witness — read your growth archive

    1

    Teaches your agent when and how to read a local witness archive: a dated history of how its subject changed.

    Free

    1 useSecurity scanned

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    witness — read your growth archive

    Example session with this skill installed

    What does my witness archive actually record about how I verify work?

    • Read your context and instructions
    • Compiled the witness

    Searched the default lens with search_observations. The tool returns matches
    most-relevant-first and stamps each with a timestamp, so I re-sorted by date — the ordering
    below is the archive's, not my inference.

    What is on the record

    • [2026-07-03T06:51:38Z · habits] Defaults to an independent, adversarial verification pass
      over work that has already been reported finished and green.
      evidence: "now run full parallel adversarial agents to verify our work and the code repo's
      status against best practices."
      id: obs_5590d966b6b9
    • [2026-07-03T16:02:40Z · feedback] Does not take "fixed, verified, CI-green" claims at face
      value; re-inspects the live result and pushes back when it doesn't satisfy him.
      evidence: After the assistant declared both fixes done and verified, he returned to reject
      the pill fix on sight.
      id: obs_cdc9dad892ea
    • [2026-08-27T12:36:07Z · biases] Named his own over-reverence for golden/reference fixtures
      as a bias and issued a standing correction against it, rather than waiting to be talked out
      of it.
      evidence: "twice. im too worried about goldens. they are not perfect. we are improving the
      world so be brave on changing goldens"
      id: obs_eb51fc2420a6

    Reading it

    Seven weeks separate the first two records from the third. In early July the pattern is a
    refusal to accept "done and verified" without re-checking it personally. By late August the
    same scepticism has been turned on the habit itself — over-protection of reference fixtures
    named as the defect, with a standing correction issued rather than a one-off waiver.

    Every line carries its own date and its own evidence quote, so any of it can be checked.
    get_facets on this lens returns the present-tense synthesis of the same material, e.g.

    • habits/gates_golden_baseline_regeneration_on_explicit_signoff (confidence 0.55): …

    What I did not do

    • I did not claim the July records caused the August correction. They are dated and
      adjacent; nothing in the archive records a causal link, so I am not supplying one.
    • Observations are the dated layer. Facets are current-state — get_facets returns what holds
      now, not superseded values. For full version history, witness export --all writes the
      whole archive and every prior value is in it.
    • This is the default lens only.

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

    About this skill

    What this does

    witness is a local, pull-only archive that distills any text stream — your coding sessions, or a corpus you feed it — into a dated history of how its subject changed. A fact whose value changes isn't overwritten: the old value is closed with a date and a successor opens, so the archive answers how did this get here, not just what is true now.

    witness never pushes anything into a session. That's deliberate, and it means your agent has to decide to read it. This skill is that decision: the trigger conditions, the right tool for each question, and how to read a bi-temporal result without flattening away the history.

    What's inside

    • When to read it — questions about yourself, advice that should be personalized, retrospectives, vague references to past work, and the non-person corpus case.
    • When not to — it isn't project memory, and writing to it by hand is almost always wrong.
    • Which tool to reach for — get_profile, get_facets, search_observations and the two write tools, with what each is actually good for.
    • Reading results well — facets outrank the generated narrative, quote the change rather than only the current value, cite the evidence each facet carries, filter by lens, and treat an empty archive as empty instead of inferring.

    Deliberately version-agnostic: it defers to the live MCP tool descriptions for exact arguments, so it stays correct across witness releases.

    Requirements

    witness installed locally and wired into your agent — a single Go binary, no runtime or container, MIT licensed. Install and setup:

    https://github.com/IngTian/witness

    This skill teaches an agent to use that install; it does not install anything itself.

    Requires a local witness install with its MCP server configured in your agent. This skill is
    guidance only — it installs nothing.

    A new archive has little to say. Observations are mined from your sessions as you work, and
    facets appear only after a review pass, so expect thin answers for the first while.

    Pull-only by design: witness never injects anything into a session, so your agent has to
    choose to read it.

    It is not project memory. It records things about the subject, not what your code does.

    Coverage follows your lenses — a domain you have not written a lens for will not appear.

    It reports what changed, when, and on what evidence. Not why: causation is not stored.

    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

    1 use

    Times used to date

    Free forever

    No account required to browse

    Trust & safety

    Security scan

    Security scan passed 28 days ago

    • Free with an Agensi account

    Requires

    witness

    Permissions

    Read Files
    Network Access
    Listed28 days ago

    What's inside

    Frequently Asked Questions

    Most memory tools store the current state of a fact and overwrite it when it changes. witness closes the old value with a date and opens a successor, so the archive answers "how did this get here", not only "what is true now". This skill teaches your agent to read that history — to quote the change and its evidence instead of flattening it into a current-state summary.

    A local witness install with its MCP server configured in your agent. witness is a single Go binary for macOS, Linux and Windows with a SQLite archive. There is no daemon to start: your agent launches the MCP server over stdio when it needs it.

    Reading an archive works from any agent that speaks MCP. Filling one does not — witness captures sessions from Claude Code and OpenCode specifically, through their hooks and plugin. So another agent can query an archive, but Claude Code or OpenCode has to be what populates it.

    Your archive fills itself: witness runs a background worker that mines observations from your sessions. This skill is for reading. It does document witness's two write tools, but instructs the agent not to use them unless you explicitly ask — speculative writes pollute the archive.

    Not much at first. Observations accumulate as you work and facets appear only after a review pass, so a new archive has little history to report. The skill tells the agent to say the archive is empty rather than infer something from nothing.

    New in Agents & Orchestration