is `workflow.evidence-bound-adjudication`.
+
 ## Agent Use Pattern
 
 1. Run `omnius discover "<need>"` or search `GET /v1/discovery`.

package/docs/discovery/agent-map.json

@@ -29,6 +29,7 @@
       {"intent": "Run a long coding task asynchronously", "query": "long horizon coding REST run poll cancel", "expand": "workflow.async-agent-run"},
       {"intent": "Understand where state is stored", "query": "project global state sessions memory config", "expand": "store.project"},
       {"intent": "Use web search", "query": "web search tool exposure agent bound", "expand": "workflow.agent-bound-tools"},
+      {"intent": "Resolve an evidence-based decision impasse", "query": "adjudicate impartial evidence quorum verdict", "expand": "workflow.evidence-bound-adjudication"},
       {"intent": "Debug a stale or unhealthy daemon", "query": "daemon version port logs health debug", "expand": "workflow.debug-runtime"},
       {"intent": "Add or change Omnius code", "query": "module ownership extend command endpoint tool", "expand": "workflow.extend-omnius"}
     ],
@@ -109,6 +110,7 @@
     {"id":"workflow.stateful-chat","kind":"workflow","title":"Use stateful daemon chat","summary":"Create or select a real chat session, send conversational turns, and load its history without treating control commands such as /quit as chats.","aliases":["chat sessions","history"],"keywords":["session","conversation","history"],"maturity":"stable","layer":"memory","audiences":["integrator","service-agent"],"workflow":[{"step":"1","action":"List or create sessions through the documented chat/session routes.","interface":"GET /v1/chats and chat creation route"},{"step":"2","action":"Send user content through POST /v1/chat with the selected session identity.","interface":"POST /v1/chat"},{"step":"3","action":"Load message history when selecting the session and distinguish UI/control events from conversational turns.","expected":"The selected chat displays its saved conversation"}],"verification":[{"check":"Reload the selected session","expected":"History is restored and control-only commands are absent from the chat list"}],"failure_modes":[{"symptom":"Chats named quit or duplicate Last task summaries appear","likely_cause":"Control/task metadata was projected as a chat session","recovery":"Use the canonical session registry and filter non-conversational control records"}],"source_of_truth":["packages/cli/src/api/chat-session.ts","packages/cli/src/api/session-summary.ts","packages/cli/src/api/web-ui.ts"]},
     {"id":"workflow.direct-tool-call","kind":"workflow","title":"Call a directly exposed tool","summary":"Inspect live metadata, confirm direct-call exposure and safety, submit the exact schema, and verify the tool result.","aliases":["tool REST call"],"keywords":["direct_callable","schema"],"maturity":"stable","layer":"execution","audiences":["integrator","service-agent"],"workflow":[{"step":"1","action":"Inspect the tool metadata and direct_callable flag.","interface":"GET /v1/tools/{name}","expected":"A rest-call interface is explicitly present"},{"step":"2","action":"Validate arguments against the returned parameter schema and call the exact route.","interface":"POST /v1/tools/{name}/call"},{"step":"3","action":"Inspect the structured output and any side effects.","expected":"Tool-specific verified result"}],"avoid_when":["The tool is agent-bound, unavailable, profile-gated, or lacks a rest-call interface"],"verification":[{"check":"Metadata, schema, and result all agree","expected":"No inferred route or unvalidated arguments"}],"failure_modes":[{"symptom":"Direct call returns not found or not callable","likely_cause":"The route was inferred or live exposure changed","recovery":"Re-read GET /v1/tools/{name}; use an agent-bound workflow when no rest-call interface exists"}],"source_of_truth":["packages/cli/src/api/direct-tool-registry.ts","packages/execution/src/tools/tool-manifest.ts"]},
     {"id":"workflow.agent-bound-tools","kind":"workflow","title":"Use agent-bound tools such as web_search","summary":"Offer a non-direct tool to an Omnius agent loop through run/chat instead of inventing a direct REST call.","aliases":["web search","daemon tools"],"keywords":["agent loop","web_search","tool exposure"],"maturity":"stable","layer":"execution","audiences":["integrator","coding-agent"],"workflow":[{"step":"1","action":"Inspect live tool metadata, security classification, availability, and schema.","interface":"GET /v1/tools/web_search"},{"step":"2","action":"Offer the tool through an agent-capable surface and a compatible tool profile.","interface":"POST /v1/run or POST /v1/chat/completions with agent_loop=true"},{"step":"3","action":"Require source/provenance verification appropriate to the research task.","expected":"The agent executes the bound tool and returns evidence"}],"avoid_when":["Calling POST /v1/tools/web_search/call unless live metadata explicitly adds direct exposure"],"verification":[{"check":"Inspect run/chat tool events and returned source evidence","expected":"The intended tool actually ran and its claims are traceable"}],"failure_modes":[{"symptom":"Direct tool URL is missing","likely_cause":"The tool is intentionally agent-bound","recovery":"Use /v1/run or agent-loop chat with the tool offered"}],"source_of_truth":["docs/guides/tools-and-web-search.md","packages/execution/src/tools/web-search.ts"]},
+    {"id":"workflow.evidence-bound-adjudication","kind":"workflow","title":"Resolve an evidence-bound decision impasse","summary":"Freeze one exact decision record outside the accumulated parent narrative, run fresh evidence-scoped constituent review, validate citations and quorum, synthesize a verdict through a final judge, and preserve an auditable durable receipt without granting the panel tools or mutation authority.","aliases":["adjudicate impasse","impartial review","decision court","constituent panel"],"keywords":["evidence","arguments","quorum","judge","verdict","dissent","receipt","anchoring"],"maturity":"stable","layer":"orchestration","audiences":["coding-agent","maintainer","interactive-user"],"prerequisites":["Top-level full-agent TUI, one-shot CLI, /v1/run, or default /v1/chat path","One exact unresolved decision","At least two allowed outcomes","Caller-supplied admissible evidence"],"inputs":["question","allowed outcomes","evidence","optional arguments, decision rules, burden, constituents, and quorum"],"outputs":["validated findings","verdict or held status","durable receipt"],"workflow":[{"step":"1","action":"Confirm that a genuine impasse exists and submit an immutable record with stable evidence and argument IDs.","expected":"Host admission succeeds and creates a record hash"},{"step":"2","action":"Frame or validate distinct constituent assignments with explicit evidence subsets.","expected":"Every admitted evidence item is covered without unknown IDs"},{"step":"3","action":"Run tools-free constituent assessments in fresh scoped contexts that do not receive the parent transcript.","expected":"Public findings cite only assigned evidence"},{"step":"4","action":"Validate findings and require quorum before invoking the judge.","expected":"Invalid findings cannot create facts or satisfy quorum"},{"step":"5","action":"Have the judge apply the burden and rules to the immutable record and validated findings, preserve material dissent, and select an allowed outcome or insufficient_evidence.","expected":"A host-validated verdict or held case"},{"step":"6","action":"Inspect the available run events and durable receipt before acting on the decision.","expected":"The decision and remaining framing risks are traceable to the admitted record, validated findings, and receipt hashes"}],"avoid_when":["Using adjudication to replace normal judgment, gather evidence, or let a child agent spawn nested panels","Treating clean contexts as a guarantee against biased record selection, framing, rules, or model behavior","Inferring a direct registry, /v1/chat/completions, Telegram, voice, or child-agent exposure that is not registered"],"verification":[{"check":"Run pnpm harness:adjudication","expected":"The deterministic panel overlaps constituent work, streams attributed assessments, synthesizes a verdict, and persists a receipt"},{"check":"Inspect .omnius/adjudications/<case-id>/<run>/receipt.json","expected":"Hashes, quorum, validation failures, elapsed time, and final status are present"}],"failure_modes":[{"symptom":"The case is held before judgment","likely_cause":"The record, framing, findings, or quorum failed host validation","recovery":"Use receipt validation codes to correct the admissible record; do not guess an outcome"},{"symptom":"A direct REST call is attempted","likely_cause":"Agent-bound full-agent exposure was confused with direct registry exposure","recovery":"Use top-level TUI, one-shot CLI, /v1/run, or default full-agent /v1/chat"}],"source_of_truth":["docs/ADJUDICATION.md","packages/orchestrator/src/adjudication.ts","packages/cli/src/tui/interactive.ts","scripts/adjudication-impasse-harness.mjs"],"related":["tool.adjudicate","layer.orchestration","layer.observability","module.orchestrator","guide.adjudication-uppercase"]},
     {"id":"workflow.provider-selection","kind":"workflow","title":"Select an inference provider and model","summary":"Resolve an explicit provider protocol, credentials, endpoint, and model; verify live reachability and hardware placement for local inference.","aliases":["BYOI","model selection"],"keywords":["provider","endpoint","protocol"],"maturity":"stable","layer":"inference","audiences":["operator","integrator","coding-agent"],"workflow":[{"step":"1","action":"Discover and expand the provider descriptor; do not infer protocol from a label or API key."},{"step":"2","action":"Configure endpoint/protocol/credential using the documented scope."},{"step":"3","action":"For local model work, perform the required hardware preflight before any token-generating request."},{"step":"4","action":"Verify the selected provider and exact model through live metadata."}],"verification":[{"check":"Live model/provider status matches the intended endpoint, protocol, and hardware","expected":"No silent fallback"}],"failure_modes":[{"symptom":"Model listing works but inference fails or uses the wrong protocol","likely_cause":"Endpoint display label was used instead of the provider descriptor","recovery":"Resolve the stable provider ID/protocol and re-test the exact endpoint before execution"}],"source_of_truth":["packages/backend-vllm/src/providerRegistry.ts","docs/guides/bring-your-own-inference.md"]},
     {"id":"workflow.voice-asr-tts","kind":"workflow","title":"Select and use ASR/TTS engines","summary":"Discover installed and supported ASR/TTS systems, perform managed setup when needed, activate one exact engine/model/device, and use the documented REST or TUI surface.","aliases":["speech","voice engines"],"keywords":["ASR","TTS","VibeVoice","transcribe_cli","LuxTTS"],"maturity":"stable","layer":"media","audiences":["integrator","operator","user"],"workflow":[{"step":"1","action":"List engines/models and inspect status before activation.","interface":"GET /v1/asr/engines; GET /v1/asr/status; voice model routes"},{"step":"2","action":"Run explicit managed setup for missing runtimes/weights and an exact accelerator when required."},{"step":"3","action":"Activate the selected engine/model and verify active status."},{"step":"4","action":"Transcribe or synthesize through the OpenAPI-documented route and validate the output artifact."}],"verification":[{"check":"Status reports the requested active engine/model/device and a small non-live test succeeds","expected":"No interpreter override or fallback to a different engine"}],"failure_modes":[{"symptom":"transcribe_cli is missing although a managed environment exists","likely_cause":"TRANSCRIBE_PYTHON points at an older Whisper environment","recovery":"Use the canonical managed transcribe runtime selection and re-check ASR status"}],"source_of_truth":["packages/execution/src/asr/registry.ts","packages/execution/src/transcribe-python-runtime.ts","packages/cli/src/api/voice-runtime.ts"]},
     {"id":"workflow.daemon-tray-update","kind":"workflow","title":"Operate daemon, tray, and updates","summary":"Ensure one current daemon owns the service port, start the tray against it, install updates through the real global npm flow, stream progress, restart components, and verify the target runtime.","aliases":["update Omnius","indicator update"],"keywords":["npm global","restart","version"],"maturity":"stable","layer":"operations","audiences":["operator","coding-agent"],"workflow":[{"step":"1","action":"Read installed and running identities from /version; diagnose port ownership before restart."},{"step":"2","action":"Start/reclaim the daemon through its managed lifecycle and confirm health."},{"step":"3","action":"Start the indicator and require daemon-online state before enabling service actions."},{"step":"4","action":"Run the update service, stream its live progress, restart the daemon/indicator, and compare /version with the target."}],"verification":[{"check":"Installed package, daemon /version, and indicator version all equal the update target","expected":"Verified target runtime, not merely queued or process-started"}],"failure_modes":[{"symptom":"UI remains on updating/queued","likely_cause":"The update worker was never executed or progress was not connected","recovery":"Inspect update job status/log stream and fail explicitly if no worker owns it"},{"symptom":"Restart verification fails","likely_cause":"Old daemon retained port ownership or new runtime did not become ready","recovery":"Resolve exact port PID, preserve unrelated processes, restart, then verify /health and /version"}],"source_of_truth":["packages/cli/src/update-service.ts","packages/cli/src/update-worker.ts","packages/cli/src/daemon.ts","packages/cli/src/tray.ts"]},

package/docs/discovery/catalog-overrides.json

@@ -428,6 +428,50 @@
       ],
       "related": ["api.tools", "tool.tool-search", "operation.version-compatibility"]
     },
+    {
+      "id": "tool.adjudicate",
+      "kind": "tool",
+      "title": "Evidence-bound adjudication",
+      "summary": "Resolve one genuine decision impasse outside the accumulated parent narrative through a frozen evidence record, fresh scoped constituent review, host-validated citations and quorum, a final judge, and a durable verdict receipt.",
+      "aliases": ["adjudicate", "adjudication", "decision impasse", "impartial decision", "evidence-bound decision"],
+      "keywords": ["agent-bound", "impasse", "evidence", "constituents", "quorum", "judge", "verdict", "dissent"],
+      "maturity": "stable",
+      "layer": "orchestration",
+      "audiences": ["coding-agent", "maintainer", "interactive-user"],
+      "direct_callable": false,
+      "use_when": [
+        "A top-level full agent has one exact unresolved decision with at least two materially plausible outcomes",
+        "The admissible evidence supports conflicting conclusions or risk tradeoffs",
+        "The decision changes the next action and benefits from fresh evidence-scoped review that reduces parent-context anchoring",
+        "The caller can make record selection and framing explicit and auditable"
+      ],
+      "avoid_when": [
+        "The question is an ordinary implementation choice, factual lookup, or substitute for gathering missing evidence",
+        "The caller is a child agent, daemon direct-tool client, bounded /v1/chat/completions loop, Telegram runner, or voice-call runner"
+      ],
+      "interfaces": [
+        {"type": "tui-agent-tool", "target": "omnius", "description": "Top-level interactive agent with the specialized streamed adjudication block"},
+        {"type": "cli-agent-tool", "target": "omnius \"<task>\"", "description": "Top-level one-shot agent with ordinary tool events"},
+        {"type": "agent-run", "target": "POST /v1/run", "description": "Spawned top-level full agent; subject to the active tool profile"},
+        {"type": "agent-chat", "target": "POST /v1/chat", "description": "Default full-agent chat mode; subject to the active tool profile"}
+      ],
+      "references": [
+        {"type": "guide", "target": "docs/ADJUDICATION.md", "relation": "canonical-guide"},
+        {"type": "source", "target": "packages/orchestrator/src/adjudication.ts", "relation": "implementation"},
+        {"type": "source", "target": "packages/cli/src/tui/interactive.ts", "relation": "top-level-registration"},
+        {"type": "source", "target": "packages/cli/src/tui/adjudication-live-block.ts", "relation": "live-projection"}
+      ],
+      "related": ["workflow.evidence-bound-adjudication", "layer.orchestration", "layer.observability", "module.orchestrator", "guide.adjudication-uppercase"],
+      "verification": [
+        {"check": "Run pnpm harness:adjudication", "expected": "Parallel constituent streams, judge synthesis, and a durable receipt complete without loading a real model"},
+        {"check": "Inspect the parent-runner and child-agent tool registration tests", "expected": "Top-level full agents keep adjudicate inline and host scope enforcement excludes it from child runners"}
+      ],
+      "failure_modes": [
+        {"symptom": "A caller attempts /v1/tools/adjudicate/call", "likely_cause": "Static tool discovery was mistaken for daemon direct-tool exposure", "recovery": "Use a top-level TUI, one-shot CLI, /v1/run, or default full-agent /v1/chat path"},
+        {"symptom": "The case returns held instead of a verdict", "likely_cause": "Admission, citation validation, quorum, or final verdict validation failed", "recovery": "Inspect the receipt and provide a corrected admissible record; do not infer a decision"}
+      ],
+      "source_of_truth": ["docs/ADJUDICATION.md", "packages/orchestrator/src/adjudication.ts", "packages/cli/src/tui/interactive.ts"]
+    },
     {
       "id": "tool.tool-search",
       "kind": "tool",

package/docs/guides/tools-and-web-search.md

@@ -85,6 +85,41 @@ origin, network/off-device policy, and risk.
 See [Tools, MCP, Hooks, Agents, And Code Graph](../rest/endpoints/tools.md) for
 the complete REST contract.
 
+## `adjudicate` Is A Top-Level Full-Agent Tool
+
+`adjudicate` resolves a genuine decision impasse through an isolated,
+evidence-scoped panel and final judge. It is agent-bound, but it is not a
+daemon direct tool. The interactive TUI, one-shot CLI, `POST /v1/run`, and the
+default full-agent `POST /v1/chat` path receive it. REST execution remains
+subject to the active full-agent tool profile.
+
+The parent freezes an admissible record before deliberation. Fresh constituent
+contexts receive only assigned evidence and rules, and the judge receives the
+record plus validated findings instead of the parent transcript. This reduces
+anchoring from accumulated working context. It does not erase bias from record
+selection or framing, which remain visible in the durable receipt.
+
+The tool is intentionally excluded from sub-agents. This prevents nested
+adjudication panels, recursive live blocks, and unbounded inference fan-out.
+It is also absent from the daemon direct-tool registry, the bounded
+`/v1/chat/completions` mini-loop, Telegram, and voice-call agents. Do not infer
+`GET /v1/tools/adjudicate`, `POST /v1/tools/adjudicate/call`, or a Telegram
+tool route from its catalog entry. Only the interactive TUI renders the
+specialized streamed adjudication block. Headless and REST full-agent paths
+use ordinary run/tool events and still persist the same receipt artifacts.
+
+Use these offline discovery commands:
+
+```bash
+omnius discover "evidence-bound decision impasse"
+omnius show tool.adjudicate
+omnius show workflow.evidence-bound-adjudication
+```
+
+See [Evidence-Bound Adjudication](../ADJUDICATION.md) for admission rules,
+panel isolation, quorum, verdict validation, durable receipts, CLI
+observability, and the deterministic harness.
+
 ## Failure Rules
 
 - Unknown tool: refresh discovery/metadata; do not retry the same name.

package/docs/index.md

@@ -32,6 +32,7 @@ should additionally expand the owning layer and module.
 - [Media Generation](./guides/media-generation.md)
 - [System Tray Indicator](./guides/system-tray.md)
 - [Tools And Web Search](./guides/tools-and-web-search.md)
+- [Evidence-Bound Adjudication](./ADJUDICATION.md)
 - [Categorized OSINT Research](./guides/osint-research.md)
 
 ## Reference

package/npm-shrinkwrap.json

@@ -1,12 +1,12 @@
 {
   "name": "omnius",
-  "version": "1.0.675",
+  "version": "1.0.677",
   "lockfileVersion": 3,
   "requires": true,
   "packages": {
     "": {
       "name": "omnius",
-      "version": "1.0.675",
+      "version": "1.0.677",
       "bundleDependencies": [
         "image-to-ascii"
       ],