Sakshi

Sakshi (साक्षी) is the Vedantic name for the consciousness that observes consciousness without acting.

A metacognitive runtime for Python agents: the Witness pattern. A watching process embedded in your agent that monitors plan execution, evaluates against expectations, and steers cognition without doing the cognition itself.

Docs: https://bionicbutterfly13.github.io/sakshi/

Where Sakshi fits (vs LangGraph, AutoGen, raw event loops): see docs/positioning.md.

Status

0.14.0 pre-1.0. The typed Witness surface is present and package-local tests cover the main DTO, goal, plan, interpret, meta, motivation, trust, uncertainty, guard primitives, and reference integrations. The public API may still change in 0.x minor releases; production hosts must verify their adapters before relying on a new minor version.

⚠ The defaults are inert. NoOpEventBus, NoOpBasinHook, and AlwaysPermitWriteGuard are intentional no-ops for tests and the quickstart below. A production host must inject real implementations of EventBus, GoalStateStore, and WriteGuard — otherwise events drop on the floor, world state never resolves, and every Sakshi-originated write is silently permitted. See sakshi.protocols for the interfaces and your host's adapter layer for examples.

For production scaffolds that have not yet finished wiring real host implementations, prefer the deny-by-default safe seams over the permissive defaults: DenyByDefaultWriteGuard (in sakshi.protocols) and DenyByDefaultPolicy (in sakshi.meta). They fail closed — every write or intervention is denied with an explicit audit reason — so a half-finished integration cannot silently permit unsafe side effects.

Install

Sakshi is published to PyPI under the distribution name pysakshi. The importable module is still sakshi:

pip install pysakshi

import sakshi

The bare name sakshi is reserved by an unrelated PyPI account with no releases; once that name is recovered, pysakshi will become an alias.

To install the latest unreleased commit from source:

pip install git+https://github.com/bionicbutterfly13/sakshi

Release process: see docs/release-checklist.md. 1.0 stability commitment: see docs/1.0-stability.md.

Reference integration

Two examples live under examples/:

• toy_blocks_agent/ — a stdlib-only 3-phase plan/act/observe loop on a tiny blocks world. The smallest concrete answer to "how do I wire Sakshi into an existing agent?"
• llm_research_agent/ — the bigger sibling: cognition is an actual Anthropic Claude model (deterministic mock available for tests). Adds dynamic GoalGraph decomposition, typed low-confidence anomaly recovery, and a WriteGuard-gated publish step. Install with the llm-examples extra to enable the live path: pip install pysakshi[llm-examples].

What it is

The Witness pattern means the host keeps control of cognition while Sakshi observes it through typed seams. The host owns the planner, memory system, tool runtime, world model, and side effects. Sakshi receives declared expectations and runtime records, checks what happened against those expectations, and emits typed signals a host can route to dashboards, gates, audits, or meta-control policies.

That boundary is useful because it makes an agent inspectable without making Sakshi the agent's brain. Hosts can add contract monitoring, confidence calibration, goal lineage, motivation envelopes, risk records, trust repair recommendations, and defensive guards while keeping their domain logic and runtime dependencies outside the package.

Sakshi exposes a small, opinionated set of seams a host application supplies:

• An EventBus for emitting cycle and goal events.
• A Clock for deterministic time.
• A GoalStateStore for fetching world state and recording goal outcomes.
• An optional BasinHook for hosts that maintain an attractor-basin field.
• A WriteGuard for routing Sakshi-originated writes through the host's safety policy.

Around these protocols Sakshi assembles a generic six-phase cognitive cycle (PERCEIVE, INTERPRET, EVAL, INTEND, PLAN, ACT) and a meta-loop (MONITOR, ASSESS, CONTROL) framed as general metacognitive phases.

Quickstart

import asyncio

from sakshi.registries import PhaseRegistry
from sakshi.protocols import NoOpEventBus


async def main() -> None:
    registry = PhaseRegistry(event_bus=NoOpEventBus())

    await registry.start_cycle("cycle-001")
    await registry.record_phase_output("PERCEIVE", {"basins": ["attention"]})
    await registry.record_phase_output("INTERPRET", {"confidence": 0.82})
    trace = await registry.finalize_cycle()

    print(trace.cycle_id)
    print([phase.phase_name for phase in trace.phase_results])


asyncio.run(main())

Hosts that need production side effects implement the small protocols in sakshi.protocols and pass those implementations into Sakshi runtime classes.

Influences

Sakshi draws inspiration from MIDCA-style metacognitive control loops, active-inference monitoring, and production cognitive-runtime experience. See INSPIRATION.md for the full provenance.

Sakshi is not a MIDCA implementation, MIDCA-compatible, or a fork of any MIDCA codebase. The MIDCA reference architecture is one influence among several.

License

Apache-2.0. See LICENSE.