More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Production Incident Investigator

    1

    Turn exported logs and deployment events into a source-linked timeline, competing incident hypotheses and a blameless action plan, without connecting production accounts.

    $15

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Production Incident Investigator

    Production Incident Investigator

    Example session with this skill installed

    Use the three Production Incident Investigator skills on the supplied synthetic fixtures: fixtures/webhook.jsonl, fixtures/receiver.log, fixtures/deployments.csv and fixtures/changes.gitlog, in that order. Write intake.json, timeline.md and postmortem.md to a fresh report directory. Keep source line references, explain the out-of-order warnings, compare the webhook rollout and retry/idempotency hypotheses, and leave unknown owners as TBD. Do not connect production services or perform remediation. Return the complete generated timeline and review.

    • Read your context and instructions
    • Compiled the production incident investigator

    Incident timeline and hypotheses

    Offline evidence review. Timestamps are UTC; original ordinals break ties. Source references use the intake input-position map.

    Confirmed observations

    • deploy: 2 exported records (source-03:3, source-03:2).
    • duplicate_error: 2 exported records (source-01:1, source-01:4).
    • git_change: 1 exported records (source-04:1).
    • http_200: 2 exported records (source-02:1, source-01:2).
    • http_500: 3 exported records (source-01:5, source-02:2, source-01:6).
    • other: 1 exported records (source-01:8).
    • retry: 2 exported records (source-01:3, source-02:3).
    • timeout: 1 exported records (source-01:7).

    Counts describe this export only. No request denominator, error rate, customer count or full outage duration is established.

    UTCSource:lineOriginal ordinalServiceObservation
    2026-10-01T09:50:00Zsource-04:114gitgit_change
    2026-10-01T09:59:50Zsource-02:19webhookhttp_200
    2026-10-01T09:59:55Zsource-03:313analyticsdeploy
    2026-10-01T09:59:58Zsource-01:11webhookduplicate_error
    2026-10-01T10:00:00Zsource-03:212webhookdeploy
    2026-10-01T10:00:01Zsource-01:22webhookhttp_200
    2026-10-01T10:00:03Zsource-01:33gatewayretry
    2026-10-01T10:00:04Zsource-01:44webhookduplicate_error
    2026-10-01T10:00:04.500000Zsource-01:55webhookhttp_500
    2026-10-01T10:00:05Zsource-01:77analyticstimeout
    2026-10-01T10:00:06Zsource-02:210webhookhttp_500
    2026-10-01T10:00:07Zsource-01:66webhookhttp_500
    2026-10-01T10:00:08Zsource-02:311gatewayretry
    2026-10-01T10:00:09Zsource-01:88webhookother

    Competing hypotheses

    H1: A rollout-related change in analytics may contribute to that service's errors.

    Confidence: Low: same-service errors occur within 15 minutes after a deploy; no code or controlled comparison establishes cause.
    Supporting observations: source-03:3, source-01:7
    Counterevidence / caveats: No pre-deploy errors in this export; baseline coverage is unknown.
    Suggested check (not executed): Compare old/new version cohorts and review the approved change diff; test a replay in an isolated environment.

    H2: A rollout-related change in webhook may contribute to that service's errors.

    Confidence: Low: same-service errors occur within 15 minutes after a deploy; no code or controlled comparison establishes cause.
    Supporting observations: source-03:2, source-01:4, source-01:5, source-02:2, source-01:6
    Counterevidence / caveats: Errors also precede this deploy: source-01:1
    Suggested check (not executed): Compare old/new version cohorts and review the approved change diff; test a replay in an isolated environment.

    H3: Redelivery plus an idempotency or acknowledgement gap may permit duplicate side effects.

    Confidence: Moderate for a retry/duplicate association: matching pseudonymous correlation IDs; low for the unverified internal mechanism.
    Supporting observations: source-01:1, source-01:3, source-01:4
    Counterevidence / caveats: Correlation IDs may be reused; logs do not establish transaction ordering, delivery identity or whether the side effect committed.
    Suggested check (not executed): Inspect delivery IDs and transaction/acknowledgement boundaries in an approved export; reproduce concurrent redelivery locally.

    Service boundaries and evidence gaps

    • Shared correlation across services: source-01:1, source-01:3, source-01:4
    • A shared correlation value connects records, not proven network calls or causal edges.
    • Clock skew, sampling, missing exports, duplicated log records and correlation reuse remain unverified.
    • Commits are change context only; no commit is assumed deployed. Deployment proximity does not establish root cause.
    • Intake warnings: 2. See intake.json for source/line and diagnostic codes.

    Containment options requiring human approval

    • If duplicate processing is active, consider pausing the affected consumer after assessing backlog and contractual impact.
    • Consider a staged rollback only after checking version-specific error cohorts and rollback/data compatibility.
    • No containment, production query, code edit or ticket creation was executed.

    $ python -B scripts/investigate.py postmortem --input run-a/intake.json --out run-a/postmortem.md
    exit=0

    Blameless incident review — draft for human validation

    Observed scope: 6 exported error-class records (source-01:1, source-01:4, source-01:5, source-01:7, source-02:2, source-01:6). Records may describe the same request; this is not a count of affected requests.
    Customer impact, detection time, mitigation time and recovery status: unknown from the supplied evidence.
    Root cause: unconfirmed. This report identifies system behaviors to investigate, without attributing blame.

    Evidence and competing explanations

    • H1: A rollout-related change in analytics may contribute to that service's errors. Support: source-03:3, source-01:7. Low: same-service errors occur within 15 minutes after a deploy; no code or controlled comparison establishes cause.
      Caveat: No pre-deploy errors in this export; baseline coverage is unknown.
    • H2: A rollout-related change in webhook may contribute to that service's errors. Support: source-03:2, source-01:4, source-01:5, source-02:2, source-01:6. Low: same-service errors occur within 15 minutes after a deploy; no code or controlled comparison establishes cause.
      Caveat: Errors also precede this deploy: source-01:1
    • H3: Redelivery plus an idempotency or acknowledgement gap may permit duplicate side effects. Support: source-01:1, source-01:3, source-01:4. Moderate for a retry/duplicate association: matching pseudonymous correlation IDs; low for the unverified internal mechanism.
      Caveat: Correlation IDs may be reused; logs do not establish transaction ordering, delivery identity or whether the side effect committed.

    The full source-linked timeline is in timeline.md; retain intake.json alongside it.

    Action plan — proposed, not created as tickets

    PriorityTask and evidenceOwnerDueAcceptance criterion
    P1Verify atomic deduplication under concurrent redelivery; source-01:1, source-01:4TBDTBDIsolated replay yields one committed side effect per delivery ID, including a crash before acknowledgement.
    P1Compare version cohorts before choosing rollback; source-03:3, source-03:2TBDTBDDocument cohort denominators and errors over equal windows; record whether change-specific impact is supported.
    P2Validate clocks, sampling and impact denominator; source-04:1TBDTBDRecord export coverage, clock offset bounds and total requests; calculate impact only with a verified denominator.
    P2Review findings and assign accountable ownersTBDTBDA human validates each causal claim, assigns owners/dates and records unresolved gaps.

    Follow-up and containment

    Containment options appear in timeline.md and require human approval. No remediation or additional investigation has been performed.
    Next: validate evidence coverage, run the proposed isolated checks, then update this draft with confirmed results and citations.

    2. See intake.json

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Reconstruct the evidence before writing the explanation

    An incident can leave you with several exports, conflicting timestamps and a plausible story that is hard to verify. A rollout appears near a burst of errors. A retry looks like a second failure. An unrelated service logs a timeout at almost the same moment. Turning those fragments directly into polished prose can hide the questions you still need to answer.

    Production Incident Investigator gives your coding agent a repeatable offline workflow for checking the evidence first. Supply supported log and deployment exports, choose a report directory, and receive a normalized evidence file, a source-linked timeline and a blameless review with concrete follow-up tasks. You keep the source files and decide what happens next.

    Three focused skills, one investigation

    Incident evidence intake validates timestamps and record shapes, normalizes timezone-aware dates to UTC and reports malformed or unordered evidence. It preserves physical source line references and original ordering. Rejected evidence blocks downstream analysis instead of disappearing silently.

    Timeline and hypotheses sorts accepted events, counts the observed categories and compares candidate explanations. Same-service errors near a deployment can suggest a rollout-related investigation. Matching pseudonymous correlation values can connect retry and duplicate-error records across service labels. Each hypothesis explains its qualitative confidence, supporting references, counterevidence and a check that could help resolve it.

    Postmortem action plan turns accepted evidence into a review draft with proposed priorities and testable acceptance criteria. Owners and dates remain TBD unless you provide them. It does not invent customer impact, recovery time or a confirmed root cause.

    What you receive

    The pack includes exactly three SKILL.md workflows, three small Python scripts, an umbrella README, input and safety references, a version file, an MIT license and synthetic examples with expected findings and negative cases. All required product resources are text. The scripts use Python's standard library and run in your agent session; no permanent plugin, local server or dependency bundle is installed.

    The output files are intake.json, timeline.md and postmortem.md. Timeline entries contain UTC timestamps, source aliases and physical line numbers. The review includes tasks such as testing atomic deduplication under concurrent redelivery, with an explicit success criterion rather than a vague instruction to improve reliability.

    A small demo with a useful trap

    The included synthetic webhook incident contains duplicate-side-effect exceptions, retries, three HTTP 500 records, deployment events and a Git change. A duplicate exception occurs before the webhook rollout. The implementation preserves that observation as counterevidence to a rollout-only explanation. A separate analytics deployment and timeout demonstrate why service boundaries matter. The complete executed local demo is shown below; these are invented test inputs, not a customer incident.

    Who this is for

    This pack suits software engineers and small teams investigating bounded exported evidence, preparing an incident review or checking an explanation before sharing it. It is particularly useful when you want citations and reproducible parsing without setting up a monitoring integration. It is not a replacement for live observability, a forensic acquisition system or an autonomous incident responder.

    Requirements and boundaries

    You need a coding agent that can read and write selected files and run Python 3.10 or later. Local execution was tested in Codex on Windows with Python 3.14.4. Agensi delivery, other agents and other platforms have not been tested. Supported inputs are the documented UTF-8 JSONL schema, one single-line plain-text log format, a simple deployment CSV and optional explicitly supplied Git log text. Vendor-specific exports may need conversion before use.

    The scripts make no network requests and do not query production, deploy, roll back or create tickets. Raw messages and payload fields are withheld from reports. Correlation hashes remain pseudonymous data, and service identifiers must already be non-personal. Review outputs locally before sharing; this is not a universal secret or personal-data detector. Your agent provider handles supplied content under your existing account settings.

    Classification and correlation use documented heuristics. Missing exports, clock skew, reused IDs, quoted error text and sampling can mislead the analysis. Counts describe exported records, not unique failed requests or error rates. Causal explanations and priorities need human review.

    Start with the included fixtures, inspect the cited results, then apply the same workflow to a small, approved export from your own incident.

    • Narrow documented JSONL/plain-log/CSV/Git export formats; no multiline stack traces or automatic vendor adapters.
    • Maximum 32 selected files, 10 MB per file and 50,000 input records; bounded exports only.
    • Textual event classification can misread quotes or negations and miss unfamiliar errors.
    • One category per record; no denominator, unique failure count, impact rate or recovery calculation.
    • Correlation links are not proven calls. Rollout proximity is not proof of causality; candidate checks need validation.
    • Unknown free text and versions are withheld, limiting automated diagnostics; inspect original sources locally when approved.
    • Clock skew, sampling and source authenticity are not repaired or certified.
    • Personal service identifiers can survive label validation. Correlation hashes are guessable for low-entropy IDs. No universal PII/secret guarantee.
    • Parent path symlinks/junctions and concurrent filesystem attacks are outside the hardening scope; use trusted report directories.
    • No production service, autonomous remediation, ticketing, monitoring subscription or hosted dashboard.
    • Tested in local Codex/Windows/Python 3.14.4 only; no marketplace-delivery or broad-agent compatibility claim.
    • MIT license permits reuse, modification and redistribution; purchase pays for the prepared pack, not exclusive rights.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 21 days ago

    Requires

    Python 3.10+
    Terminal access
    Python
    Git

    Permissions

    Terminal / Shell
    Read Files
    Write Files

    File Scopes

    <selected-evidence>/*.jsonl
    <selected-evidence>/*.log
    <selected-evidence>/*.csv
    <selected-evidence>/*.gitlog
    <temporary-scripts>/investigate.py
    <temporary-scripts>/evidence.py
    <temporary-scripts>/reasoning.py
    <selected-reports>/intake.json
    <selected-reports>/timeline.md
    <selected-reports>/postmortem.md
    <selected-reports>/agent-review.md

    Reads only the supplied offline exports and its own support files. Writes new reports to the selected directory and temporary unchanged scripts for in-session execution. Raw messages, unknown payload fields, email/token/payment strings embedded there, commit subjects, versions and raw correlation identifiers are withheld. Correlation hashes are pseudonyms and service names must already be non-personal; review locally before sharing. The implementation uploads no evidence and sends nothing to the creator. Content processed by the coding agent remains subject to the user's existing agent-provider configuration. No telemetry or external service is part of this product. Explicit approval is required before connections, external browsing, installs or commands with material side effects. There is no remediation integration, production query or ticket creation. Human review of privacy and proposed causal explanations remains necessary.

    Listed21 days ago

    What's inside

    Frequently Asked Questions

    The skill requires local Python 3.10+ and accepts standard offline formats like JSONL logs, CSV deployment exports, and Git text logs. No external database or cloud connection is needed, ensuring your data remains in your controlled environment.

    Unlike generic AI prompts that may hallucinate timestamps, this skill uses a deterministic Python intake process. It anchors every event in the generated timeline to a specific line in your raw evidence, ensuring the resulting postmortem is verifiable and accurate.

    The investigator includes a built-in privacy-gating step that strips PII and sensitive payloads during the intake phase. It normalizes all data to UTC and uses pseudonymous IDs to maintain a blameless, security-conscious analysis environment.

    The package contains the core Python intake scripts, the three-stage analysis workflow (Intake, Analysis, Synthesis), and standardized templates for generating cited timelines and blameless postmortem reports.

    It is designed for post-incident reconstruction using offline exports rather than live monitoring. It does not perform automated rollbacks or integrate directly with live ticketing systems, focusing instead on high-fidelity reporting and root cause analysis.

    New in Testing & Debugging