Pisama

Find and fix failures in AI agent systems. No LLM calls required.

Pisama ships heuristic detectors that apply across frameworks including n8n, LangGraph, Dify and OpenClaw, with per-platform gating (for example coordination runs only on multi-agent platforms). They run locally with zero LLM cost on the heuristic tier.

Install

pip install pisama

Hosted first diagnosis

The basic Usage example below runs locally. A founder-issued Pisama Cloud API key is for the hosted service, not a requirement for offline analyze(). If you arrived here after redeeming an invitation, use this section first.

1. Save the one-time API key in your secret manager. Invitation redemption does not create dashboard access. Never paste the key into an issue, shared trace, source file, or chat.
2. Export one representative run as OpenTelemetry JSON (resourceSpans) and remove credentials and data you are not authorized to upload. Hosted ingestion stores the run and counts toward your agreed usage allowance.
3. Save the following as hosted_first_diagnosis.py and run python3 hosted_first_diagnosis.py. It uses only Python's standard library, prompts for the key without echoing it, and never prints the bearer token. It sends your selected file to Pisama Cloud; it is not an offline example.

import getpass
import json
from pathlib import Path
from urllib.error import HTTPError, URLError
from urllib.request import Request, urlopen

BASE = "https://api.pisama.ai/api/v1"


def post(path, payload, token=None):
    headers = {"Content-Type": "application/json"}
    if token:
        headers["Authorization"] = f"Bearer {token}"
    request = Request(
        BASE + path,
        data=json.dumps(payload).encode("utf-8"),
        headers=headers,
        method="POST",
    )
    try:
        with urlopen(request, timeout=120) as response:
            return json.load(response)
    except HTTPError as error:
        # Do not dump request headers or response bodies containing private data.
        raise SystemExit(f"{path}: HTTP {error.code}; stop and check access/format.")
    except URLError:
        raise SystemExit(f"{path}: network failure; do not blindly retry ingestion.")


trace_path = Path(input("Path to your redacted OTLP JSON file: ").strip())
trace = json.loads(trace_path.read_text(encoding="utf-8"))
if not isinstance(trace, dict) or not isinstance(trace.get("resourceSpans"), list):
    raise SystemExit("Expected an OTLP JSON object containing resourceSpans.")

key = getpass.getpass("One-time Pisama Cloud API key: ").strip()
auth = post("/auth/token", {"api_key": key, "scope": "full"})
del key
token = auth.get("access_token")
if not isinstance(token, str) or not token:
    raise SystemExit("No bearer token returned; stop and contact Pisama.")

ingest = post("/traces/ingest", trace, token)
print("Ingestion:", {name: ingest.get(name) for name in ("accepted", "rejected", "traces")})
if not ingest.get("accepted") or ingest.get("rejected"):
    raise SystemExit("Ingestion was empty or partial; inspect the export before continuing.")

result = post("/diagnose/why-failed", {
    "content": json.dumps(trace), "format": "otel", "include_fixes": False,
}, token)
del token
print("Failure signals:", result.get("failure_count"))
for finding in result.get("all_detections", []):
    print(json.dumps({name: finding.get(name) for name in (
        "category", "mistake_agent", "affected_spans", "evidence", "suggested_fix",
    )}, indent=2))

Ingestion accepts work asynchronously. The diagnosis request above analyzes the submitted content; it does not prove that background analysis of the stored run has completed. Review the reported agent, spans, evidence and next action against your run. Zero signals is not proof of success, and a suggested fix is not proof that the task will work after a change. include_fixes=False avoids requesting optional generated fixes; it does not promise that every hosted detector is free of model calls or usage charges. Hosted scope, retention and pricing are defined in your founder-led agreement, not this package's MIT license.

For 401, obtain a fresh token using an active key. For 403, check your invitation, scope and entitlement with Pisama. For 429, stop and check quota or rate limits. Do not repeatedly upload the same run to work around an error. Contact team@pisama.ai if the flow cannot produce an inspectable finding. Do not send the key or an unredacted trace by email.

Usage

Local analyze() accepts a nonempty Pisama native trace (spans), an ATIF trajectory, or native span JSONL. OTLP resourceSpans exports are not a local input format; use the hosted workflow above for those exports. Unsupported or empty inputs raise an error rather than reporting a clean analysis. JSONL may contain native span rows, or exactly one native trace envelope; mixing envelopes with other rows is rejected rather than dropping later evidence.

from pisama import analyze

result = analyze("trace.json")  # also accepts dicts and JSON strings

for issue in result.issues:
    print(f"[{issue.type}] {issue.summary} (severity: {issue.severity})")
    print(f"  Fix: {issue.recommendation}")

CLI

Validated response-contract accounting, when the core supplies it, is preserved under each detector assessment. Version 1 includes positional span records and checked/unsupported/outside-scope counts for eligible response pairs only. Raw span IDs and arbitrary metadata are omitted. Invalid schemas, indices, counts or enum values are rejected with response_coverage_status="invalid"; they never establish whole-trace coverage. Business semantics remain unassessed.

AnalyzeResult.detector_assessments and JSON CLI output preserve explicit detector coverage. abstained is not a passed check; older core versions that provide no assessment are marked unknown. Terminal output separates explicit contract passes, abstentions, errors and unspecified coverage. Detector execution counts are not counts of validated requirements. No findings is not proof of task success. Coverage metadata does not turn heuristic confidence into calibration. check --json schema 3 includes per-file assessments and assessment_reporting_complete: one identified report per executed detector. That flag includes explicitly unknown/abstained reports and does not describe trace coverage. Duplicate/missing detector identities make reporting incomplete. trace_coverage is currently unassessed; files_clean remains zero because core does not provide whole-trace/span/contract coverage accounting. Even a detector reporting a satisfied contract may have skipped other requests. passed remains a severity-threshold gate, not complete validation; explicit detector errors fail even with --fail-on never. Unknown coverage and abstention are not detector errors but are never counted as checked-clean files. Replay comparisons do not label disappeared findings fixed: current assessment metadata lacks comparable contract/input provenance, even when a later check passes.

pisama analyze trace.json          # Analyze a trace
pisama watch python my_agent.py    # Watch a live agent (pip install "pisama[auto]")
pisama replay <trace-id>           # Re-run detection on stored traces
pisama smoke-test --last 50        # Batch test recent traces
pisama detectors                   # List all core detectors
pisama mcp-server                  # Start MCP server (pip install pisama[mcp])

MCP Server

Works in Cursor, Claude Desktop, and Windsurf. No API key is needed:

{
  "mcpServers": {
    "pisama": { "command": "pisama", "args": ["mcp-server"] }
  }
}

Optional extras

The base pisama install has zero-cost heuristic detection covered. Two extras add opt-in functionality on top.

Auto-instrumentation: pisama[auto]

Zero-code tracing for LLM calls. init() patches supported clients (Anthropic, OpenAI) so every call after it emits an OTEL trace Pisama can analyze, no manual instrumentation needed.

pip install "pisama[auto]"

import pisama.auto

pisama.auto.init(api_key="ps_...")

# All subsequent LLM calls are automatically traced
import anthropic
client = anthropic.Anthropic()
response = client.messages.create(...)  # traced automatically

This used to require the standalone pisama-auto package. That package still works and stays fully supported for existing installs; pisama[auto] is the same code, folded into the base package so there is one less dependency to track. New projects should install it this way.

Agent hooks and tools: pisama[agents]

Real-time hooks, tools, and self-check utilities for agent runtimes (built for the Claude Agent SDK), wired to Pisama's detection infrastructure for in-loop failure prevention rather than after-the-fact analysis.

pip install "pisama[agents]"

from pisama.agents import pre_tool_use_hook, post_tool_use_hook

agent.hooks.pre_tool_use = pre_tool_use_hook
agent.hooks.post_tool_use = post_tool_use_hook

Active self-check is available the same way:

from pisama.agents import check

result = await check(
    output="The server is healthy based on the metrics.",
    context={"query": "Is auth-service down?", "sources": [...]},
)
if not result["passed"]:
    ...  # revise output based on result["issues"]

This used to require the standalone pisama-agent-sdk package. That package still works and stays fully supported for existing installs; pisama[agents] is the recommended path for new projects, one package instead of two.

Detectors

Core detectors, gated per platform (n8n, LangGraph, Dify, OpenClaw and others). A representative selection:

Detector	What It Catches
loop	Infinite loops, retry storms, stuck patterns
coordination	Deadlocked handoffs, message storms
hallucination	Factual errors, fabricated tool results
injection	Prompt injection, jailbreak attempts
corruption	State corruption, type drift
persona_drift	Persona drift, role confusion
derailment	Task deviation, goal drift
context	Context neglect, ignored instructions
specification	Output vs. requirement mismatch
communication	Inter-agent message breakdown
decomposition	Poor task breakdown, circular dependencies
workflow	Unreachable nodes, missing error handling
completion	Premature completion, unfinished work
withholding	Suppressed findings, hidden errors
convergence	Metric plateau, regression, thrashing
overflow	Context window exhaustion
propagation	Silent error propagation across steps
citation	Fabricated citations and source misattribution
routing	Inputs misrouted to the wrong specialist agent
mcp_protocol	MCP tool-communication failures

Links

• Offline usage
• Hosted first diagnosis
• GitHub
• Platform

License

MIT

Source boundary

This repository is the public source for the MIT-licensed pisama Python package. It does not contain the Pisama Cloud backend, dashboard, calibration data, managed detection tiers, or paid automation.