agentprovenance

AgentProvenance is a local-first security analysis and provenance control plane for autonomous, tool-using agents, especially sandboxed coding agents.

It is not a generic sandbox runtime, generic telemetry collector, Kubernetes/Ray replacement, RL trainer, trace dashboard, or version-control system (it borrows Git's content-addressing and verification model, not its branch/merge workflow). It owns a narrower primitive:

Execution Context
  -> Evidence Ingest
  -> Runtime Causality Graph
  -> Provenance DAG
  -> State Diff / Blame / Artifact Lineage
  -> Security Analysis / Risk Decision
  -> Taint / Response Action
  -> Replay / Forensics / Audit Manifest

The goal is to answer questions ordinary traces do not answer well:

• Which snapshot did this execution start from?
• Which attempt produced this artifact?
• Which tool call started this process?
• Which child process caused this runtime event?
• Which process changed this file?
• Which behavior is anomalous for this agent or task profile?
• Which execution branch was tainted, quarantined, interrupted, or blocked by a response gate?
• Which evidence supports a risk decision?
• What response action should be triggered: audit, deny, kill, quarantine, taint, export forensics, or notify a human through Feishu/DingTalk?
• What exact behavior evidence, deviation signal, and risk context should an external evaluator, RL pipeline, or human reviewer inspect?
• Can this execution be diffed, blamed, verified, replayed, and audited later?

Contents

• Why
• Security Loop
• Core Model
  • White-box mode
  • Zero-SDK mode
• Relationship To Existing Systems
• Quickstart
• Deployment Modes
  • Falco-compatible Receiver
• Security Evidence Commands
• External Evaluator Protocol
• Compliance Evidence, Not Certification
• Graph Commands
• Current Capability
• Core Demo Acceptance
• Architecture
• Substrate Signals
• Boundaries
• Repository Layout
• Roadmap
• Development

Why

Modern agent execution is not one prompt and one tool call. Coding agents and autonomous workflows fork attempts, edit files, run tests, create artifacts, spawn subprocesses, touch external systems, and trigger runtime telemetry. Logs, traces, metrics, and sandbox events each capture pieces of that story, but they rarely produce a Git-like causal record of execution state.

AgentProvenance turns sandboxed execution into a security-oriented evidence graph:

base snapshot
  -> attempt
  -> execution context
  -> tool_call
  -> process / child process
  -> runtime_event
  -> file_diff / artifact
  -> baseline feature / risk signal
  -> taint / response action
  -> replay / forensics / audit manifest

The primary path is recording and explaining sandboxed agent execution. The branch-heavy coding-agent script is only a stress demo: it creates many branches, artifacts, runtime events, and risk cases quickly enough to exercise the graph. AgentProvenance does not choose the reward winner. It emits structured trajectory evidence and expectation-deviation signals so an external evaluator or training pipeline can turn them into reward, penalty, filtering, or human-review decisions.

For RL pipelines, the useful primitive is not "best-of-one" or automatic winner selection. The useful primitive is observability over each trajectory: what the agent did, which subprocesses and files were touched, which network/runtime events appeared, which behavior violated safety or task expectations, and which risk/baseline signals should contribute to reward shaping or rejection.

Security Loop

The security model is intentionally simple and concrete:

application context
  run / session / attempt / tool_call / user / task / workspace

system telemetry
  process / file / network / resource / sandbox / eBPF event

correlation
  container_id / cgroup_id / pid / ppid / cwd / timestamp / file diff

security analysis
  behavior baseline / suspicious event / taint lineage / risk decision

response
  audit / deny / kill / quarantine / taint / forensics / Feishu or DingTalk notification

This makes AgentProvenance closer to an AI-era HIDS/control-plane layer than a pure LLM trace dashboard. Traditional host monitoring asks "what did this process do?" AgentProvenance adds the agent execution context needed to ask "which agent/tool/task caused it, what state did it change, what evidence proves it, and what response should happen?"

The project currently implements the evidence graph, runtime correlation, diff/blame, telemetry batch manifests, policy decisions, normalized risk signals, baseline deviation records, response action records, taint, quarantine, and forensics/export foundations. Deeper eBPF receivers and Feishu/DingTalk response adapters belong to the next security-control phases.

Core Model

AgentProvenance supports two context modes.

White-box mode

An agent harness, SDK, tool router, or framework explicitly provides context:

run_id / session_id / attempt_id / tool_call_id / tool_name / args_hash

This gives high precision and fits custom coding-agent systems, Agentix-style harnesses, LangGraph-like workflows, and internal tool routers.

Zero-SDK mode

The user can run an agent command directly:

agentprov record -- <agent command>

The current MVP records a command in a working directory, snapshots the pre-execution file state, runs the command, computes post-execution file changes, and emits runtime file evidence into the DAG. The long-term zero-SDK path adds deeper process-tree and kernel telemetry capture.

Zero-SDK inference uses runtime facts:

root process / process tree / cwd / timestamp / container_id / cgroup_id
  / file diff / artifact refs

Raw system-side telemetry should not be required to carry tool_call_id. Kernel and runtime signals usually know PID, cgroup, namespace, container ID, timestamp, and process tree. AgentProvenance correlates those substrate facts back to execution context.

Today, the CLI exposes the underlying binding primitive:

agentprov telemetry bind --run <run_id> --session <session_id> \
  --attempt <attempt_id> --tool-call <tool_call_id> --process <process_id> \
  --container-id <container_id> --cgroup-id <cgroup_id> --pid <pid>

Then raw events can be ingested without tool_call_id:

agentprov telemetry ingest --raw-event raw-execve-1 --pid <pid> \
  --timestamp <event_time> --source tetragon_jsonl --type execve \
  --payload '{"argv":["./async_child.sh"]}'
agentprov telemetry ingest-jsonl --format tetragon --file tetragon-events.jsonl
agentprov telemetry ingest-falco --file falco-events.jsonl

ingest-jsonl records a telemetry batch manifest with the input file hash, mapped event IDs, event ID hash, receiver summary, and row-level mapping results. By default it also evaluates runtime policy for ingested events, so metadata-IP, private-CIDR, and secret-path rows become policy_decisions, risk_signals, response_actions, graph edges, and timeline rows. Use --no-policy when the receiver should only normalize and store telemetry. This gives the DAG an audit handle for external Falco/Tetragon/LoongCollector evidence without turning AgentProvenance into a long-term log store.

ingest-falco is the Falco-compatible receiver path. It reads Falco JSON/stdout from a file or stdin stream, maps recognized execve, open/openat, and connect events into normalized runtime events, correlates them by PID/container/cgroup/time evidence, and then evaluates policy by default. Raw Falco rows do not need tool_call_id; ToolCallScope is recovered from the binding table when possible.

Relationship To Existing Systems

AgentProvenance is designed to coexist with system-level observability projects, LLM tracing systems, and sandbox runtimes.

So the differentiation is not "another zero-SDK eBPF observer." The narrow primitive is:

system-side telemetry + application-side agent context
  -> evidence DAG
  -> security analysis and risk judgment
  -> automated response and audit trail

Quickstart

Prerequisites:

• Go 1.23+
• Docker Desktop or a compatible Docker daemon

git clone https://github.com/ByteYellow/AgentProvenance
cd AgentProvenance

go build ./cmd/agentprov

mkdir -p /tmp/agentprov-record-demo
printf 'value = 1\n' > /tmp/agentprov-record-demo/app.py
./agentprov record --run run-record-demo --workdir /tmp/agentprov-record-demo -- \
  sh -lc 'printf "value = 2\n" > app.py && echo artifact > artifact.txt'
./agentprov observe summary --run run-record-demo
./agentprov graph explain --run run-record-demo --file app.py

./agentprov adapter list
./agentprov adapter inspect filtered-jsonl --json
./scripts/demo_telemetry_jsonl.sh
./agentprov telemetry batches --run run-telemetry-jsonl-demo
./agentprov timeline --run run-telemetry-jsonl-demo
./agentprov timeline --run run-telemetry-jsonl-demo --view causality
./agentprov timeline --run run-telemetry-jsonl-demo --json
./scripts/accept_phase1.sh

The quick path builds agentprov, records a command, explains the changed file, ingests filtered substrate telemetry, and runs the Phase 1 acceptance gate. observe summary is the run-level observability entry point: it summarizes application context, runtime telemetry coverage, risk, baseline, response, and top evidence refs before you drill into timeline or graph queries.

demo_telemetry_jsonl.sh is the minimal substrate telemetry path. It binds a ToolCallScope, ingests Tetragon/Falco/LoongCollector fixture JSONL from examples/telemetry/, lists normalized events, and explains how one substrate event entered the DAG.

accept_phase1.sh is the machine-checkable gate for the current MVP.

timeline is the execution timeline surface. It merges application context, runtime telemetry, evidence, policy decisions, risk signals, baseline deviations, response actions, and external effects into one time-ordered view. --view causality groups rows into agent context, runtime process, runtime telemetry, evidence, risk/policy, and external-effect lanes, with correlation status and drill-down commands. The JSON output is designed to feed a future UI.

Deployment Modes

AgentProvenance is intentionally usable in three deployment shapes. RL, benchmark, and evaluator users should start with the first shape; enterprise security and audit users can move toward the later shapes when they need shared ingest, retention, and query services.

For RL and evaluator pipelines, the default contract is lightweight and offline-first:

• Install: one Go binary plus an optional thin Python package.
• Call: wrap an existing command first; SDK/framework integration is optional.
• Batch: every trajectory gets stable run_id / evidence manifest / signal context output, and query surfaces are paged.
• Overhead: default capture focuses on process/file/diff/artifact/exit/resource evidence; heavier Falco/eBPF-style telemetry is an opt-in substrate.
• Ownership: AgentProvenance emits evidence, deviation, risk, and trajectory signals. The RL system owns reward, ranking, dataset policy, and winner selection.
• Policy: RL mode does not require online deny/kill/quarantine. Those actions are opt-in security controls; offline scoring can run later over captured EvalContext JSONL.

Python usage stays thin and CLI-backed. For Deploy 1, the intended RL/evaluator entry point is one function that runs the local offline loop end to end:

from agentprov import Registry, Signal, run_batch_pipeline

registry = Registry(name="rl-reward-signals")

@registry.rule("file_change_reward")
def file_change_reward(ctx):
    return Signal.reward_feature(
        "file_change_reward",
        float(len(ctx.file_changes())),
        "reward feature from file state changes",
    )

@registry.rule("metadata_penalty")
def metadata_penalty(ctx):
    if ctx.has_event_type("metadata_ip"):
        return Signal.penalty("metadata_ip", -1.0, "metadata service access")
    return None

result = run_batch_pipeline(
    [
        {"run_id": "traj-0001", "workdir": "/tmp/job1", "command": ["pytest", "-q"]},
        {"run_id": "traj-0002", "workdir": "/tmp/job2", "command": ["pytest", "-q"]},
    ],
    registry,
    binary="./agentprov",
    data_dir=".agentprov-rl",
    engine="rl-reward-signals",
    import_signals=True,
    include_forensics=True,
)

print(result.batch_id, result.signal_count)

For RL users, custom "rules" are ordinary Python evaluator functions. They run offline over evidence; Go keeps ownership of capture, correlation, manifests, and query integrity. The same workflow can be split into lower-level calls when the pipeline already owns scheduling or sharding:

from agentprov import Client, evaluate_batch

client = Client(binary="./agentprov", data_dir=".agentprov-batch")
batch = client.record_batch(
    [
        {"run_id": "traj-0001", "workdir": "/tmp/job1", "command": ["pytest", "-q"]},
        {"run_id": "traj-0002", "workdir": "/tmp/job2", "command": ["pytest", "-q"]},
    ],
)
contexts = client.batch_eval_contexts(batch_id=batch["batch_id"])
reports = evaluate_batch(contexts, registry=registry)
client.import_signal_reports(reports, engine=registry.name)

Later, the same local store can be queried by batch, shard, job, or run:

./agentprov evidence batch-summary --latest --json
./agentprov evidence batch-summary --shard shard-0 --json
./agentprov evidence batch-summary --run traj-0001 --json
./agentprov signal batch-context --shard shard-0 --latest > eval-contexts.jsonl
./agentprov forensics export-batch --latest --json

Falco-compatible Receiver

The dedicated Falco receiver is useful when Falco is already filtering kernel or runtime events on the host:

./agentprov telemetry bind --run run-falco-demo --session session-falco-demo \
  --attempt attempt-falco-demo --tool-call tool-falco-demo \
  --process process-falco-demo --container-id container-falco-demo --pid 4242 \
  --started-at 2026-01-01T00:00:00Z

./agentprov telemetry ingest-falco \
  --file examples/telemetry/falco-risk-events.jsonl --json

./agentprov telemetry list --run run-falco-demo
./agentprov telemetry list --run run-falco-demo --limit 100 --json
./agentprov telemetry list --run run-falco-demo --limit 100 --cursor <next_cursor> --json
./agentprov timeline --run run-falco-demo
./agentprov security risks --run run-falco-demo --json
./agentprov security responses --run run-falco-demo --json

For a live stream, pipe Falco JSON output directly:

sudo falco -o json_output=true -o json_include_output_property=true | \
  ./agentprov telemetry ingest-falco --file -

The receiver maps Falco process, file, and network rows into normalized runtime events. Metadata IP, private CIDR, and secret-path rows are promoted into security evidence: RiskSignal, ResponseAction, policy graph edges, and timeline entries. graph explain --risk <policy_decision_id> --json links the risk back to the raw runtime event and forward to the response action. Falco remains the substrate collector; AgentProvenance owns correlation, causality, provenance, and risk/audit linkage.

The main smoke path is telemetry correlation and graph explanation:

./scripts/demo_telemetry_jsonl.sh

It binds a ToolCallScope, ingests raw Falco/Tetragon/LoongCollector-style runtime events that do not carry tool_call_id, normalizes them into the event store, correlates them back to application context, and explains the resulting causal graph.

Security Evidence Commands

./agentprov observe summary --run <run_id>
./agentprov observe summary --run <run_id> --json
./agentprov observe coverage --run <run_id>
./agentprov observe coverage --run <run_id> --json
./agentprov observe scopes --run <run_id>
./agentprov observe scopes --run <run_id> --json
./agentprov observe event --run <run_id> --event <event_id>
./agentprov observe event --run <run_id> --event <event_id> --json
./agentprov observe process --run <run_id> --process <process_id>
./agentprov observe process --run <run_id> --process <process_id> --json
./agentprov observe flow --run <run_id>
./agentprov observe flow --run <run_id> --json
./agentprov evidence manifest --run <run_id>
./agentprov evidence manifest --run <run_id> --json
./agentprov evidence manifest --run <run_id> --materialize --json
./agentprov telemetry correlations --run <run_id>
./agentprov telemetry correlations --run <run_id> --json
./agentprov telemetry correlations --event <event_id> --json
./agentprov timeline --run <run_id>
./agentprov timeline --run <run_id> --view causality
./agentprov timeline --run <run_id> --limit 100 --cursor <next_cursor> --json
./agentprov timeline --run <run_id> --tool-call <tool_call_id> --json
./agentprov timeline --run <run_id> --process <process_id> --json
./agentprov timeline --run <run_id> --type risk_signal --json
./agentprov security risks --run <run_id>
./agentprov security risks --run <run_id> --json
./agentprov security deviations --run <run_id>
./agentprov security deviations --run <run_id> --json
./agentprov security responses --run <run_id>
./agentprov security responses --run <run_id> --json
./agentprov baseline learn --template <template_name> --run <run_id>
./agentprov baseline check --template <template_name> --run <run_id>
./agentprov signal context --run <run_id>
./agentprov signal batch-context --batch <batch_id>
./agentprov signal batch-context --shard <shard_id>
./agentprov signal batch-context --runs runs.jsonl
./agentprov signal run --run <run_id>
./agentprov signal run --run <run_id> --json
./agentprov signal run --run <run_id> \
  --external "PYTHONPATH=python python3 examples/evaluators/python_signal_eval.py" --json
./agentprov signal import --run <run_id> --file external-signals.json --json
./agentprov signal import-batch --file signal-reports.jsonl --engine python-sdk --json
./agentprov compliance frameworks
./agentprov compliance map --framework owasp-asi --run <run_id>
./agentprov compliance explain --framework owasp-asi --run <run_id> --item ASI05
./agentprov compliance gaps --framework owasp-asi --run <run_id>
./agentprov compliance report --framework nist-rfi-2026-00206 --run <run_id>
./agentprov policy test examples/events/metadata-egress.jsonl
./agentprov policy decisions --run <run_id>
./agentprov forensics export <run_id>
./agentprov forensics export-batch --batch <batch_id>
./agentprov forensics export-batch --latest --include-eval-contexts --json

These commands are now part of the mainline security evidence surface:

External Evaluator Protocol

AgentProvenance exposes evidence to external scoring systems without owning their reward, ranking, or dataset policy.

./agentprov signal context --run <run_id> > eval-context.json

./agentprov signal run --run <run_id> \
  --external "PYTHONPATH=python python3 examples/evaluators/python_signal_eval.py" \
  --json

./agentprov signal import --run <run_id> --file external-signals.json --json

The protocol is intentionally small:

• EvalContext contains trajectories, file changes, runtime events, risk signals, and response actions.
• External evaluators read EvalContext from stdin and return { "signals": [...] }.
• EvalSignal can represent reward features, penalties, dataset labels, or quality signals.
• python/agentprov_eval and the agentprov import alias provide a thin helper SDK with Registry, @rule, evaluate_batch, reports_jsonl, and CLI-backed capture/query helpers. They do not encode a reward function.
• signal import-batch accepts JSONL EvalReport records so RL pipelines can import many offline signal reports without one command per run.

This lets a benchmark harness, RL pipeline, red-team harness, or data filtering job decide how evidence becomes score, rejection, or review.

In daemon mode, the same protocol is available through the local API:

GET  /v1/signal/context?run=<run_id>
POST /v1/signal/run
POST /v1/signal/import

The daemon does not expose an HTTP endpoint that runs arbitrary external shell commands. A client can fetch EvalContext, execute its evaluator in its own process boundary, and import the resulting signals back into the daemon for validation. The CLI follows that shape when --daemon-url is set.

Compliance Evidence, Not Certification

AgentProvenance can map run evidence to security framework profiles such as OWASP Agentic Security and NIST AI agent security assessment questions:

./agentprov compliance frameworks
./agentprov compliance frameworks --ruleset examples/compliance/custom-ruleset.yaml
./agentprov compliance validate --ruleset examples/compliance/custom-ruleset.yaml
./agentprov compliance map --framework owasp-asi --run <run_id>
./agentprov compliance map --framework owasp-asi --run <run_id> --only ASI05,ASI10,TRACE
./agentprov compliance explain --framework owasp-asi --run <run_id> --item ASI05
./agentprov compliance gaps --framework owasp-asi --run <run_id>
./agentprov compliance gaps --framework owasp-asi --run <run_id> --missing-only --json
./agentprov compliance map --framework enterprise-agent-review --ruleset examples/compliance/custom-ruleset.yaml --run <run_id>
./agentprov compliance report --framework nist-rfi-2026-00206 --run <run_id> --json

The output is an evidence-backed self-assessment. It does not certify compliance, provide legal advice, or replace qualified third-party audit. Each check item is derived from evidence already present in the run: timeline events, runtime telemetry, ToolCallScope bindings, policy decisions, risk signals, baseline deviations, response actions, forensics bundles, content-addressed provenance objects, and graph edges.

Each item reports:

covered | partial | missing | not_applicable

with concrete evidence_refs, a gap when evidence is incomplete, and a recommended next step. This makes agent execution evidence usable for security reviews without turning AgentProvenance into a GRC platform. compliance gaps turns that same mapping into an actionable backlog of missing or partial evidence items for a run.

Custom rulesets can add local frameworks and rules without replacing built-ins. The YAML model separates rules, frameworks, and mappings; mappings can also select built-in items such as ASI05, ASI10, or TRACE and reuse them inside an enterprise-specific review profile. JSON keeps control_id for compatibility and also emits item_id for the current terminology.

Graph Commands

./agentprov graph trace --run run-demo-bugfix
./agentprov graph refs --run run-demo-bugfix
./agentprov graph log --run run-demo-bugfix
./agentprov graph materialize --run run-demo-bugfix
./agentprov graph objects --run run-demo-bugfix
./agentprov graph objects --run run-demo-bugfix --limit 50 --json
./agentprov graph objects --run run-demo-bugfix --limit 50 --cursor <next_cursor> --json
./agentprov graph verify --run run-demo-bugfix
./agentprov graph verify --run run-demo-bugfix --json
./agentprov graph replay --run run-demo-bugfix
./agentprov graph replay --run run-demo-bugfix --json
./agentprov graph trajectories --run run-demo-bugfix --json
./agentprov graph diff --run run-demo-bugfix --file calculator.py
./agentprov graph diff --run run-demo-bugfix --file calculator.py --json
./agentprov graph blame --run run-demo-bugfix --file calculator.py
./agentprov graph blame --run run-demo-bugfix --file calculator.py --json
./agentprov graph explain --run run-demo-bugfix --file calculator.py
./agentprov graph explain --run run-demo-bugfix --file calculator.py --json
./agentprov graph explain --run run-demo-bugfix --file calculator.py --depth 4 --limit 200 --json
./agentprov graph explain --run run-demo-bugfix --file calculator.py --depth 4 --limit 200 --cursor <next_cursor> --json
./agentprov graph explain --tool-call <tool_call_id>
./agentprov graph explain --risk <policy_decision_id> --json

What these mean:

Current Capability

Core Demo Acceptance

The main demo must prove:

• Multiple attempts fork from the same clean snapshot.
• Raw telemetry does not need tool_call_id.
• Paged graph objects and graph explain responses expose stable result_set_id and per-page page_hash integrity metadata.
• PID, cgroup, container, and time-window bindings can resolve execution context.
• Native runtime causality records tool_call -> process -> runtime_event.
• PID/PPID/TGID telemetry creates process-tree causality edges.
• Runtime-observed file_write can appear in the same trajectory that produced a file diff.
• Runtime-observed file events create workspace_file/<path> graph nodes and can be explained together with diff/blame.
• Zero-SDK process observations expose outlived child processes and verify that orphan lifecycle evidence and policy decisions exist.
• Timeline JSON shows zero-SDK process_observed events with pid, ppid, command, first/last seen timestamps, outlived_root, and scope boundary metadata.
• Risk events can create taint and response records, but Phase 1 does not make final reward or winner decisions.
• graph diff emits unified diff and JSON.
• graph blame emits created/modified/deleted/unchanged state attribution.
• graph trajectories --json emits a structured evidence package for external evaluators.

Run:

./scripts/demo_telemetry_jsonl.sh
./scripts/demo_provenance_trace.sh

Architecture

flowchart TD
    Agent["Agent / Harness / Benchmark / Red-team / RL Pipeline"] --> CLI["agentprov CLI"]
    Agent --> SDK["Optional SDK / Tool Router"]
    Agent --> Recorder["Zero-SDK Recorder\nagentprov record -- <cmd>"]

    CLI --> Boundary
    SDK --> Boundary
    Recorder --> Boundary
    RuntimeTelemetry["Runtime Telemetry\nFalco / Tetragon / auditd / future eBPF"] --> Boundary
    SandboxIdentity["Sandbox Identity\ncontainer / cgroup / pid / cwd / time"] --> Boundary
    AppContext["Application Context\nrun / session / attempt / tool_call"] --> Boundary
    ExternalSignals["External Evaluator Signals\nreward_feature / penalty / label / quality"] --> Boundary

    subgraph Boundary["API / Ingest Boundary"]
        Daemon["Daemon API\ncontrol + query"]
        Validation["Validation / Normalization\nschema / identity / redaction"]
        Spool["Spool / Backpressure / Retention"]
    end

    Daemon --> Validation
    Validation --> Spool
    Spool --> Core

    subgraph Core["Observability + Provenance Core"]
        Correlation["ToolCallScope Correlation\npid / cgroup / container / time window"]
        Timeline["Execution Timeline\napplication context + runtime events"]
        Causality["Runtime Causality Graph\nprocess / file / network / event"]
        Provenance["Git-like Provenance DAG\nrefs / objects / diff / blame"]
        Evidence["Evidence Manifest\ncontent-addressed refs / hashes"]
    end

    Core --> Correlation
    Correlation --> Timeline
    Timeline --> Causality
    Causality --> Provenance
    Provenance --> Evidence
    Evidence --> Query["Evidence Query Surface\nobserve / timeline / explain / verify / audit"]
    Evidence --> Security["Security Analysis\nbaseline / policy / risk / response / forensics"]

Capability gating is a hard design rule. Upper layers must query the runtime, snapshot, telemetry, and isolation capabilities before assuming fast fork, memory snapshot, restore, identity, or enforcement semantics. Docker-only execution degrades to directory/filesystem provenance instead of pretending to provide VM-level resume.

All producers enter through the API/Ingest Boundary. Zero-SDK recorders, SDK/tool routers, telemetry receivers, sandbox adapters, and external evaluator signals are producer inputs; they should not bypass validation, normalization, identity binding, redaction, spool/backpressure, or retention controls to write directly into the core evidence graph.

Substrate Signals

Substrate integrations are downstream of the provenance model:

• Docker is the active local runtime.
• OpenSandbox, gVisor, Firecracker, and Kata are future runtime substrates.
• Kubernetes, Ray, Batch, and cloud systems are orchestration substrates.
• Falco, Tetragon, LoongCollector, auditd, and eBPF are telemetry substrates. The current MVP consumes already-filtered Tetragon/Falco/LoongCollector JSONL through agentprov telemetry ingest-jsonl, adds a dedicated agentprov telemetry ingest-falco receiver for Falco JSON/stdout streams, records a hashable batch manifest, and does not run kernel probes.
• A future agentprov-sensor should be capability-gated: prefer modern eBPF where the kernel and permissions support it, fall back to a legacy eBPF path where appropriate, and only use a kmod-style collector when an operator explicitly accepts that deployment model. The control plane must treat sensor capability as data, not as a hidden assumption.

The project value is not collecting more logs. The value is correlating substrate signals with execution context and making them affect diff, blame, taint, replay, and auditability.

Boundaries

These boundaries are intentional:

• AgentProvenance does not implement a general sandbox runtime.
• It does not replace Kubernetes, Ray, OpenSandbox, Firecracker, gVisor, Kata, Falco, Tetragon, LoongCollector, or eBPF.
• It is not a LangSmith clone, LLM gateway, or general observability dashboard.
• It does not promise memory snapshot or VM-level instant clone in Phase 1.
• It does not perform arbitrary branch auto-merge.
• It does not roll back real external side effects. External actions are recorded, gated, and optionally linked to compensation hooks.
• It does not make final reward, penalty, or winner decisions for RL pipelines; it emits the behavior evidence and deviation signals those systems can score.

See docs/product.md for the product direction and docs/deployment-modes.md for deployment shapes. docs/comparisons.md covers adjacent-system boundaries.

Repository Layout

cmd/agentprov/        CLI entrypoint
internal/cli/         command parsing and output

internal/record/      zero-SDK command recorder
internal/telemetry/   normalized runtime event schema, JSONL ingest, correlation inputs
internal/correlation/ ToolCallScope and runtime identity binding
internal/provenance/  timeline, graph trace, refs, objects, diff, blame, verify, replay
internal/evidence/    compact evidence records and external effects
internal/security/    policy decisions, risk signals, baseline deviations, response actions
internal/signals/     unified graph-attached signal model (behavior/cost/quality/security)
internal/baseline/    behavior baseline learning and deviation records
internal/attest/      in-toto/DSSE ed25519 evidence signing (tamper-evidence)
internal/forensics/   evidence bundle export (optional signed attestation)

internal/substrate/   Docker/runtime/snapshot adapters used as execution substrates
internal/control/     local lease/session control for substrate-backed demos
internal/computerapi/ file/tool API over local sandbox sessions
internal/ports/       local preview proxy support

internal/stressdemo/  branch-heavy fanout demos that exercise provenance under load
internal/experimental/ resource windows, scheduler, node metadata, warm-pool experiments
internal/store/       SQLite schema and repositories
examples/             tasks, events, policies
scripts/              runnable demos
docs/                 product direction, MVP details, comparisons

The main product path lives in record, telemetry, correlation, provenance, evidence, security, signals, baseline, attest, and forensics. substrate contains runtime facts AgentProvenance can consume. stressdemo and experimental are intentionally separated so branch fanout and old resource experiments do not define the project identity.

Roadmap

Recently landed (cross-cutting, ahead of full Phase 5/6):

• Unified signal model (internal/signals, agentprovenance.signals/v1) - one graph-attached row type for behavior/cost/quality/security, replacing the per-dimension silos; security and quality are live producers, with idempotent backfill from legacy tables.
• Signed evidence attestation (internal/attest) - in-toto/DSSE ed25519 signing of evidence digests, wired into forensics export, giving post-compromise tamper-evidence rather than integrity-only hash recompute.
• Concurrency correctness - SQLite pragmas (incl. busy_timeout) applied via DSN to every pooled connection, fixing silent SQLITE_BUSY under concurrent writers; correlation bindings bound stale-open over-matching.

Near-term hardening:

• Deeper graph integrity checks for process-tree and file-event causality.
• Realistic zero-SDK acceptance for file modification, file creation, file deletion, child process observation, delayed runtime-event correlation, diff/blame, timeline, evidence manifest, replay, and graph verification.
• Realistic Falco risk acceptance for raw execve, metadata-IP, private-CIDR, and secret-path rows becoming correlated telemetry, risk signals, response actions, graph explanations, evidence manifests, and verified DAG state.
• Forensics bundle acceptance for risky runtime streams, including bundle sha256 verification and embedded evidence/risk/response/graph-edge checks.
• Daemon evidence API acceptance for the same risk path through HTTP endpoints, proving the first control/query boundary without making the CLI own every lifecycle.

Development

go test ./...
./scripts/accept_phase1.sh
./scripts/accept_zero_sdk_realistic.sh
./scripts/accept_falco_risk_realistic.sh
./scripts/accept_forensics_bundle.sh
./scripts/accept_batch_forensics.sh
./scripts/accept_evidence_query_pagination.sh
./scripts/accept_daemon_evidence_api.sh
./scripts/accept_telemetry_spool_backpressure.sh
./scripts/accept_telemetry_100k_pressure.sh
./scripts/accept_telemetry_event_windows.sh
./scripts/accept_signal_engine.sh
./scripts/accept_python_helper.sh
./scripts/accept_deploy1_batch_pipeline.sh

The acceptance scripts are the main machine-checkable gates for Phase 1 observability and provenance correlation. accept_phase1.sh validates the cross-layer telemetry path. accept_zero_sdk_realistic.sh validates a more realistic no-SDK command path with process-tree observation, delayed event correlation, file diff/blame, evidence materialization, replay, and graph verification. accept_falco_risk_realistic.sh validates the system-side security path from Falco-style runtime rows to risk, response, explain, evidence, and graph verification. accept_forensics_bundle.sh validates that a risky runtime stream can be exported into a hashed forensics bundle with embedded evidence manifest, risk/response records, graph edges, and a clean graph verify result. accept_batch_forensics.sh validates that a recorded batch can be exported into a sha256-verified batch audit bundle with per-run bundle refs, optional EvalContext records, summary hashes, and replay/query commands. accept_daemon_evidence_api.sh validates the daemon API path for ToolCallScope binding, async Falco spool ingest, control API responsiveness while a batch is queued, graph verify, evidence manifest materialization, and forensics export. accept_python_helper.sh validates the thin Python helper path for CLI-backed record, batch record, evidence manifest export, EvalContext export, batch forensics export, and signal import. accept_telemetry_event_windows.sh validates that Falco ingest automatically builds 10s/60s telemetry event windows with resolved and high-risk counts. accept_daemon_evidence_api.sh also validates daemon GET /v1/telemetry/windows and the window aggregate embedded in observe summary. accept_deploy1_batch_pipeline.sh validates the one-call Deploy 1 Python flow: record a batch, export EvalContext records, run registered evaluator functions, import EvalSignal reports, export batch forensics, and return a summary. accept_evidence_query_pagination.sh validates timeline cursor pagination, daemon-client cursor propagation, and evidence manifest query refs with result/page hashes.