omnius 1.0.589 → 1.0.591

package/.aiwg/addons/omnius-docs/README.md

@@ -3,3 +3,5 @@
 This project-local addon exposes Omnius documentation as small discoverable skills. It is for agents that need to answer questions about Omnius without loading the entire docs tree.
 
 Start with `omnius-docs`, then load a more specific skill when the task is about realtime, sponsors, Telegram, operations, or browser interaction validation.
+
+For cross-cutting delivery work, the add-on also exposes `evidence-directed-delivery`, `workspace-reality-audit`, `test-surface-audit`, `runtime-provenance-audit`, `hardware-evidence-audit`, `secrets-and-config-audit`, and `agent-failure-recovery`. These skills are feedback-oriented: they help an agent discover evidence gaps and recover from actual failures without treating planning telemetry as a completion veto.

package/.aiwg/addons/omnius-docs/manifest.json

@@ -32,6 +32,41 @@
       "name": "browser-interaction-validation",
       "description": "Validate a user-facing browser flow with canonical Playwright interactions, postconditions, diagnostics, and honest blocked/not-applicable evidence.",
       "triggers": ["browser interaction validation", "UI interaction test", "Playwright validation", "browser flow", "dashboard validation", "form interaction"]
+    },
+    {
+      "name": "evidence-directed-delivery",
+      "description": "Use generic claim/evidence, anti-reward-hacking, bounded recovery, and feedback-only completion rules for non-trivial work.",
+      "triggers": ["evidence based completion", "delivery audit", "completion evidence", "verify claims", "remaining gaps"]
+    },
+    {
+      "name": "workspace-reality-audit",
+      "description": "Audit repository truth, generated artifacts, documentation drift, and delivery ownership before claiming progress.",
+      "triggers": ["workspace audit", "repository reality", "stale artifact", "untracked source", "documentation drift"]
+    },
+    {
+      "name": "test-surface-audit",
+      "description": "Map canonical tests to claims and expose placeholder, mock-only, copied-logic, and un-wired verification gaps.",
+      "triggers": ["test audit", "test coverage", "test surface", "canonical test", "mock only"]
+    },
+    {
+      "name": "runtime-provenance-audit",
+      "description": "Prove active process, artifact, port, configuration, and restart/update ownership for runtime claims.",
+      "triggers": ["runtime provenance", "deployed version", "process audit", "service ownership", "running artifact"]
+    },
+    {
+      "name": "hardware-evidence-audit",
+      "description": "Classify static, simulation, integration, and physical-device evidence without overstating hardware proof.",
+      "triggers": ["hardware validation", "device evidence", "simulation versus hardware", "firmware test", "physical device"]
+    },
+    {
+      "name": "secrets-and-config-audit",
+      "description": "Audit credential and configuration provenance without exposing secret values.",
+      "triggers": ["secret audit", "credential provenance", "endpoint credential", "configuration security"]
+    },
+    {
+      "name": "agent-failure-recovery",
+      "description": "Convert failed calls into bounded, evidence-backed recovery actions and stop blind retries.",
+      "triggers": ["recover tool failure", "stale hash", "unknown tool", "retry loop", "agent recovery"]
     }
   ]
 }

package/.aiwg/addons/omnius-docs/skills/agent-failure-recovery/SKILL.md

@@ -0,0 +1,23 @@
+---
+compaction_strategy: structured
+---
+
+# agent-failure-recovery
+
+Use after a failed tool, rejected mutation, unknown capability, stale hash,
+timeout, or contradictory evidence.
+
+## Procedure
+
+1. Preserve the raw failure and classify it: retryable, blocked, permanent, or
+   unknown.
+2. Name the authoritative state needed to change strategy. For a stale edit,
+   re-read the target; for an unavailable tool, inspect the capability contract;
+   for a failed verifier, inspect its exact diagnostic.
+3. Create one recovery action with evidence requirements and an exit condition.
+4. Retry only after a distinct observation or correction. Do not repeat an
+   identical failure.
+5. Resolve the failure only with evidence that addresses the same tool family,
+   claim, and current state. Keep it as feedback; do not use it as a terminal
+   completion gate.
+

package/.aiwg/addons/omnius-docs/skills/evidence-directed-delivery/SKILL.md

@@ -0,0 +1,63 @@
+---
+compaction_strategy: structured
+---
+
+# evidence-directed-delivery
+
+Use this skill for any non-trivial implementation, diagnosis, review, release,
+or completion claim. It keeps the agent oriented to observed system state
+without making verification a veto on `task_complete`.
+
+## Required operating contract
+
+1. Start with a bounded, read-only discovery pass. Inspect only surfaces
+   relevant to the request: workspace, build/test, runtime, delivery boundary,
+   configuration/security, and existing failures.
+2. State which observations are direct evidence and which are only proxies.
+   Source edits, command exit codes, screenshots, simulations, and task
+   completion are not proof by themselves.
+3. For every material claim, identify the most likely counterexample and check
+   it when the capability is available and in scope.
+4. Classify each result as `verified`, `unverified`, `contradicted`, or
+   `blocked`. Preserve failed and inconclusive evidence until newer evidence
+   addresses the same claim and state.
+5. `task_complete` remains callable. Before calling it, provide the main loop
+   structured feedback: verified claims, unverified claims, contradictions,
+   active recovery items, and surfaces not checked.
+
+## Anti-reward-hacking rules
+
+- Do not use “fixed”, “complete”, “verified”, “deployed”, “healthy”, or
+  “end-to-end” unless the corresponding claim has fresh, scope-matched
+  evidence.
+- Do not count tool volume, todo volume, created files, plans, or persuasive
+  prose as implementation evidence.
+- Do not convert simulation or mock evidence into real-device evidence.
+- Do not hide a failed tool call because a later unrelated check passed.
+- Do not retry the same failure unchanged. Classify it, inspect authoritative
+  state, choose a distinct recovery action, and retry once only when justified.
+
+## Capability contract
+
+Use only tools in the current schema or `[RUNTIME CAPABILITY CONTRACT v1]`.
+When a tool is unavailable, read its structured recovery output, use an
+advertised equivalent if present, and do not repeat the unavailable call.
+
+## Handoff shape
+
+```text
+claim: active artifact contains the change
+- status: verified
+- evidence: current executable hash matched built artifact
+- counter-check: process path and working directory inspected
+
+claim: device path works
+- status: unverified
+- evidence: simulation harness passed
+- boundary: no physical device/transport observation was available
+
+recovery: stale edit target
+- status: resolved
+- evidence: re-read current target, rebuilt the edit, then ran a fresh verifier
+```
+

package/.aiwg/addons/omnius-docs/skills/hardware-evidence-audit/SKILL.md

@@ -0,0 +1,25 @@
+---
+compaction_strategy: structured
+---
+
+# hardware-evidence-audit
+
+Use for physical-device, transport, firmware, sensor, accelerator, or browser
+permission claims.
+
+## Evidence ladder
+
+Classify every relevant result exactly once:
+
+- static source inspection
+- mock or simulation
+- emulator or software-in-the-loop
+- integration against a live service
+- connected physical device
+- production/field observation
+
+Never promote an item to a higher level without a fresh observation at that
+level. Before live local inference or hardware work, follow the project’s
+hardware preflight rules. If the required device, endpoint, permission, or
+accelerator cannot be proven, report `blocked` or `not_applicable`.
+

package/.aiwg/addons/omnius-docs/skills/runtime-provenance-audit/SKILL.md

@@ -0,0 +1,20 @@
+---
+compaction_strategy: structured
+---
+
+# runtime-provenance-audit
+
+Use for any claim about a running service, deployed code, background worker,
+port, endpoint, or update.
+
+## Procedure
+
+1. Inspect active process identity, executable/module path, version or hash,
+   working directory, start time, and active listener/port.
+2. Compare the active artifact with the intended built artifact. Source state
+   alone cannot satisfy this claim.
+3. Identify service ownership, restart/update procedure, configuration origin,
+   and whether a process is project-supervised or ad hoc.
+4. Report unavailable access as `blocked` or `unverified`, never as a failed
+   implementation claim.
+

package/.aiwg/addons/omnius-docs/skills/secrets-and-config-audit/SKILL.md

@@ -0,0 +1,22 @@
+---
+compaction_strategy: structured
+---
+
+# secrets-and-config-audit
+
+Use before remote inference, deployment, endpoint changes, or when inspecting
+configuration provenance.
+
+## Procedure
+
+1. Discover whether required credential sources exist without displaying their
+   values. Prefer a named environment/credential reference over raw settings.
+2. Inspect tracked and ignored configuration separately; ignored state can
+   still be an operational or security finding.
+3. Never print tokens, authorization headers, secret values, or full URLs that
+   embed credentials.
+4. For remote inference, record endpoint, exact model identifier, credential
+   source indicator, timestamp, and non-secret receipt only.
+5. Treat 401/403, unavailable models, or changed free/pricing status as a
+   configuration outcome requiring action—not an invitation to retry blindly.
+

package/.aiwg/addons/omnius-docs/skills/test-surface-audit/SKILL.md

@@ -0,0 +1,22 @@
+---
+compaction_strategy: structured
+---
+
+# test-surface-audit
+
+Use when a task changes behavior, claims a fix, or asks for release readiness.
+
+## Procedure
+
+1. Identify the canonical build and test commands from project manifests and
+   documentation. Flag placeholders and commands that intentionally fail.
+2. Map each requested claim to source, unit, integration, browser, and
+   hardware evidence levels. State which level is actually exercised.
+3. Check whether tests invoke production code or duplicate the implementation
+   under test. Copied protocol/decoder logic is a drift risk, not independent
+   proof.
+4. Run only relevant safe checks, capture fresh results, and distinguish
+   passing execution from coverage of the requested behavior.
+5. On failure, retain the exact diagnostic and create a distinct recovery
+   action; do not blanket-retry.
+

package/.aiwg/addons/omnius-docs/skills/workspace-reality-audit/SKILL.md

@@ -0,0 +1,24 @@
+---
+compaction_strategy: structured
+---
+
+# workspace-reality-audit
+
+Use for discovery of repository and delivery gaps before implementation or a
+completion claim.
+
+## Procedure
+
+1. Discover available read-only tools from the current capability contract.
+2. Inspect tracked, modified, untracked, ignored, and generated paths.
+3. Compare documentation claims with actual source and repository layout.
+4. Compare changed source with generated artifacts by timestamp/hash where
+   applicable; do not infer a build occurred from an old artifact existing.
+5. Report only observed gaps. An empty exploratory scan is an advisory, not a
+   missing-evidence failure unless the task explicitly named required artifacts.
+
+## Output
+
+For each finding emit: `status`, `evidence`, `impact`, `counter-check`, and
+`smallest next action`. Keep workspace findings separate from runtime claims.
+