Skip to content

fix(v10): wire self-healing runtime into compute and agent execution #23

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
jayDevCodes merged 5 commits into from
Aug 25, 2026
Merged

fix(v10): wire self-healing runtime into compute and agent execution #23

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
3e73580
fix(v10): integrate self-healing repair with compute and agent runtime
jayDevCodes Aug 25, 2026
e3b9fd0
feat(runtime): allow self-healing compute policy to pin worker tier
jayDevCodes Aug 25, 2026
fe30b4f
fix(v10): honor compute policy tier during repair workers
jayDevCodes Aug 25, 2026
7a8ea79
test(v10): cover integrated self-healing repair path
jayDevCodes Aug 25, 2026
ac4ba48
docs(state): record V10 self-healing capabilities and evidence targets
jayDevCodes Aug 25, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
37 changes: 19 additions & 18 deletions MODEL_STATE.json
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,55 +1,56 @@
{
"schema_version": 5,
"model_version": "v10-cognitive-compute-report",
"schema_version": 6,
"model_version": "v10.0-self-healing-runtime",
"purpose": "Canonical architecture and measurement baseline for evidence-driven MY-AI upgrades.",
"reporting": {
"entrypoint": "myai.build_model_report / AIEngine.model_report",
"source_of_truth": ["live engine feature presence", "CapabilityLedger", "CapabilityBenchmark cases", "EvolutionMemory strategy scores", "CognitiveComputeController policies"],
"source_of_truth": ["live engine feature presence", "CapabilityLedger", "CapabilityBenchmark cases", "EvolutionMemory strategy scores", "CognitiveComputeController policies", "SelfHealingRuntime episodes", "CodeHealthStore"],
"manual_capability_scores": false,
"numeric_claim_rule": "Numeric capability claims require executable evidence; architecture presence alone is not a score."
},
"architecture": {
"core": "recursive multi-model cognitive mesh with bounded cognitive context, persistent evidence-gated memory, capability-aware compute allocation, and self-derived reporting",
"core": "recursive multi-model cognitive mesh with bounded cognitive context, persistent evidence-gated memory, capability-aware compute allocation, causal repository intelligence, and bounded self-healing runtime",
"model_routing": ["fast", "balanced", "frontier"],
"compute_control": ["capability-gap pressure", "uncertainty pressure", "risk pressure", "historical strategy reliability", "bounded execution budgets"],
"repository_representation": ["persistent symbol graph", "repository twin", "program graph", "call edges", "data-flow references", "control-flow edges"],
"runtime_evidence": ["persistent trace events", "causal links", "failure neighborhoods"],
"memory": ["semantic knowledge", "typed cognitive memory", "persistent cognitive memory", "evidence-gated lifecycle", "repair memory", "evolution memory", "capability ledger"],
"verification": ["specialist workers", "frontier judge", "targeted validation", "promotion gates", "deterministic capability checks"],
"self_improvement": ["strategy scoring", "verified strategy feedback", "baseline comparison", "capability deltas", "regression detection", "executable architecture benchmarks", "compute-policy adaptation"]
"compute_control": ["capability-gap pressure", "uncertainty pressure", "risk pressure", "historical strategy reliability", "bounded execution budgets", "repair-specific tier enforcement"],
"repository_representation": ["persistent symbol graph", "repository twin", "program graph", "call edges", "data-flow references", "control-flow edges", "causal impact slices"],
"runtime_evidence": ["persistent trace events", "causal links", "failure neighborhoods", "failure signatures", "repair episodes"],
"memory": ["semantic knowledge", "typed cognitive memory", "persistent cognitive memory", "evidence-gated lifecycle", "repair memory", "evolution memory", "capability ledger", "failure signature memory"],
"self_healing": ["failure detection", "causal diagnosis", "bounded reproduction", "policy-governed recursive repair workers", "independent frontier judge", "fault injection", "stable-code health", "validation-gated promotion"],
"verification": ["specialist workers", "frontier judge", "targeted validation", "promotion gates", "deterministic capability checks", "fault recovery verification"],
"self_improvement": ["strategy scoring", "verified strategy feedback", "baseline comparison", "capability deltas", "regression detection", "executable architecture benchmarks", "compute-policy adaptation", "repair lessons"]
},
"capability_dimensions": {
"reasoning": {"status": "architecture_benchmarked", "score_source": "executable architecture evidence"},
"coding": {"status": "architecture_benchmarked_via_code_localization", "score_source": "executable repository evidence"},
"repository_understanding": {"status": "architecture_benchmarked", "score_source": "program/repository twin evidence"},
"debugging": {"status": "architecture_benchmarked", "score_source": "impact/dependency evidence"},
"debugging": {"status": "architecture_benchmarked_with_causal_trace", "score_source": "impact/dependency + runtime trace evidence"},
"memory": {"status": "architecture_benchmarked", "score_source": "typed lifecycle + bounded retrieval evidence"},
"tool_use": {"status": "multi_model_provider_layer_unmeasured", "score": null},
"planning": {"status": "architecture_benchmarked", "score_source": "cognitive planning + bounded graph evidence"},
"verification": {"status": "architecture_benchmarked", "score_source": "deterministic cognitive verification evidence"},
"efficiency": {"status": "architecture_benchmarked", "score_source": "bounded-context evidence"},
"efficiency": {"status": "architecture_benchmarked", "score_source": "bounded-context + stable-code health evidence"},
"self_improvement": {"status": "architecture_benchmarked", "score_source": "evidence-gated strategy promotion"}
},
"evidence_runner": {
"entrypoint": "myai.run_architecture_benchmark",
"mode": "deterministic",
"requires_provider_model": false,
"measured_cases": ["frontier-routing", "recursive-agent-graph", "cross-episode-repair-memory", "trace-verification", "program-graph-localization", "causal-trace-diagnosis", "targeted-context", "strategy-promotion"],
"unmeasured_provider_capabilities": ["tool_use"],
"measured_cases": ["frontier-routing", "recursive-agent-graph", "cross-episode-repair-memory", "trace-verification", "program-graph-localization", "causal-trace-diagnosis", "targeted-context", "strategy-promotion", "self-healing-runtime", "fault-injection-recovery", "stable-code-inspection"],
"unmeasured_provider_capabilities": ["tool_use", "provider-backed task quality"],
"principle": "Executed evidence, not architecture presence, is required for numeric capability claims."
},
"compute_policy": {
"entrypoint": "myai.CognitiveComputeController",
"inputs": ["capability ledger gap", "task uncertainty", "task risk", "historical strategy reliability", "routing decision"],
"outputs": ["reasoning depth", "parallelism", "retry budget", "verification passes", "preferred tier"],
"constraints": ["bounded depth", "bounded parallelism", "bounded retries", "never overrides provider-independent evidence requirements"],
"constraints": ["bounded depth", "bounded parallelism", "bounded retries", "repair worker tier is enforced while independent judges remain frontier", "never overrides provider-independent evidence requirements"],
"principle": "Use more compute when measured capability gaps or uncertainty justify it; avoid redundant exploration when verified strategies are reliable."
},
"benchmark_protocol": {
"required_before_claiming_improvement": ["fixed task set", "same model/provider configuration", "same tool budget where applicable", "multiple runs for reliability", "task success", "regression rate", "latency", "tokens/cost", "unnecessary context/code-read volume", "memory transfer across episodes", "persisted evidence for every numeric claim"],
"current_executable_suite": ["frontier-routing", "recursive-agent-graph", "cross-episode-repair-memory", "trace-verification", "program-graph-localization", "causal-trace-diagnosis", "targeted-context", "strategy-promotion"],
"current_executable_suite": ["frontier-routing", "recursive-agent-graph", "cross-episode-repair-memory", "trace-verification", "program-graph-localization", "causal-trace-diagnosis", "targeted-context", "strategy-promotion", "self-healing-runtime", "fault-injection-recovery", "stable-code-inspection"],
"external_reference_directions": ["STATE-Bench", "EvoMemBench", "EvoAgentBench", "PAST-Bench"]
},
"known_limits": ["The deterministic runner measures architecture capabilities, not frontier-model task quality.", "Tool-use capability remains unmeasured by the provider-independent runner.", "Fresh GitHub Actions execution is not guaranteed to be visible through the current integration at every commit.", "Program/data/control-flow coverage is strongest for Python and should be extended to other languages before broad repo claims.", "Compute policy is a bounded controller; it does not train model weights or guarantee task-level quality improvements.", "Self-evolution changes remain evidence-gated; automatic model-weight training is not implemented.", "Self-derived reporting summarizes live architecture state but does not replace independent benchmark evaluation."]
,"next_upgrade_targets": ["provider-backed end-to-end benchmark harness", "real runtime instrumentation and automatic trace capture", "sandbox patch execution with automatic rollback", "tool-use benchmark with explicit capability boundaries", "longitudinal benchmark runner with repeated runs and confidence intervals", "ability-transfer memory and curriculum generation", "model-vs-model and architecture-vs-architecture comparison harness", "predictive world-state and counterfactual simulation"]
"known_limits": ["The deterministic runner measures architecture capabilities, not frontier-model task quality.", "Tool-use capability remains unmeasured by the provider-independent runner.", "Fresh GitHub Actions execution is not guaranteed to be visible through the current integration at every commit.", "Program/data/control-flow coverage is strongest for Python and should be extended to other languages before broad repo claims.", "Compute policy is a bounded controller; it does not train model weights or guarantee task-level quality improvements.", "Self-evolution changes remain evidence-gated; automatic model-weight training is not implemented.", "Self-healing generates and validates repair episodes but production patch promotion remains an external gated step.", "Self-derived reporting summarizes live architecture state but does not replace independent benchmark evaluation."],
"next_upgrade_targets": ["provider-backed end-to-end benchmark harness", "real runtime instrumentation and automatic trace capture", "sandbox patch execution with automatic rollback", "tool-use benchmark with explicit capability boundaries", "longitudinal benchmark runner with repeated runs and confidence intervals", "ability-transfer memory and curriculum generation", "model-vs-model and architecture-vs-architecture comparison harness", "predictive world-state and counterfactual simulation"]
}
46 changes: 28 additions & 18 deletions src/myai/agent_runtime.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,10 +20,17 @@ class AgentRuntimeResult:
class MultiModelAgentRuntime:
"""Connect the recursive graph to real tier-specific model endpoints."""

def __init__(self, pool: TieredModelPool, router: AdaptiveModelRouter, budget: ExecutionBudget) -> None:
def __init__(
self,
pool: TieredModelPool,
router: AdaptiveModelRouter,
budget: ExecutionBudget,
worker_tier_override: ModelTier | None = None,
) -> None:
self.pool = pool
self.router = router
self.budget = budget
self.worker_tier_override = worker_tier_override
self._tiers: dict[str, ModelTier] = {}

def run(
Expand Down Expand Up @@ -68,19 +75,23 @@ def _decompose(self, node: TaskNode, budget: ExecutionBudget) -> Sequence[TaskNo
)

def _choose_tier(self, node: TaskNode, context: Sequence[ChatMessage]) -> ModelTier:
decision = self.router.choose(
RoutingRequest(
task_kind=node.role,
complexity=0.85 if node.depth == 0 else 0.7,
uncertainty=0.75 if node.role in {"critic", "reviewer", "countercheck"} else 0.45,
context_size=sum(len(m.content) for m in context),
risk="high" if node.role in {"reviewer", "countercheck", "synthesis"} else "medium",
latency_sensitive=False,
quality_priority=node.role in {"critic", "reviewer", "synthesis"},
if self.worker_tier_override is not None and node.role not in {"critic", "reviewer", "countercheck", "synthesis"}:
tier = self.worker_tier_override
else:
decision = self.router.choose(
RoutingRequest(
task_kind=node.role,
complexity=0.85 if node.depth == 0 else 0.7,
uncertainty=0.75 if node.role in {"critic", "reviewer", "countercheck"} else 0.45,
context_size=sum(len(m.content) for m in context),
risk="high" if node.role in {"reviewer", "countercheck", "synthesis"} else "medium",
latency_sensitive=False,
quality_priority=node.role in {"critic", "reviewer", "synthesis"},
)
)
)
self._tiers[node.id] = decision.tier
return decision.tier
tier = decision.tier
self._tiers[node.id] = tier
return tier

def _worker(
self,
Expand All @@ -98,10 +109,9 @@ def _worker(
system = ChatMessage(
role="system",
content=(
"You are a specialist worker in MY-AI V9.1. Return only useful work for the "
"assigned role. Do not claim tools or evidence you did not receive. Preserve "
"uncertainty and disagreements. Use the shared cognitive state as context, not as "
"proof; distinguish beliefs from observations and avoid inventing missing facts."
"You are a specialist worker in MY-AI V10. Return only useful work for the assigned role. "
"Do not claim tools or evidence you did not receive. Preserve uncertainty and disagreements. "
"Use the shared cognitive state as context, not as proof; distinguish beliefs from observations and avoid inventing missing facts."
),
)
prompt = ChatMessage(
Expand Down Expand Up @@ -138,7 +148,7 @@ def _judge(
ChatMessage(
role="system",
content=(
"You are an independent judge for MY-AI V9.1. Evaluate whether the worker output is "
"You are an independent judge for MY-AI V10. Evaluate whether the worker output is "
"consistent, useful, evidence-aware and complete. Treat the cognitive state as "
"context rather than ground truth. Return JSON only: "
'{"passed":true|false,"confidence":0.0-1.0,"feedback":"..."}'
Expand Down
Loading