showcase: add Kleos Research

showcase/kleos-research/README.md

@@ -0,0 +1,99 @@
+# Kleos Research
+
+The layer between agents and the models they run on. Two things live here:
+**0xCopilot**, a local-first desktop harness, and **Kaleidoscope**, memory for
+agents that outlives the session.
+
+## 0xCopilot — the harness
+
+An Electron desktop app that supervises an embedded PostgreSQL and three Python
+services on `localhost`. Nothing is hosted: the runtime, the store, and the
+conversation history sit on the user's own disk.
+
+Relevant to EconomyOS: it runs on **your own key or a Virtuals ACP key**. The
+model picker is built from Virtuals' live catalogue — roughly sixty models from
+ten vendors behind one OpenAI-wire endpoint — and carries **real per-Mtok
+pricing on every row**, so the cost of a task is legible before it is spent
+rather than after. The catalogue is discovered rather than hardcoded, so a model
+Virtuals adds appears without a code change.
+
+`$CPILOT` is tokenised through Virtuals
+([listing 113720](https://app.virtuals.io/virtuals/113720)).
+
+## Kaleidoscope — the memory
+
+Filesystem-native, continually adapting knowledge for agent harnesses:
+user-owned inspectable files, no database server, no memory-owned model, and
+harness-neutral integration. It preserves decisions, evidence, and outcomes
+across sessions, then retrieves a small task-scoped context under explicit
+latency and token bounds.
+
+It is measured on **BEAM 100K** — ten memory abilities over conversations from
+100K tokens up — through an
+[open-source evaluation suite](https://github.com/kleos-research/kaleidoscope-benchmarks)
+that shells out to the real `kscope` binary rather than reimplementing it, so
+what gets measured is what ships.
+
+Held-out results are published in
+[Optimising for memory recall](https://kleosresearch.xyz/research/optimising-for-memory-recall.pdf):
+
+- a full scan of **100,000 memories in 3.04 ms**, with no index
+- **32 bytes per memory retains 99.7%** of the retrieval quality of 1024
+- **96% of the recall of a model four times its size**, from 7.2 MB
+
+Caveats travel with the numbers: the encoder results are configured on a train
+split and reported on held-out test over externally-authored corpora, and the
+timing is one machine — the ratios travel, the absolute milliseconds do not.
+
+## The skill
+
+[`skills/0xcopilot-local-workspace`](skills/0xcopilot-local-workspace) — stand up
+a local agent workspace and point it at Virtuals compute, so work runs on
+EconomyOS credits with per-Mtok pricing visible before a task is spent.
+
+It is deliberately a workspace skill and not a task-runner: the CLI is a
+launcher and installer with no `copilot run "<task>"` verb, and the skill says so
+under "when NOT to use this" rather than implying a pipeline that does not exist.
+Two steps are marked as human-only handoffs — signing in, and entering the
+provider key — because an agent should not be typing an operator's credential.
+
+**There is no Kaleidoscope skill, and that is not an oversight.** The evaluation
+suite is open source, but it requires the `kscope` binary on `PATH`, and
+Kaleidoscope itself is not yet published. A skill whose first step no reader can
+perform is worse than no skill. (Note also that the `kscope` package on PyPI is
+an unrelated project by another author — it is not this one.) When Kaleidoscope
+ships publicly, a memory skill belongs here.
+
+## Why the proof is shaped like this
+
+There is no demo video here, and the benchmark artifact is a findings file
+rather than a headline. That is deliberate.
+
+[`tools/harness-bench/FINDINGS.md`](https://github.com/0x-copilot-dev/0x-copilot/blob/main/tools/harness-bench/FINDINGS.md)
+is the harness benchmark, measured on the packaged app against a live model and
+scored from the same records the product bills from. Its headline is a real
+win — raising an inherited step ceiling took task completion from **3-of-4 to
+4-of-4 for +0.1% tokens**, because the ceiling had been terminating real work
+with `recursion_limit_exceeded`. It also carries, in the same file, the
+correction of an earlier version of itself that had declared that finding
+falsified on the strength of a metric structurally blind to the failure it was
+measuring.
+
+The same discipline is why nothing here is quoted against another memory
+system. A comparison that looked significant was withdrawn once an aggregation
+fault was found — one side had been scored across nine abilities and the other
+across ten — and the corrected interval crosses zero. It is not published
+because it is not settled.
+
+Claims that survive replication are the ones on the card.
+
+## Links
+
+| | |
+| --- | --- |
+| Harness source | https://github.com/0x-copilot-dev/0x-copilot |
+| CLI | https://www.npmjs.com/package/@0x-copilot/cli |
+| Desktop app | https://copilot.kleosresearch.xyz |
+| Kaleidoscope | https://memory.kleosresearch.xyz |
+| Evaluation suite | https://github.com/kleos-research/kaleidoscope-benchmarks |
+| The lab | https://kleosresearch.xyz |

showcase/kleos-research/showcase.json

@@ -0,0 +1,110 @@
+{
+  "slug": "kleos-research",
+  "title": "Kleos Research",
+  "tagline": "Runs agents locally on your own keys or Virtuals compute, meters what every run actually costs, and gives them a filesystem-native memory that outlives the session",
+  "description": "Kleos Research builds the layer between agents and the models they run on. 0xCopilot is the local-first desktop harness: bring your own key, or a Virtuals ACP key, and pick from a catalogue carrying live per-Mtok pricing so the cost of a task is visible before you spend it. Kaleidoscope is filesystem-native memory for agents — harness-agnostic, user-owned files, no database server and no memory-owned model — which scans 100,000 memories in 3.04 ms with no index, and holds 99.7% of its retrieval quality at 32 bytes per memory. It is measured on the BEAM 100K tier through an open-source evaluation suite that shells out to the real binary rather than reimplementing it. Both publish their measurements in the open, retractions included: the harness benchmark records a step-ceiling fix worth 3-of-4 to 4-of-4 task completion for +0.1% tokens, and prints the claims it has withdrawn beside the ones that held.",
+  "status": "live",
+  "topic": "agents",
+  "topics": [
+    "agents",
+    "memory",
+    "acp",
+    "compute",
+    "virtuals",
+    "token",
+    "robinhood-chain",
+    "developer-tools",
+    "privacy"
+  ],
+  "builder": {
+    "name": "Kleos Research",
+    "url": "https://kleosresearch.xyz"
+  },
+  "links": {
+    "repo": "https://github.com/0x-copilot-dev/0x-copilot",
+    "share": "https://x.com/Copilot0x/status/2078188003479195781",
+    "feedback": "https://github.com/Virtual-Protocol/acp-cli-demos/issues/new?title=Feedback%3A%20Kleos%20Research",
+    "demo": "https://copilot.kleosresearch.xyz"
+  },
+  "primitives": [
+    "acp",
+    "token"
+  ],
+  "visual": {
+    "kind": "local-first agent harness with metered runs and portable memory",
+    "eyebrow": "local · byok · virtuals compute",
+    "title": "the layer between agents and models",
+    "posterUrl": "https://raw.githubusercontent.com/Virtual-Protocol/acp-cli-demos/main/showcase/kleos-research/assets/hero-card.png"
+  },
+  "skills": [
+    {
+      "name": "0xcopilot-local-workspace",
+      "href": "https://github.com/Virtual-Protocol/acp-cli-demos/tree/main/showcase/kleos-research/skills/0xcopilot-local-workspace",
+      "sourcePath": "showcase/kleos-research/skills/0xcopilot-local-workspace",
+      "summary": "Stand up a local-first agent workspace on the operator's own machine and point it at Virtuals compute, so work runs on EconomyOS credits with per-Mtok pricing visible before a task is spent. Covers install, the human-only credential handoff, validation, and the stop conditions — including that the CLI is a launcher, not a batch runner.",
+      "install": "cp -R showcase/kleos-research/skills/0xcopilot-local-workspace ~/.agents/skills/\ncp -R showcase/kleos-research/skills/0xcopilot-local-workspace ~/.claude/skills/"
+    }
+  ],
+  "artifacts": [
+    {
+      "label": "Harness benchmark findings — a measured completion win, the real cost structure, and two published retractions",
+      "href": "https://github.com/0x-copilot-dev/0x-copilot/blob/main/tools/harness-bench/FINDINGS.md",
+      "kind": "proof"
+    },
+    {
+      "label": "0xCopilot source — Electron desktop app supervising an embedded Postgres and three Python services on localhost",
+      "href": "https://github.com/0x-copilot-dev/0x-copilot",
+      "kind": "repo"
+    },
+    {
+      "label": "Virtuals compute integration — catalog discovery against compute.virtuals.io with live per-Mtok pricing",
+      "href": "https://github.com/0x-copilot-dev/0x-copilot/blob/main/services/ai-backend/src/agent_runtime/api/virtuals_model_source.py",
+      "kind": "proof"
+    },
+    {
+      "label": "0xCopilot CLI releases — install the desktop app from npm",
+      "href": "https://github.com/0x-copilot-dev/0x-copilot/releases",
+      "kind": "demo"
+    },
+    {
+      "label": "Optimising for memory recall — held-out results behind the 3.04 ms scan and the 32-byte memory (PDF)",
+      "href": "https://kleosresearch.xyz/research/optimising-for-memory-recall.pdf",
+      "kind": "proof"
+    },
+    {
+      "label": "Kaleidoscope — filesystem-native, continually adapting memory for agent harnesses",
+      "href": "https://memory.kleosresearch.xyz",
+      "kind": "docs"
+    },
+    {
+      "label": "Kaleidoscope BEAM evaluation suite — open-source harness that runs the BEAM 100K tier against the real binary",
+      "href": "https://github.com/kleos-research/kaleidoscope-benchmarks",
+      "kind": "repo"
+    },
+    {
+      "label": "0xCopilot CLI on npm — install the desktop app with one command",
+      "href": "https://www.npmjs.com/package/@0x-copilot/cli",
+      "kind": "demo"
+    },
+    {
+      "label": "Kaleidoscope memory-recall paper — announcement",
+      "href": "https://x.com/Copilot0x/status/2089381387048472685",
+      "kind": "post"
+    },
+    {
+      "label": "Kleos Research — the lab behind 0xCopilot and Kaleidoscope",
+      "href": "https://kleosresearch.xyz",
+      "kind": "docs"
+    },
+    {
+      "label": "$CPILOT on Virtuals — the live listing",
+      "href": "https://app.virtuals.io/virtuals/113720",
+      "kind": "demo"
+    }
+  ],
+  "feedbackPrompts": [
+    "Does live per-Mtok pricing in the model picker change which model you would pick for a task?",
+    "What would you need to see before trusting an agent memory that lives as plain files on your own disk?",
+    "Which matters more in a harness you pay for: a lower cost per task, or a higher share of tasks that actually finish?"
+  ]
+}

showcase/kleos-research/skills/0xcopilot-local-workspace/SKILL.md

@@ -0,0 +1,105 @@
+# 0xCopilot local workspace on Virtuals compute
+
+Stand up a local-first agent workspace on the operator's own machine and point
+it at Virtuals compute, so agent work runs on EconomyOS credits instead of a
+personal provider bill — with the model catalogue carrying live per-Mtok pricing
+so the cost of a task is legible before it is spent.
+
+## When to use this
+
+- The operator has Virtuals credits (or their own provider key) and wants a
+  local harness rather than a hosted one.
+- Conversations, runs, and artifacts must stay on the operator's disk.
+- You are asked to verify or repair an existing 0xCopilot install.
+
+## When NOT to use this
+
+- **You need a headless, scriptable agent run.** This CLI is a launcher and
+  installer, not a batch runner — there is no `copilot run "<task>"`. Work is
+  given to the agent in the app. If you need automation, drive the product's
+  HTTP API instead; do not pretend this skill gives you a pipeline.
+- The target is Linux, or Windows on ARM. Not supported.
+- The operator wants a shared/multi-tenant deployment. This is single-user and
+  local by design.
+
+## Preconditions
+
+| | |
+| --- | --- |
+| Node.js | 20+ |
+| OS | macOS (Apple Silicon or Intel) or Windows x64 |
+| Network | required on first launch |
+| Disk | several hundred MB for the runtime, plus app data |
+| Credential | a Virtuals ACP key, or an OpenAI / Anthropic / Google key |
+
+## Approval gates — stop and ask before each
+
+1. **First install downloads several hundred megabytes.** Get explicit consent
+   before running `copilot` or `copilot install` on a metered or constrained
+   connection.
+2. **Running agent tasks spends credits.** Virtuals compute draws down the
+   operator's balance. Confirm a budget before starting work, not after.
+3. **`copilot uninstall` deletes all local 0xCopilot data** — conversations,
+   runs, artifacts. It is not recoverable. Never run it to "clean up" without an
+   explicit instruction naming the consequence. Prefer `copilot repair`, which
+   recovers a stuck launch while keeping data.
+4. **Never type, echo, or store the operator's provider key.** It is entered by
+   a human, in the app, and lives in the OS keychain.
+
+## Steps
+
+```bash
+npm install -g @0x-copilot/cli   # or: bun add -g @0x-copilot/cli
+copilot                          # prepares the runtime, then launches the app
+```
+
+Then **hand off to the human** for two steps an agent must not perform:
+
+1. Sign in (wallet, or Google where enabled).
+2. **Settings → Models & keys → Provider keys** → add the Virtuals key. Selecting
+   Virtuals as the provider exposes its catalogue — roughly sixty models across
+   ten vendors behind one endpoint, each row carrying real per-Mtok pricing.
+
+Give the agent a goal in the composer once a key is saved.
+
+## Stop conditions
+
+- `copilot doctor` reports an unmet requirement → report its output verbatim and
+  stop. Do not attempt to work around a missing runtime.
+- The launch hangs → `copilot repair` (keeps data). Add `--session` only if the
+  operator agrees to clear saved sign-ins. Escalate rather than reinstalling.
+- A run fails with a provider or service error → this is a credential or
+  upstream-availability problem, not a task problem. Re-check the key before
+  retrying, and never retry in a loop against a paid endpoint.
+
+## Validation
+
+```bash
+copilot version    # prints the installed version
+copilot doctor     # diagnoses the setup and prints what it finds
+```
+
+The install is good when `doctor` reports no unmet requirements and the app
+launches to a workspace. The *key* is good when the composer's model pill
+resolves to a real model rather than a placeholder — a saved key is not proof of
+a working one, because the catalogue endpoint answers publicly and says nothing
+about your credential.
+
+## Output contract
+
+On success, report:
+
+- installed CLI version,
+- `doctor` verdict (unmet requirements, if any),
+- whether a provider key is configured — **as a boolean, never the value**,
+- the local data location, so the operator knows what `uninstall` would remove.
+
+On failure, report the failing command, its exact output, and which stop
+condition above was hit. Do not summarise an error into prose — the product's
+typed errors are more precise than a paraphrase of them.
+
+## Links
+
+- CLI package — https://www.npmjs.com/package/@0x-copilot/cli
+- Source — https://github.com/0x-copilot-dev/0x-copilot
+- App — https://copilot.kleosresearch.xyz