Artifacts - Noēsis

Every Noēsis episode produces a set of structured artifacts that capture the complete cognitive trace. These files enable replay, debugging, auditing, and analysis.

Artifact structure

.noesis/episodes/
  <label>/                     # e.g., "demo", "production"
    <episode_id>/              # e.g., "ep_20251108_120000_abc123_def4_s0"
      events.jsonl             # cognitive event timeline with lineage
      summary.json             # metrics and KPIs (insight.metrics)
      state.json               # plan, beliefs, memory, outcomes
      final.json               # sealed terminal result
      manifest.json            # SHA-256 catalog + optional HMAC
      learn.jsonl              # learning signals (optional)
      prompts.jsonl            # prompt provenance (opt-in, ADR-005)
      checkpoints/             # resume anchors for interrupted runs
      tool_invocations/        # prepared side-effect state (when used)
        prepared/
        approvals/
        idempotency/

Episode IDs use ULID format (monotonic, sortable, 48-bit timestamp + 80-bit entropy). Directive and governance IDs use deterministic UUIDv5 for reproducible lineage tracking.

Paused approval workflows are intentionally not sealed. While a run is waiting for approval, expect events.jsonl, state.json, learn.jsonl, and checkpoint or tool-invocation records to exist, but not terminal-only files such as final.json, summary.json, or manifest.json. Those terminal artifacts are written after the run resumes and completes.

Tool invocation state

Protocol-first tool integrations persist review and idempotency state under the episode directory:

Path	Purpose
`tool_invocations/prepared/*.json`	Reviewable prepared tool intent keyed by `run_id + draft_id`
`tool_invocations/approvals/*.json`	Human or policy decisions bound to the reviewed draft
`tool_invocations/idempotency/*.json`	Execution fingerprints used to replay or reject duplicate side effects

For approval-gated subprocess tools, the runtime bridge prepares the draft, emits tool.approval.pending, interrupts the run, and writes a checkpoint. After an approved ToolApprovalDecision is persisted, ns.resume_run(...) continues the same run and executes the existing draft without re-preparing it. Constraints to check during incident response:

there should be exactly one pending_approval prepared draft for the run
unsupported bridge protocols (http, mcp) fail before draft persistence on prepare or before run.resume is emitted on resume
missing final.json or manifest.json is expected until the resumed run reaches a terminal state

summary.json

The summary captures episode outcomes, metrics, and cross-references.

Schema

{
  "schema_version": "1.3.0",
  "episode_id": "ep_2024_abc123_s0",
  "task": "Draft release notes for v1.2.0",
  "seed": 42,
  "started_at": "2024-01-15T10:30:00Z",
  "duration_sec": 5.12,
  "flags": {
    "intuition": true,
    "mode": "meta",
    "using": "langgraph",
    "direction": {
      "applied": 1,
      "vetoed": 0,
      "policy": "SafetyPolicy@1.0",
      "threshold": 0.75,
      "last_diff": [
        "plan.steps[0].params.limit: null → 100"
      ]
    }
  },
  "ports": {
    "model": "openai:gpt-4o-mini"
  },
  "agents_config_hash": "sha256:9f7d...",
  "answer": {},
  "metrics": {
    "success": 1,
    "plan_count": 2,
    "act_count": 3,
    "reflect_count": 1,
    "veto_count": 0,
    "latencies": {
      "first_action_ms": 150,
      "total_ms": 5000
    }
  },
  "insight": {
    "plan_adherence": 0.95,
    "tool_coverage": 1.0,
    "branching_factor": 2
  },
  "tags": {
    "environment": "staging",
    "team": "platform"
  }
}

Key fields

Field	Description
`schema_version`	Version of the summary schema
`episode_id`	Unique identifier
`task`	The original task/goal
`seed`	Seed used for deterministic runs (if provided)
`duration_sec`	Wall-clock duration of the episode
`flags`	Configuration flags for the run
`ports`	Adapter labels bound to the run (e.g., model provider)
`agents_config_hash`	Hash of adapters + intuition config for reproducibility
`metrics`	Computed metrics
`insight`	Advanced insight metrics
`tags`	User-provided metadata

Reading summaries

import noesis as ns

episode_id = ns.last()
summary = ns.summary.read(episode_id)

print(f"Task: {summary['task']}")
print(f"Success: {summary['metrics']['success']}")
print(f"Actions: {summary['metrics']['act_count']}")

state.json

The state captures the cognitive context at the end of the episode.

Schema

{
  "version": "1.0",
  "state_schema_version": "1.0.0",
  "episode": {
    "id": "ep_2024_abc123_s0",
    "seed": 0,
    "adapter": "baseline",
    "started_at": "2024-01-15T10:30:00Z",
    "completed_at": "2024-01-15T10:30:05Z"
  },
  "goal": {
    "task": "Draft release notes for v1.2.0",
    "context": {
      "changelog_path": "CHANGELOG.md"
    }
  },
  "beliefs": [
    {
      "statement": "Version 1.2.0 includes 3 new features",
      "confidence": 0.9,
      "provenance": "changelog_analysis"
    }
  ],
  "plan": {
    "steps": [
      {
        "id": "step_1",
        "kind": "detect",
        "description": "Read changelog entries",
        "status": "done",
        "inputs": {"file": "CHANGELOG.md"},
        "outputs": {"entries": 15}
      },
      {
        "id": "step_2",
        "kind": "act",
        "description": "Generate summary",
        "status": "done"
      }
    ]
  },
  "memory": {
    "facts": [
      {
        "key": "changelog_format",
        "value": "keep-a-changelog",
        "timestamp": "2024-01-15T10:30:02Z"
      }
    ],
    "scratchpad": "Found 3 breaking changes to highlight"
  },
  "outcomes": {
    "status": "ok",
    "actions": [
      {
        "tool": "changelog_reader",
        "input": "CHANGELOG.md",
        "output": {"entries": 15},
        "timestamp": "2024-01-15T10:30:01Z"
      }
    ],
    "metrics": {
      "task_score": 0.95
    },
    "artifacts": ["release_notes.md"]
  }
}

Step kinds

Kind	Purpose
`detect`	Gather information
`analyze`	Process or categorize
`plan`	Sub-planning
`act`	Execute action
`verify`	Check results
`review`	Human review point

Step statuses

Status	Meaning
`pending`	Not yet started
`running`	Currently executing
`done`	Completed successfully
`skipped`	Intentionally skipped
`failed`	Failed with error
`vetoed`	Blocked by policy

Outcome statuses

Status	Meaning
`ok`	Completed successfully
`error`	Failed with error
`vetoed`	Blocked by policy
`aborted`	Manually stopped
`partial`	Partially completed
`pending`	Still running

Reading state

import noesis as ns

episode_id = ns.last()
state = ns.state.read(episode_id)

print(f"Goal: {state['goal']['task']}")
print(f"Steps: {len(state['plan']['steps'])}")
print(f"Status: {state['outcomes']['status']}")

events.jsonl

The event timeline records every phase transition with timing and lineage.

Format

Events are stored as newline-delimited JSON (JSONL):

{"id": "evt_1", "phase": "observe", "payload": {...}, "metrics": {...}}
{"id": "evt_2", "phase": "interpret", "payload": {...}, "caused_by": "evt_1", "metrics": {...}}
{"id": "evt_3", "phase": "plan", "payload": {...}, "caused_by": "evt_2", "metrics": {...}}

Event structure

{
  "id": "evt_abc123",
  "phase": "plan",
  "agent_id": "planner.minimal",
  "payload": {
    "steps": [
      "detect:Gather data",
      "act:Execute task"
    ],
    "step_records": [
      {
        "id": "step-1",
        "kind": "detect",
        "description": "Gather data",
        "status": "pending"
      },
      {
        "id": "step-2",
        "kind": "act",
        "description": "Execute task",
        "status": "pending"
      }
    ],
    "source": "planner.minimal"
  },
  "metrics": {
    "started_at": "2024-01-15T10:30:00.500Z",
    "completed_at": "2024-01-15T10:30:00.512Z",
    "duration_ms": 12.7
  },
  "caused_by": "evt_xyz789"
}

Event metrics always include started_at, completed_at, and duration_ms, which you can aggregate for per-phase latency. Plan derivability contract: state.json.plan can be reconstructed from the latest plan event (steps + step_records + source) plus subsequent act events carrying step_id and step_status.

Event phases

The phase field stores the phase emitted into the trace.

Phase	Purpose
`observe`	Capture raw input
`intuition`	Policy guidance
`interpret`	Extract signals
`plan`	Decide actions
`direction`	Policy directives
`governance`	Audit decisions
`act`	Execute actions
`reflect`	Evaluate outcomes
`learn`	Propose updates
`terminate`	Episode end
`insight`	Metrics and KPIs
`memory`	Persistence status

Reading events

import noesis as ns

episode_id = ns.last()
events = list(ns.events.read(episode_id))

# Filter by phase
plan_events = [e for e in events if e["phase"] == "plan"]
act_events = [e for e in events if e["phase"] == "act"]

# Reconstruct causal chain
for event in events:
    print(f"{event['phase']}: caused_by={event.get('caused_by', 'none')}")

CLI access

# All events
noesis events ep_abc123

# Filter by phase
noesis events ep_abc123 --phase act

# As JSON
noesis events ep_abc123 -j | jq '.[] | select(.phase == "plan")'

manifest.json

The manifest provides integrity verification for all artifacts with SHA-256 hashes and optional signatures.

Schema

{
  "schema_version": "manifest/1.0",
  "episode_id": "ep_2024_abc123_s0",
  "created_at": "2024-01-15T10:30:05Z",
  "files": [
    {"name": "summary.json", "sha256": "sha256:abc123...", "size_bytes": 1234, "kind": "summary"},
    {"name": "state.json", "sha256": "sha256:def456...", "size_bytes": 5678, "kind": "state"},
    {"name": "events.jsonl", "sha256": "sha256:ghi789...", "size_bytes": 9012, "kind": "events"}
  ],
  "signature": {
    "alg": "hs256",
    "kid": "ops-key-1",
    "value": "base64sig",
    "ts": "2024-01-15T10:30:05Z"
  }
}

Field	Description
`schema_version`	Manifest schema (`manifest/1.0`)
`files`	Array of files with `name`, `sha256`, `size_bytes`, `kind` (`summary`, `state`, `events`, `learn`, `attachment`, `custom`)
`signature`	Optional signature block with algorithm, key ID, value, timestamp

Artifact immutability guarantees

All artifacts are written atomically using this pattern:

Write to temporary file
Call fsync() to ensure durability
Atomic rename to final path

This ensures no partial writes survive system crashes.

Verifying integrity

import hashlib
import json
from pathlib import Path


def verify_manifest(episode_dir: Path) -> bool:
    """Verify artifact integrity using manifest."""
    manifest_path = episode_dir / "manifest.json"
    
    with open(manifest_path) as f:
        manifest = json.load(f)
    
    for filename, info in manifest["files"].items():
        file_path = episode_dir / filename
        
        # Check size
        if file_path.stat().st_size != info["size"]:
            return False
        
        # Check hash
        with open(file_path, "rb") as f:
            actual_hash = hashlib.sha256(f.read()).hexdigest()
        
        if actual_hash != info["sha256"]:
            return False
    
    return True

JSON serialization uses canonical_dumps() with sorted keys for byte-identical output, ensuring consistent hashes across runs.

learn.jsonl (optional)

Learning signals are stored separately when the learn phase emits proposals.

Schema

{"proposal_id": "prop_1", "type": "threshold", "target": "veto_confidence", "suggested_value": 0.8}
{"proposal_id": "prop_2", "type": "memory", "key": "pattern_detected", "value": "sql_injection_attempt"}

Reading learn signals

import json
from pathlib import Path

learn_path = Path(f".noesis/episodes/{episode_id}/learn.jsonl")

if learn_path.exists():
    with open(learn_path) as f:
        proposals = [json.loads(line) for line in f]
    
    for p in proposals:
        print(f"Proposal: {p['type']} -> {p.get('target', p.get('key'))}")

prompts.jsonl (opt-in)

When prompt provenance is enabled (ADR-005, experimental), all prompts are recorded for debugging and auditing.

Schema

{
  "prompt_id": "pmt_abc123",
  "timestamp": "2024-01-15T10:30:01Z",
  "model": "gpt-4",
  "template": "Generate release notes for:\n{changelog}",
  "variables": {"changelog": "..."},
  "response": "## Release Notes...",
  "tokens": {"input": 150, "output": 200},
  "latency_ms": 1250
}

Experimental feature (ADR-005): Enable prompt logging carefully—prompts may contain sensitive data. Use ns.set(log_prompts=True) only when needed for debugging or compliance.

Use cases

Debugging: Trace exactly what prompts were sent to LLMs
Compliance: Audit trail for regulated industries
Cost analysis: Track token usage per episode
Prompt optimization: Analyze prompt patterns across runs

Storage configuration

Runs directory

import noesis as ns

# Set custom runs directory
ns.set(runs_dir="./my-runs")

# Or via environment
# NOESIS_RUNS_DIR=./my-runs

Labels

Organize episodes with labels:

import noesis as ns

ns.set(label="production")
# Artifacts go to .noesis/episodes/production/ep_.../

ns.set(label="staging")
# Artifacts go to .noesis/episodes/staging/ep_.../

Retention

Control artifact retention:

from noesis.episode import EpisodeIndex

store = EpisodeIndex("./.noesis/episodes/_episodes", ttl_days=14)
store.vacuum()  # Remove episodes older than 14 days

Best practices

Back up production artifacts. The runs directory contains valuable audit data—include it in your backup strategy.

Redact sensitive data. Tasks and prompts may contain PII. Consider scrubbing before long-term storage.

Use manifests for compliance. The SHA-256 checksums in manifest.json provide tamper evidence for audits.

Next steps

Events reference

Complete event schema documentation.

State reference

Complete state schema documentation.

Export metrics

Send metrics to observability tools.

Summary reference

Complete summary schema documentation.

​Artifact structure

​Tool invocation state

​summary.json

​Schema

​Key fields

​Reading summaries

​state.json

​Schema

​Step kinds

​Step statuses

​Outcome statuses

​Reading state

​events.jsonl

​Format

​Event structure

​Event phases

​Reading events

​CLI access

​manifest.json

​Schema

​Artifact immutability guarantees

​Verifying integrity

​learn.jsonl (optional)

​Schema

​Reading learn signals

​prompts.jsonl (opt-in)

​Schema

​Use cases

​Storage configuration

​Runs directory

​Labels

​Retention

​Best practices

​Next steps

Events reference

State reference

Export metrics

Summary reference

Artifact structure

Tool invocation state

summary.json

Schema

Key fields

Reading summaries

state.json

Schema

Step kinds

Step statuses

Outcome statuses

Reading state

events.jsonl

Format

Event structure

Event phases

Reading events

CLI access

manifest.json

Schema

Artifact immutability guarantees

Verifying integrity

learn.jsonl (optional)

Schema

Reading learn signals

prompts.jsonl (opt-in)

Schema

Use cases

Storage configuration

Runs directory

Labels

Retention

Best practices

Next steps