Skip to content

SCRATCHPAD.md

Latest commit

 

History

History
764 lines (597 loc) · 50.8 KB

SCRATCHPAD.md

File metadata and controls

764 lines (597 loc) · 50.8 KB

SCRATCHPAD — Session Memory

This file persists context between Claude Code sessions.

Current Session (2026-05-09 evening) — PEEKABOO INTEGRATION: AX-FIRST NATIVE AUTOMATION + ISSUE 380 (THREE IDIOMS)

Headline

Wired Peekaboo (https://github.com/openclaw/Peekaboo) into kbot end-to-end across three layers in a single sprint via parallel-agent swarm: provider-agnostic adapter, six registered peekaboo_* tools, additive AX-first fallback in the existing mouse_click / keyboard_type flow, and the editorial that names what just changed. Six agents dispatched; all returned green.

What shipped (kbot v4.4.0)

1. src/adapters/peekaboo/ — bidirectional adapter (5 files, 629 LOC)

  • types.ts — PeekabooSnapshot, PeekabooElement, the five command result types, PeekabooOutcome<T> discriminated-union helper.
  • runner.ts — runPeekaboo(args, opts) via execFile, peekabooAvailable() probe, PEEKABOO_BIN env override.
  • commands.ts — see, click, type_, setValue, performAction, agent — all return PeekabooOutcome<T>.
  • index.ts — public surface.
  • adapter.test.ts — 13 deterministic tests, vi.mock('node:child_process') for stubbing.

2. src/tools/peekaboo.ts — six registered kbot tools

  • peekaboo_see / peekaboo_click / peekaboo_type / peekaboo_set_value / peekaboo_perform_action / peekaboo_agent.
  • All tier: 'free', all return string, never throw — outcomeToString collapses PeekabooOutcome to either pretty JSON or Error: ....
  • Binary presence guarded via peekabooAvailable().
  • Approval flow: requireApproval(app) checks ~/.kbot/computer-use/<app>.lock (Coordinator-managed). computer.ts does not export approvedApps/isAppApproved/claimApp, so this is a fail-closed check rather than a direct in-process integration. Limitation documented inline.
  • 12 vitest tests. Registered in swarm-2026-04.ts.

3. src/tools/computer.ts — additive AX-first fallback

  • Imports peekabooAvailable, see, click, type_ from the adapter.
  • Module-level peekabooReady() cached probe with KBOT_DISABLE_PEEKABOO=1 escape hatch.
  • mouse_click (lines ~520–538): on darwin + binary present, captures snapshot, attempts AX click via peekabooClick; falls through to AppleScript+cliclick on any failure.
  • keyboard_type (lines ~739–752): same pattern with peekabooType.
  • New peekaboo_status tool reports available / disabled-via-env / not-on-PATH.
  • 6/6 existing computer.test.ts pass — no regressions, no signature changes.

4. skills/native-automation/peekaboo-snapshot-act/SKILL.md (91 lines)

  • Names the doctrine: web=DOM, audio=OSC, native=AX. Snapshot-then-act over screenshot-and-guess.
  • Iron Laws: ONE SNAPSHOT, MANY ACTIONS / ELEMENT ID OVER COORDINATES / PERFORM-ACTION OVER CLICK.
  • Five phases: Approve & focus → Capture surface → Choose the right verb → Reuse the snapshot → Fall back gracefully.

What shipped (magazine)

ISSUE 380 — THREE IDIOMS (cobalt + cream + asymmetric-left + asterisk-stamp + NOTED · IDIOMS · V·26)

  • Essay spread, three numbered sections (DOM / OSC / Accessibility), closing recognition that surface-published structure is the right input shape, kbot v4.4.0 / Peekaboo news in a small — MARGIN — block at the end.
  • Adjacent to the 376/377/378 filed-pattern arc but verb shifts from FILED to NOTED — a recognition, not a filing.
  • First binding of cobalt as the structural-argument accent.

Collision recovery

ISSUE 379 — ON BECOMING A REAL MAGAZINE was deployed by the user (commit 5f16579e) mid-swarm; the issue agent's brief was already in flight and overwrote 379.ts with THREE IDIOMS content. Recovered: git checkout HEAD -- src/content/issues/379.ts restored the deployed essay byte-for-byte, THREE IDIOMS preserved at 380.ts with all identifiers renumbered, registered behind 379 in index.ts. No data loss.

Test math

  • New: 25 tests (13 adapter + 12 tool wrapper), all green.
  • Regression: 6/6 computer.test.ts still pass; no existing tests modified.
  • Type-check: clean across both site and kbot package.

Install state (this Mac)

  • peekaboo 3.0.0-beta4 at /opt/homebrew/bin/peekaboo (via brew install steipete/tap/peekaboo).
  • Project .mcp.json registers peekaboo: peekaboo mcp (committed, travels with repo).
  • User-level ~/.claude.json registers same (every Claude Code session on this Mac picks it up). Connection verified ✓.
  • Macos perms inherit from Claude Code's existing Screen Recording + Accessibility grants — no extra prompt.

Decisions worth remembering

  • npx -y @steipete/peekaboo does NOT spawn the MCP server — that's the CLI. The MCP entry point is peekaboo mcp (subcommand), not a separate bin. Both .mcp.json and the claude mcp add invocation must use peekaboo mcp (or the explicit peekaboo-mcp.js bin from the npm package if going npx-only).
  • The kbot adapter is provider-agnostic by design (no runtime dep on @steipete/peekaboo) — matches the agent-sdk adapter posture from the prior session.
  • AX-first fallback is gated behind darwin + binary present + non-empty app arg — Linux behavior bit-for-bit unchanged. Escape hatch is KBOT_DISABLE_PEEKABOO=1.

Open ends for next session

  • Apply the AX-first pattern to additional computer.ts tools (currently only mouse_click and keyboard_type get it). Candidates: mouse_drag, keyboard_combo, app_focus.
  • Surface peekaboo_agent through a dedicated SKILL.md so its handoff semantics (delegate to peekaboo's own agent loop) are explicit.
  • The requireApproval limitation in peekaboo.ts calls for a small computer.ts refactor: export isAppApproved so direct in-process gating is possible. Low risk, high clarity gain.

Files touched (this session)

A packages/kbot/src/adapters/peekaboo/types.ts
A packages/kbot/src/adapters/peekaboo/runner.ts
A packages/kbot/src/adapters/peekaboo/commands.ts
A packages/kbot/src/adapters/peekaboo/index.ts
A packages/kbot/src/adapters/peekaboo/adapter.test.ts
A packages/kbot/src/tools/peekaboo.ts
A packages/kbot/src/tools/peekaboo.test.ts
M packages/kbot/src/tools/computer.ts            (additive AX-first + peekaboo_status)
M packages/kbot/src/tools/swarm-2026-04.ts       (register peekaboo tools)
M packages/kbot/package.json                      (4.3.0 → 4.4.0)
A packages/kbot/skills/native-automation/peekaboo-snapshot-act/SKILL.md
A src/content/issues/380.ts                       (THREE IDIOMS)
M src/content/issues/index.ts                     (register ISSUE_380)
M .mcp.json                                       (peekaboo entry)
M ~/.claude.json                                  (user-scope MCP, outside repo)
M SCRATCHPAD.md                                   (this entry)

Earlier Today (2026-05-09 morning) — MAY-NEWS RESPONSE: SECURITY-AUDIT SKILLS + AGENT SDK ADAPTER + REVIEW SPREAD

Headline

Started as a "where's AI at right now" question, ended as a build sprint against the May 2026 news cycle. Two new things land on claude/ai-current-state-8f4vr: a security-audit skill family (the BYOK/local-first answer to Project Glasswing + Claude Mythos) and an Agent SDK adapter (the schema-only response to Anthropic opening the Agent SDK to external developers).

What shipped (kbot)

1. skills/security-audit/ — four new skills

  • local-vulnerability-hunt/SKILL.md — Mythos-style audit workflow, BYOK frontier model, audit trail at ~/.kbot/security-audits/<session>/. Five phases: Scope → Surface map → Hypothesize → Confirm → File. Iron laws: no remote scan without consent; no finding without evidence; no fix without verification.
  • dependency-audit/SKILL.md — npm audit + lockfile diff + provenance review. Lockfile is the truth; transitive counts double.
  • secrets-leak-scan/SKILL.md — working-tree + git-history sweep with named patterns (AWS, sk-, Slack, GitHub PAT, JWT, private keys). Rotate before scrub; never rely on rewriting public history.
  • threat-model-quickdraw/SKILL.md — STRIDE-lite, 30 minutes, produces docs/threat-models/<feature>.md artifact.

2. src/tools/security-audit-local.ts — substrate tool

  • security_audit_local registered tool. Walks a tree, applies pattern set across JS/TS/Py/Go/Rust/Ruby/Java/PHP/Shell, persists JSONL surface map. Detects: eval-shaped sinks, subprocess calls, route registrations, weak crypto, JWT verify-skip, SQL string concat, FS writes near user input, TLS-skip flags, non-constant-time comparisons.
  • 16 tests, all stub-driven, deterministic. Caps at 5,000 files / 2,000 signals / 1MB per file.
  • Wired into swarm-2026-04.ts registration.

3. src/adapters/agent-sdk/ — bidirectional adapter

  • types.ts — AgentSdkTool / AgentSdkExecutableTool / AgentSdkInputSchema / JSON Schema property types. NO runtime dependency on @anthropic-ai/sdk — kbot stays provider-agnostic.
  • to-agent-sdk.ts — kbot ToolDefinition → Agent SDK tool. Options: preserveName, renameTool, strict.
  • from-agent-sdk.ts — Agent SDK tool → kbot ToolDefinition. Two flavors: schema-only (with optional fallbackExecutor) and executable (handler embedded). Type-union fallback for ["string","null"].
  • index.ts — public surface.
  • adapter.test.ts — 20 tests, including round-trip preservation.

Test math

  • New: 36 tests (16 security-audit-local + 20 agent-sdk adapter), all green.
  • Regression check: futures/ suite still 95/95 green.
  • Type-check: zero errors in new files. Pre-existing chalk/ora type warnings unchanged.

Decisions worth remembering

  • Forecast module was already shipped end-to-end (4.2.0 wire-up via forecast-summary.ts was done) — pulled it from the build list when the README revealed it.
  • Adapter is schema-only by design — kbot's BYOK contract means we don't ship @anthropic-ai/sdk as a dep just to translate JSON.
  • Skill family pairs with existing pentest/hacker-toolkit (remote, authorized) and existing agents/security-agent.ts (the runtime under security_agent_scan). The new skills are the narrative layer on top of the substrate, not duplicates.

Why the Mythos echo is on-brand

Glasswing only goes to ~6 organizations (AWS, Apple, Cisco, Google, JPM, Microsoft). The democratized version that fits kbot's positioning: same phased workflow, but BYOK any frontier model, against your own code, audit trail on disk you control. MIT, no phone-home, no allowlist.

Open ends for the next session

  • Optional v4.3.0: ship a "second-opinion" hook that fans the same surface signal across two providers and diffs the verdicts. The local-vulnerability-hunt skill already names this in its anti-patterns; make it real.
  • Issue 376 candidate: editorial on the Mythos / Glasswing posture vs the BYOK alternative, with the audit trail as the evidence pack. Probably writes itself off the new skill family.
  • The Agent SDK adapter could grow a thin messages-api-router.ts that takes a tool_use block and dispatches into the kbot registry — a pure convenience for users wiring kbot into Anthropic Messages API loops.

Files touched

A packages/kbot/skills/security-audit/local-vulnerability-hunt/SKILL.md
A packages/kbot/skills/security-audit/dependency-audit/SKILL.md
A packages/kbot/skills/security-audit/secrets-leak-scan/SKILL.md
A packages/kbot/skills/security-audit/threat-model-quickdraw/SKILL.md
A packages/kbot/src/tools/security-audit-local.ts
A packages/kbot/src/tools/security-audit-local.test.ts
A packages/kbot/src/adapters/agent-sdk/types.ts
A packages/kbot/src/adapters/agent-sdk/to-agent-sdk.ts
A packages/kbot/src/adapters/agent-sdk/from-agent-sdk.ts
A packages/kbot/src/adapters/agent-sdk/index.ts
A packages/kbot/src/adapters/agent-sdk/adapter.test.ts
M packages/kbot/src/tools/swarm-2026-04.ts  (register security_audit_local)
A src/components/ReviewFeature.tsx          (new editorial tool #5)
A src/components/ReviewFeature.css
M src/components/IssueFeature.tsx           (router case for review)
M src/content/issues/index.ts               (ReviewSpread + types)
M src/content/issues/accents.ts             (review→olive default)
M docs/design-language.md                   (current-tools table updated; future-moves drift fixed)
M SCRATCHPAD.md                              (this entry)

Editorial follow-on (ISSUE 378)

First use of the new review spread tool: ISSUE 378 — ON THE BENCH — ledger stock + classic + olive accent + FILED · BENCH · V·26 seal. Five subjects graded against a five-criterion rubric (access ceiling, coverage, audit trail, cost ceiling, second-opinion friction): Mythos, GPT-5.5-Cyber, Sec-Gemini v1, Llama 4 + PurpleLlama, kbot security_audit_local + BYOK frontier. Standout (BEST AVAILABLE) goes to Llama 4 + PurpleLlama, deliberately not the house toolkit. kbot scores A− with its limits written down. Continues the small filed-pattern arc 376 STANDARDS → 377 API TIER → 378 BENCH on the AI-tools beat.

Caveat written into the issue file: SpreadCommon.stock does not yet admit ledger, so the cover carries ledger and the spread falls back to ivory. Worth resolving in a future commit if the editor wants ledger inside-spreads; for now ivory under the score monuments reads cleanly enough that this isn't a regression.

Editorial follow-on (review spread)

Built on top of the security-audit work in the same session. The news-cycle response on the kbot side (security-audit skills + Agent SDK adapter) needed a magazine form to carry the editorial: the review spread is that form. Top-line italic verdict, numbered rubric, optional standout award, grid of subject cards with score monument + stars + pros/cons + per-card verdict. Olive-led by default. ISSUE 378 candidate: grade frontier model security capabilities (Mythos, Glasswing partners, BYOK alternatives) using the new tool against the new audit substrate.

Drift the doc patch caught: the design-language.md "Future moves" list still claimed cobalt/ivy/pool weren't shipped — they've been in accents.ts for weeks. The dispatch and forecast tools were also live but undocumented in the current-tools table. All five tools (essay/interview/forecast/dispatch/review) are now in the table with their grammar one-liners and default accents.

Prior Session (2026-04-28 → 04-29 overnight) — THE BIG BUILD: V5 EXECUTION + 4-ISSUE EDITORIAL DROP + KERNEL.MD SUPERSESSION

Headline

The plan from the previous V5-research session got built tonight, plus four magazine issues, plus the canonical project doc rewrite. Roughly 18 hours of execution against the work the prior session had set up.

What shipped (kbot)

Seven npm releases in sequence:

  • 3.99.32 — wave 1 of 2026-04 capability swarm: image_thoughtful, channels family (Slack full + 5 stubs), file library, workspace agents, parallel computer-use coordinator, security agent, plugin integrity manifest. 7 modules, 76 new tests.
  • 3.99.33 — wave 2 wiring: Office channel adapter, Anthropic Managed Agents adapter (managed-agents-2026-04-01 beta header), plugin-sdk integrity, hierarchical planner integration, security agent wired into guardian/hacker, computer.ts rewired to Coordinator.
  • 3.99.34 — Claude Code / Cursor / Zed integration: kbot setup-* CLIs, .claude/skills/kbot.md pre-authorization skill (the Harrison fix), 15 MCP description audit reframings.
  • 3.99.35 — LLaDA2.0-Uni local provider + local_image_thoughtful tool (no OpenAI key needed for image gen).
  • 4.0.0 — evidence-driven curation. 670+ tools → 105 specialty skills, 61 deprecated, 630 cut. Five Phase A discovery agents produced telemetry/test-refs/daemon-deps/external-surface/forge-readiness CSVs at packages/kbot/CURATION_DISCOVERY/CURATION_DECISION.csv. Source files stay on disk (reversible); only LAZY_MODULE_IMPORTS got pruned. Forge load-path bug fixed (forged tools now survive restart). Vocab refresh: user-facing copy says "skills" where the OAI/Claude Code framing fits.
  • 4.0.1 — cache-warmth (Anthropic prompt-cache TTL warning, jcode borrow), BENCHMARKS.md (methodology-explicit n=5 cold-start vs claude/codex/aider/opencode/jcode), templates/jcode-mcp-snippet.json (kbot wraps as backend for jcode users), research/jcode-analysis.md.
  • 4.1.0 — V5 futures substrate. Six modules under packages/kbot/src/futures/: harness (Sylph 2604.21003), skill-graph (Tencent 2604.25727), latent-state (Stanford+UIUC+NVIDIA+MIT 2604.25917), forecast (in-house), persona (Cequence press signal), debate (Plurai 2604.25203). 95 tests, all stub-driven, deterministic, ~150ms total. Net +338 tests session-over-session (731 → 1069).

What shipped (kernel.chat magazine)

Four issues drafted, registered, deployed:

  • 372 — THE AUDIT (postmark dateline fired for the first time: ROOM 503 · IV·26; new ledger stock + ledger-rule layout + graphite accent)
  • 373 — ON COMPOSITION (cream + asymmetric-left + no ornament; the editorial-neighbours framework applied at the AI-tools layer)
  • 374 — AGAINST VIRAL BENCHMARKS (ivory + classic + new asterisk-stamp ornament; jcode anatomy without naming the engineer)
  • 375 — THE SIX BORROWS (cream + new numbered-catalog layout + dossier element; credit-page for the v5 build)

Two corrections deployed after first publish (372's count-of-counts line dropped, 375's "Cequence Security" → "Cequence").

Design-language extensions

  • WIRED decoded as second editorial neighbour in docs/design-language.md — four transferable mechanics (methods sidebar adjacent to claim, numbered references at foot of long features, pull-quote with the number not the sentence, tomato-tinted inline footnote markers). Magazine now has two complete neighbours decoded (PAPERSKY + WIRED).
  • All four PAPERSKY-starred mechanics now in active use (postmark dateline was the unfired one until 372).
  • Type extensions: IssueStock += 'ledger'; IssueCoverLayout += 'ledger-rule' + 'numbered-catalog'; IssueCoverOrnament += 'asterisk-stamp'; IssueRecord.coverPostmark?: { place, date } (first-class support); INK_SEEDS += graphite (#3F3D3A).
  • Stock cabinet — explicit register documented for what each paper signals (cream/ivory/butter/kraft/ink/ledger).

Doc supersession

KERNEL.md (335 lines) now the canonical project reference; CLAUDE.md (64 lines, was 673) reduced to a shim pointing at KERNEL.md. Old framing was "kbot is the product, web is the companion" — wrong now that the magazine is a working publication. New framing: one publication on two media, the same editor's discipline applied to a tool registry and a table of contents.

Real user signal

Harrison McCormick (Pro user, second time he's been the canary) iMessaged 4/27 that "Claude is refusing to help me use the agent it had me build, citing academic dishonesty." That triggered wave 3 (kbot.md skill + setup CLIs + MCP description audit). Same-night fix; he can npm i -g @kernel.chat/kbot && kbot setup-claude-code to land in a working integration. Texts to him are drafted but not sent yet.

Pattern that worked

  1. Read research first (the night before's session set up V5_FUTURES_PLAN.md so this session could just execute)
  2. Plan in writing — every release commit cites numbers and reasoning
  3. Swarm-dispatch parallel agents on non-overlapping work (separate files, separate scopes); keep scope tight per agent — stream timeouts come from sprawling context
  4. Foreground the integration step (registration, typecheck, deploy)
  5. Ship with evidence (audit CSVs, RELEASE_NOTES, CHANGELOG, methodology docs)

Known timeouts / agent failures

Stream-idle timeouts hit ~5 agents this session on the bigger scopes. Recovery pattern: check what survived on disk, tight-rescope, salvage in foreground. The stalled agents were always over-scoped, never the deterministic-stub ones. Going forward: cap per-agent scope at one-module / ≤500 LOC / single-purpose.

Open in the queue

  • Skill injector (jcode borrow C) — embedding-driven skill auto-injection. Stalled this session, deferred to 4.2.
  • Smart grep (jcode borrow B) — file-structure injection. Stalled, deferred.
  • Cross-harness session import (jcode borrow D) — Claude Code → kbot session resume. Stalled, deferred.
  • forecast/ → growth_summary wiring — substrate exists, integration is 4.2 work.
  • persona/ → permissions.ts wiring — substrate exists, integration is 4.2 work.
  • Real EvolutionAgent — interface shipped, codegen implementation is 4.3+.
  • Deprecation removal runway — 4.1 (lab-) → 4.2 (stream-) → 4.3 (engine-*) → 4.4 (misc) → 4.5 (forge_tool migration).
  • Harrison upgrade text — drafted in conversation, not sent.
  • Social post about the cut — could pair the audit story (672 → 105 with public CSV) with the four-issue drop. Not drafted.

Live state at session end

  • kbot v4.1.0 on npm, latest tag, 1069 tests passing, 56 test files
  • kernel.chat ISSUE 375 — THE SIX BORROWS is the live cover at https://kernel.chat
  • GitHub: main at 7e154a64 (KERNEL.md supersession). gh-pages at b26e1f6 (post-corrections deploy). 13 stars, 2 forks, 0 open PRs.
  • npm downloads (all-time): 25,828 total since first publish 2026-03-04 (56 days). Last 30 days: 11,401.

Recommended next session opener

"Pick from: (1) Send Harrison the upgrade text + draft the social post about the 4.0 cut. (2) Wire forecast → growth_summary as the first 4.2 substrate-to-product hop. (3) Skill injector salvage with tighter agent scope. (4) Start drafting ISSUE 376 (subject open — could be the postmark mechanic at a different geography, or the WIRED decode applied to a specific data-journalism piece kbot has the receipts for)."

Previous session notes follow below

Prior Session (2026-04-29) — V5 RESEARCH + FUTURES PLAN

What happened

  • Verified live state: kbot is at v4.0.1, published 2026-04-29 (CLAUDE.md was at 3.60.0 / 3.99.21 — stale). Real npm downloads ~1,266/wk, not the 4,806 stated.
  • CLAUDE.md updated in this session to match v4.0.1 reality. Committed as b497466 on claude/project-scope-definition-Z7fbx, pushed.
  • Researched HF Daily Papers (2026-04-29) — pulled 5 high-relevance papers for kbot's architecture:
    • Sylph.AI "The Last Harness You'll Ever Build" (2604.21003) — formalizes Worker / Evaluator / Evolution Agent loop. Maps directly onto kbot's existing critic-gate + learning systems.
    • Tencent SkillSynth (2604.25727) — scenario-mediated skill graph for terminal task synthesis. Maps onto skill-router package.
    • Stanford RecursiveMAS (2604.25917) — latent-state multi-agent. Top of HF page (121 upvotes). Successor architecture to text-handoff multi-agent.
    • Plurai BARRED (2604.25203) — asymmetric debate generates guardrail training data. Recipe to ship critic-gate.ts (currently feature-flagged off awaiting FP measurement).
    • Alibaba TCOD (2604.24005) — temporal curriculum on-policy distillation. Recipe for distilling frontier models to local llama.cpp runtime.
  • Researched competitive landscape: terminal coding agents have consolidated into a real category; MCP at 97M installs (mainstream); agentic sovereignty / "Thick Agents" recognized as a tier.

Strategic positions identified

  1. Harness Evolution Loop (Sylph) — foundation for self-improving kbot
  2. Skill Graph (Tencent) — formalizes skill-router as a graph
  3. Latent-state envelope (Stanford) — prep for multi-agent v2
  4. Forecast / predictions module — addresses user framing literally
  5. Persona / privilege scoping (Cequence Apr-28) — wraps permissions.ts as identity-bound
  6. Asymmetric debate (Plurai) — generates the data that ships critic-gate

What's drafted (not built)

  • packages/kbot/V5_FUTURES_PLAN.md — full plan, sequenced into 4 phases, ~1,200 LOC TS + 400 LOC tests + 300 lines markdown. Read this when back.
  • packages/kbot/src/futures/harness/types.ts — concrete sample. Full Harness Evolution Loop type system. Pure types, no runtime, no risk. Demonstrates the contract every other module would target.
  • packages/kbot/src/futures/{harness,skill-graph,latent-state,forecast,persona,debate}/ — empty directories, placeholder for the 6 modules.

Key decisions deferred to user (see V5_FUTURES_PLAN.md "Open decisions")

  1. Module name: futures/ vs v5/ vs experimental/
  2. CLI surface: subcommand vs flag
  3. Forecast as new tool or arg on existing
  4. Public API export path
  5. Phase ordering: Phase 1 first (recommended), or do forecast first for visibility

What didn't ship this session

  • No code beyond the types.ts sample (user redirected to "prepare and plan till I get home")
  • Critic-gate still feature-flagged off
  • BARRED debate not implemented
  • Terminal-Bench evaluation not run

Recommended next session opener

"Read packages/kbot/V5_FUTURES_PLAN.md, pick a phase, ship it."

The plan is structured so any single phase is independently shippable and reversible (everything additive under src/futures/).

Branch state

  • Current branch: claude/project-scope-definition-Z7fbx
  • Pushed commits: b497466 (CLAUDE.md refresh)
  • Uncommitted at session end: V5_FUTURES_PLAN.md, futures/harness/types.ts, this SCRATCHPAD entry

Previous Session (2026-04-19) — KBOT-CONTROL + SUNO-INSPIRED ROADMAP

What shipped

  • kbot-control.amxd — single M4L device, TCP 127.0.0.1:9000, JSON-RPC 2.0, 37+ LOM methods. Supersedes AbletonOSC + AbletonBridge + kbot-bridge (all three can be decommissioned).
  • packages/kbot-control-standalone/ — open-source-ready package with client.ts, LICENSE, LAUNCH assets.
  • packages/kbot/src/growth.ts — "kbot is N% better this week" dashboard. The Suno "My Taste visible" play applied to an agent: surface measurable self-improvement to users.
  • packages/kbot/src/critic-gate.ts — always-on adversarial critic on every tool output. GAN-style generator/discriminator pattern; every agent move gets scored before it leaves the loop.
  • packages/kbot/src/planner/hierarchical/ — 4-tier planner design inspired by Suno's 3-stage transformer: Goal / Phase / Action / ToolCall.
  • packages/kbot/CURATION_PLAN.md — 670 → 52 tools plan. 87% reduction. Plan only, not executed.
  • packages/kbot/research/action-tokens/ — research proposal for an agent-action token LM. This is the Suno lesson applied literally: neural codec → audio pattern maps cleanly to agent-action tokens → agent outputs.
  • Full migration: 22 ableton_* tools now prefer kbot-control, fall back to OSC. Backwards-compatible rollout.
  • Demo session: 90s Atlanta soul — Serum 2 Mark1 Stage + 6-device stock FX chain + Cm9 / Ab△9 / Fm11 / G7♭9 progression.

Key findings

  • Ableton 12.4 beta 15 removed LOM browser access — this is a real Ableton regression, not a kbot bug. Worked around in the migration. Report upstream.
  • Suno's architecture (neural codec → audio pattern LM → audio) maps 1:1 onto agent design (action codec → action pattern LM → tool calls). That's the research moat.
  • Warner Music settled with Suno Nov 2025; licensed models launching 2026. Licensing precedent opens the door for agent-action training data deals.

Interview prep (hand-off docs)

  • .claude/INTERVIEW_CHEAT.md — the pitch.
  • .claude/STUDIO_MODE.md — the Claude Code prompt to replay the demo session.
  • TZFM one-breath defense is ready (see packages/2027/src/dsp/fm.rs).

Open gaps going forward

  • Action-token research pivoted to embedding nearest-neighbor after baseline measurement killed the transformer bet (2026-04-19) — heuristic learned-router.ts already hits 91.8% top-5, far above the transformer proposal's 40% ship bar. New direction: embedding-NN + user-specific fine-tuning; prototype in packages/kbot/research/action-tokens/embedding-nn/, transformer-era artifacts in _archive/.
  • Must add durationMs + outcome logging to tool-pipeline.ts — still relevant: the per-user fine-tune corpus for the embedding-NN approach depends on it.
  • browser.load_by_name needs UI fallback until Ableton restores the API.
  • Critic false-positive rate needs retrospective analysis to tune strictness. Currently shipping at default thresholds.
  • Hierarchical planner Phase 2 (real tier logic) not implemented yet — scaffolding only.
  • Tool curation not executed — plan only. 670 → 52 reduction remains a design doc.

Previous session notes follow below

Prior Session (2026-04-17) — HERMES PARITY + SKILLS + SELF-AWARENESS (v3.99.0 → v3.99.2)

Three ships tonight

  1. v3.99.0 — skills-loader v2 (agentskills.io format), 14 bundled native skills, Hermes import (76 skills), CLI kbot skills list | import. Outperforms Hermes on 8 of 10 axes per audit.
  2. v3.99.1 — self-awareness.ts ground-truth block injected into every system prompt. Reads version, configured provider, model, platform. Fixed introspection hallucination (kbot was claiming "GPT-4 over WebSocket" when asked about itself).
  3. v3.99.2 — stronger ground-truth directives. Explicit "NOT GPT-4, NOT Hermes, NOT Llama" disclaimers. kbot now answers self-referential questions accurately: "Kernel Bot v3.99.2, Node.js 22.18.0, Ollama's gemma3:12b."

Honest remaining gaps

  • Ground truth reflects CONFIGURED provider, not runtime fallback provider. If Ollama is down and Claude answers, kbot still says "Ollama" — needs a runtime probe.
  • Tokenizer doesn't stem — "dreams/dream" and "curate/curation" miss in relevance scoring. 2 of 10 benchmark queries still pick imported Hermes skill over native kbot skill. ~50 lines of stemming would fix.
  • ollama launch kbot not yet in Ollama's integration list. openclaw (predecessor) already is. Distribution work.

Previous session notes follow below

Prior Session (2026-04-17) — HERMES PARITY + SKILLS SHIPPED (v3.99.0)

What shipped to npm tonight

  • @kernel.chat/kbot@3.99.0 published + installed globally.
  • 14 bundled skills in packages/kbot/skills/ across 7 categories (software-development, self-improvement, orchestration, memory, music-production, deployment, emergent). Ship in the tarball via files: ["skills/**/*.md"].
  • skills-loader.ts v2 — recursive category-dir scan, YAML frontmatter (agentskills.io + Hermes + kbot.metadata), conditional activation (requires_toolsets / fallback_for_toolsets / platforms), relevance scoring with native-content boost, token budget (2000), dedup with precedence: project-local > bundled > user-global.
  • CLI commands: kbot skills list, kbot skills import --from <hermes|claude|path>. Import symlinks 76 Hermes skills into ~/.kbot/skills/imported/.

Why: Hermes Agent shipped as an Ollama 0.21 integration

Research conclusion (stored in project_hermes_adoption.md): Hermes's edge was curated skill bodies + agentskills.io format. kbot had the plumbing (skill-system.ts, skill-library.ts, skill-rating.ts) but disconnected from the standard. Now compatible and superior on 8 of 10 axes — agent orchestration, memory cascade, self-improvement, music production are native categories Hermes has nothing for.

Audit results before publish

  • Functional edge cases: 12/12 pass
  • Security: 10/10 pass (no eval, no Function(), YAML bombs stored as strings, path traversal inert)
  • Content truthfulness: 3 bugs found and fixed (train-curate → train-self; recordSkillExecution/patchSkill → skill_manage; "17 specialists" → "25+")
  • Fresh install: extracted tarball contains all 14 SKILL.md files at correct paths
  • 731/731 existing tests still pass; typecheck clean

Known follow-ups

  • Tokenizer doesn't stem — "dreams" vs "dream" miss, "curate" vs "curation" miss. 2 of 10 benchmark queries still pick imported Hermes skill over native kbot skill. Low-priority; stemming would add ~50 lines.
  • ollama launch kbot not yet in Ollama's integration list. openclaw (kbot's predecessor) already is. Distribution work, not code.

Previous session notes follow below

Prior Session (2026-04-16) — TRAIN-SELF: LOCAL FINE-TUNING PIPELINE

Clean state (all committed)

7 new files in packages/kbot/src/: teacher-logger.ts, train-curate.ts, train-self.ts, train-cycle.ts, train-agent-trace.ts, train-merge.ts, train-grpo.ts

Integration:

  • agent.ts::callProvider → logs every non-local Claude call to ~/.kbot/teacher/traces.jsonl
  • cli.ts → 5 new subcommands: train-self | train-cycle | train-merge | train-grpo | train-agent-trace
  • ~/.zshrc → export KBOT_TEACHER_LOG=1 (always-on at shell open)
  • ~/Library/LaunchAgents/com.kernel.kbot-train-self.plist → Sundays 3am (dry-run; enable with launchctl load)

Known fragility:

  • kbot-discovery-daemon (PID 2491) auto-commits "evolution: kbot proposal" every few min. It previously wiped uncommitted files. Commit work fast. (Committed: 3 commits during this session — fea4acd5 → ff0498da.)

Corpus status (first run):

  • ~/.claude/projects/**/*.jsonl → 2,537 examples examined, 200 kept, mean score 0.700.
  • ~/.kbot/teacher/dataset-default.jsonl (353KB) ready for MLX.
  • Teacher traces file: empty until future kbot sessions run through new middleware.

Live run (background task b8h2y33gf):

  • Command: kbot train-self --mode default --max-examples 150 --iters 200 --num-layers 8
  • Base: mlx-community/Qwen2.5-Coder-7B-Instruct-4bit (first-run HF download ~4GB)
  • Log: ~/.kbot/teacher/train-self.log
  • Output model: kernel-self:v<timestamp> in Ollama
  • Test when done: ollama run kernel-self:<ts>

Shipped (not yet versioned/published)

6 phases of local fine-tuning infra, end-to-end on M3 Max 36GB:

New files under packages/kbot/src/:

  • teacher-logger.ts — middleware that captures every Claude call as (prompt, response, tools, outcome) to ~/.kbot/teacher/traces.jsonl. PII/key scrubber (sk-ant, ghp_, AIza, JWT, email, IP). Size rotation at 500MB. Wired into agent.ts::callProvider at line ~820.
  • train-curate.ts — scores + dedupes traces into training JSONL. Modes: default / reasoning / agent-trace / code-only.
  • train-self.ts — end-to-end pipeline: curate → mlx_lm.lora → mlx_lm.fuse → quantize → Ollama deploy. Default bases per mode (Qwen2.5-Coder-7B / DeepSeek-R1-Distill-7B / Qwen2.5-Coder-14B).
  • train-cycle.ts — DeepSeek-R1 Distill style on-policy loop: student (Ollama) generates → Claude grades with JSON rubric → corrected pairs append to corrections.jsonl → optional retrain.
  • train-agent-trace.ts — reformats tool-use traces with explicit <think>/<tool>/<args>/<result>/<answer> tokens for Phase 4 specialization.
  • train-merge.ts — MergeKit wrapper (TIES / SLERP / DARE / linear). Default kbot triad: qwen-coder + deepseek-r1 + self. MoE swap path documented (DeepSeek-V2-Lite-16B, Qwen3-MoE).
  • train-grpo.ts — GRPO rollouts with verifiable rewards: regex-match, json-valid, build-pass, test-pass, lint-pass, custom. Group-relative advantage calc. Rollouts persist to JSONL; external runner hookup via --runner-cmd.

CLI commands registered in cli.ts before program.parse:

  • kbot train-self --mode [default|reasoning|agent-trace|code-only]
  • kbot train-cycle --student --teacher --samples --threshold --retrain
  • kbot train-merge [--default | --method ties/slerp/dare_ties/linear]
  • kbot train-grpo --student --group-size --runner-cmd
  • kbot train-agent-trace --min-tools --verified-only

Validated:

  • npm run typecheck clean
  • kbot --help shows all 5 new commands
  • Teacher-logger end-to-end test persists trace and scrubs sk-ant-api03-... → sk-ant-<REDACTED>

Research grounding (2025–2026):

  • s1/s1.1 (1K curated reasoning traces) → reasoning mode
  • DeepSeek-R1 Distill (on-policy student+teacher) → train-cycle
  • Magpie/Genstruct (instruction back-translation) → curator approach
  • Agent-R / SWE-Gym (tool-token SFT) → agent-trace mode
  • MergeKit / TIES / SLERP → train-merge
  • GRPO (DeepSeek-Math) → train-grpo
  • DeepSeek-V2-Lite / Qwen3-MoE → MoE swap path docs

Hardware target confirmed: M3 Max, 36GB unified, MLX 0.29.3 + mlx-lm 0.29.1 installed. ~350–500 tok/s expected for 7B LoRA.

Pending to fully close loop:

  • Real user sessions must run through the (updated) callProvider to populate ~/.kbot/teacher/traces.jsonl. Today: zero traces yet.
  • Observer log (~/.kbot/observer/session.jsonl) is tool-call-only, NOT prompt/response. It's a good source for agent-trace mode after grouping by session, but curator's default mode currently skips it correctly (yields 0 examples until teacher-log accumulates).
  • MergeKit / mlx_lm.fuse / llama.cpp convert binaries not yet checked; pipeline fails gracefully when missing.
  • Bench harness (Claude-as-judge on 20 held-out tasks) not yet written — plan says write it at Phase 1 ship; queued for next session.
  • npm publish + version bump (v3.98.0 "teacher logger" → v3.99.0 "train-self" → v4.0.0 "reasoning distill") not done.

Next session pickup:

  1. Use kbot for 1–2 days so teacher-log populates (~200–500 traces minimum).
  2. Run kbot train-self --dry-run --mode default to confirm curator emits a dataset.
  3. Install mergekit (pip install mergekit) to unblock train-merge.
  4. Write the bench harness.
  5. Version-bump + publish v3.98.0 with teacher-logger alone (Phase 0 ships standalone value: forever-free dataset accumulation).

Previous Session (2026-04-05 night) — STREAM V2 + INTELLIGENCE COORDINATOR

Shipped: v3.97.0 (npm published, GitHub pushed)

Stream V2 — 6 new systems (~5,700 lines, 26 new tools):

  • PCM audio engine: oscillators, ADSR, chiptune sequencer, 7 SFX types
  • Stream overlay: follow/raid/sub/donation/achievement alerts, goal bars, ticker, info bar
  • Weather system: 12 weather types, day/night cycle, mood-coupled particles
  • AI chat: Gemma 4 via Ollama, viewer memory, topic tracking, 4 response modes
  • VOD system: auto-record, highlight detection, clips, YouTube upload
  • Chat commands: 31 commands, XP economy, inventory, polls, boss fights, giveaways

Intelligence Coordinator (886 lines):

  • 4-phase cognitive loop wired into agent.ts
  • Pre-execution: learned routing, confidence gating, graph context, anticipation
  • Tool oversight: pattern warnings, destructive tool detection, graph logging
  • Post-response: heuristic self-eval (no LLM), pattern extraction, drive updates
  • Cross-session consolidation every 10 interactions (selfTrain, graph prune, behaviour rules)

Stream Renderer Improvements:

  • Pulsing red LIVE dot, weather/time in header, viewer count badge
  • Ambient scenery: trees, rocks, grass tufts, flowers, clouds (procedural, camera-relative)
  • Ground-level atmospheric haze, improved speech bubble position
  • Chat panel accent border, branded kernel.chat watermark
  • All v2 systems wired into render loop and chat processing chain
  • Stream auditor agent created (.claude/agents/stream-auditor.md)

Stats: v3.97.0 on npm | 5,105 downloads/week | 19,384 total | 787+ tools | 37 registered users

Stream: Live on Twitch (kernelchatkbot) + Kick. Rumble needs fresh key each session.

Groq: $0.25 invoice for March (llama-3.1-8b-instant). Account needs login to check which email.

Pending:

  • Tune PCM audio (enabled but untested on stream)
  • Test all 31 chat commands live on Twitch
  • X API tokens still expired (social posting blocked)
  • Video demo still needed
  • kernel.chat site needs tool count update (787+)
  • Daily stream auditor cron (session-only, needs launchd for persistence)

Previous Session (2026-04-02 full day) — MEGA BUILD + ABLETON + LEADERBOARD

Two-day summary (Apr 1-2): v3.62.0 → v3.73.3

Shipped:

  • Dream engine (7-tier memory cascade, dreaming daemon)
  • Buddy system (8 species, evolution, achievements, chat, trading cards, leaderboard)
  • Voice input, memory scanner, user behavior learning
  • Service watchdog, morning briefing
  • Multi-agent finance, music gen, AI interpretability, cyber threat intel
  • A2A protocol, Ollama 0.19 detection, DeepSeek V4 provider
  • KBotBridge Remote Script (programmatic Ableton device loading via Browser API)
  • iPhone control (Apple ecosystem + mobile-mcp)
  • Buddy leaderboard on kernel.chat/#/leaderboard
  • Coldplay Clocks session + Empire of the Sun / Tame Impala build
  • Install script (kernel.chat/install.sh)
  • CI fixed, demo GIF re-recorded, README updated everywhere

Stats: 764+ tools, 10 stars, 1,929 downloads (Apr 1), v3.73.3

Users: Jae (portfolio analysis emails), Harrison (install help), Ray (agent setup)

Ableton: KBotBridge on port 9997, AbletonOSC enabled. Two sessions saved.

Pending:

  • Collective intelligence plan (partially built, not fully executed)
  • iPhone Developer Mode needs Xcode installed for full device control
  • X API tokens expired (social posting needs manual)
  • Video demo needs better recording (current GIF has issues)
  • kernel.chat site updated: scroll fixed, 764+ tools, responsive breakpoints

Previous Session (2026-04-02 afternoon) — HTTPS FIX + USER SUPPORT + MARKETING PUSH

User Issue: Harrison (hwmccormick123@gmail.com)

  • Harrison couldn't connect kbot to Claude Cowork — form requires https://, kbot serve only spoke HTTP
  • Fixed by adding native HTTPS support to kbot serve
  • Emailed him the fix via kernel-comms MCP
  • Email agent is live and running (launchd com.kernel.email-agent) — Harrison can reply and get AI responses via local Ollama

Shipped: kbot serve --https

  • serve.ts — Added --https flag with auto-generated self-signed TLS cert (~/.kbot/certs/)
  • cli.ts — Added --https, --cert <path>, --key <path> flags
  • ensureSelfSignedCert() — EC P-256 cert via openssl, 365-day validity, localhost + 127.0.0.1 SANs
  • Users can also provide custom certs: kbot serve --cert x.pem --key x.key
  • Clean build, clean typecheck

Marketing Push

  • HN post live: https://news.ycombinator.com/item?id=47622060 (Show HN: K:BOT — 738-tool terminal AI agent, plugs into Claude Cowork)
  • X thread drafted — 4 tweets in tools/social-posts-2026-04-02.md (X API tokens expired, needs manual post)
  • LinkedIn drafted — also in tools/social-posts-2026-04-02.md
  • Demo recording script created at tools/demo-recording.sh (asciinema + vhs + agg all installed)
  • Competitor intel: Skales (BSL-1.1, desktop GUI agent from Vienna, 6 HN points) — kbot differentiates on: true MIT open source, terminal-native, 738 tools, Claude Cowork connector, deeper local AI

Discovery Daemon Status

  • Running: 1,477 total runs, 608 pulses, 70 intel scans, 0 errors today
  • Email agent: running via launchd since 6:30 AM
  • Ollama: online with qwen2.5-coder:32b + 13 other models
  • Resend webhook: active, pointing to Supabase receive-email edge function

Stats

  • 738 registered tools
  • v3.71.0 on npm
  • npm: 4,799 downloads/week, 10 GitHub stars
  • 170 npm versions published

Not done

  • X thread needs manual posting (API tokens expired — needs X_API_KEY, X_API_SECRET, X_ACCESS_TOKEN, X_ACCESS_SECRET in .env)
  • Claude-in-Chrome extension not bridged to Claude Code terminal (separate MCP — not configured in ~/.claude/settings.json)
  • npm publish with HTTPS changes not yet done
  • Video demo still pending
  • LinkedIn post needs manual posting

Previous Session (2026-04-01 overnight) — CLAUDE CODE LEAK → DREAM ENGINE → NIGHT SHIFT

Claude Code Source Leak Study + Original Builds

What happened: Claude Code's full source (~512K lines TypeScript) leaked via source maps in npm package @anthropic-ai/claude-code@2.1.88. Studied the architecture, built original features inspired by patterns found.

Shipped: v3.63.0 — Dream Engine + Rival Intel

  • Dream Engine (dream.ts, 660 lines) — post-session memory consolidation via local Ollama, exponential decay aging, dream journal auto-injected into system prompt
  • 5 dream tools — dream_now, dream_status, dream_journal, dream_search, dream_reinforce
  • Rival Intel Agent (.claude/agents/rival-intel.md) — competitive intelligence on Claude Code architecture
  • CLI — kbot dream run/status/search/journal
  • Published npm + pushed GitHub

Shipped: v3.64.0 — Night Shift (buddy, voice, scanner)

  • Buddy System (buddy.ts, 513 lines) — 8 ASCII companion species, 5 moods, deterministic assignment, persistent naming
  • Voice Input (voice-input.ts, 466 lines) — local STT via whisper.cpp + Ollama, push-to-talk
  • Memory Scanner (memory-scanner.ts, 564 lines) — passive in-session detection of corrections, preferences, project facts. Hooks into addTurn(), scans every 5 turns.
  • 6 new tools — buddy_status, buddy_rename, voice_listen, voice_status, memory_scan_status, memory_scan_toggle
  • Published npm + pushed GitHub

Stats

  • 686 registered tools (was 671 at session start)
  • v3.64.0 on npm (was v3.62.0 at session start)
  • npm: 4,806 downloads/week, 6 GitHub stars

Previous Session (2026-03-31 night) — ABLETON BEAT SESSION: Kalan.FrFr x Don Toliver

Built a full beat in Ableton Live 12 via kbot OSC + AppleScript automation

Session: 142 BPM | F minor | Fm - Db - Ab - Eb progression

Active tracks (all Roland Cloud):

  1. TR-808 DRUMS (track 5) — 81-note pattern: bouncy syncopated kick, clap on 2&4, hi-hats w/ triplet rolls, rimshot, conga
  2. ZENOLOGY 808 BASS (track 17) — 10-note sub bass pattern, F1→Db2→Ab1→Eb2 with ghost re-triggers
  3. ZENOLOGY MELODY (track 18) — 14-note dreamy pluck motif, Ab→Bb→C movement
  4. XV-5080 PAD (track 13) — 16-note wide chord voicings, one per bar
  5. ZENOLOGY COUNTER (track 20) — 8-note subtle F5/Eb5 fills

What worked:

  • kbot AbletonOSC tools: transport, track rename, clip create, MIDI write, clip fire, mixer — all solid
  • Plugin loading via AppleScript: View > Search in Browser → type name → keyboard Down arrows → Return
  • ZENOLOGY (not FX) loads with 3 Down arrows to skip past FX presets
  • TR-808 loaded via Python Quartz drag from browser to session view
  • cliclick installed via Homebrew for macOS mouse automation

What didn't work:

  • load_plugin OSC endpoint — always times out (custom kbot extension, not in standard AbletonOSC)
  • CGEvent mouse drags — coordinates didn't match screen positions (Retina scaling mismatch)
  • IDE terminal steals focus from Ableton on every bash command — solved by running clicks inside osascript blocks
  • Loading multiple heavy Roland plugins in sequence can crash Ableton

Presets still needed (user will do manually):

  • ZENOLOGY tracks need bass/pluck/texture presets selected
  • XV-5080 needs a pad preset selected
  • Add reverb + delay sends on melody and counter tracks

Previous Session (2026-03-31 afternoon) — SHIP v3.59.0 + COMPUTER-USE VERIFIED

Shipped to GitHub, npm pending auth

Published to GitHub

  • v3.59.0 committed and pushed (96 files, +8,157 lines)
  • Commit: ea31a96b + 1733988a (serum2 registration fix)

What shipped

  1. 5 bug fixes — concurrent session state (memory.ts → Map), selfTrain guard, DNS rebinding SSRF, Gemini/Cohere tool warning, edit_file full-context diff
  2. Session isolation — serve.ts creates unique session per HTTP request, destroys after
  3. 9 M4L devices — auto-pilot, bass-synth, dj-fx, drum-synth, genre-morph, hat-machine, pad-synth, riser-engine, sidechain
  4. DJ Set Builder — registered in tool index
  5. Serum 2 Preset tool — was missing from index, now registered
  6. Computer-use expansion — 866+ lines added to computer.ts
  7. Ableton Live integration — OSC-based class in integrations/

Computer-Use MCP Verified

  • list_granted_applications — works
  • request_access — works (granted Finder)
  • screenshot — works (captured desktop)
  • Significance: kbot goes from terminal-only to full desktop agent

Stats

  • 698 tests passing (vitest), 0 type errors
  • 670+ registered tools
  • npm publish blocked — token expired, needs npm login

Not done

  • npm publish (needs auth)
  • GitHub release (can do next session)
  • Show HN post
  • Video demo

Previous Session (2026-03-31 night) — CODE QUALITY & CONCURRENCY FIXES

Night shift — 5 bug fixes, 1 security fix, 8 new tests

Bug Fixes Applied

  1. Concurrent state in memory.ts — Replaced single sessionHistory array with Map<string, ConversationTurn[]> keyed by session ID. All functions accept optional sessionId param (default 'default'). Added destroySession() for serve mode cleanup. CLI unchanged.
  2. Concurrent state in learning.ts — Added concurrency docs (shared state is intentional for learning). Added selfTrainRunning guard to prevent overlapping selfTrain() runs.
  3. DNS rebinding in fetch.ts — SSRF protection now resolves hostname via dns.lookup() and checks resolved IP against blocked ranges. Domains pointing to 127.0.0.1 are now caught.
  4. Gemini/Cohere silent degradation — Added upfront warning when these providers are used with tools: "provider doesn't support native tool calling — tools will be parsed from text output".
  5. edit_file diff preview — Now passes full file content to diff preview (was passing just the matched fragment). Diff algorithm shows 3 lines of context with ... separators.

Serve Mode Session Isolation (bonus)

  • Wired sessionId through AgentOptions → runAgent → all memory calls
  • serve.ts now creates a unique session per HTTP request and destroys it after
  • Concurrent /stream requests no longer share conversation history

Tests

  • 8 new session isolation tests in memory.test.ts
  • 698 tests passing (up from 690)
  • 0 type errors (tsc --noEmit clean)
  • Clean build

Stats

  • 600 registered tools (verified: 549 in tools/ + 51 elsewhere)
  • README "600+" claim is accurate

Previous Session (2026-03-29/30) — ABLETON BRAIN + M4L + TERMINAL CONTROL

Built AI music production system + full terminal control for platform ops

Published: @kernel.chat/kbot@3.54.0, 3.55.0, 3.56.0 (npm + GitHub)

Terminal Control System (v3.56.0)

  • 6 new CLI command groups, 32 new tools — everything manageable from terminal
  • kbot admin — users, billing (Stripe), moderation, platform stats (6 tools)
  • kbot monitor — live health dashboard, logs, uptime checks, alerts (4 tools)
  • kbot deploy — all-in-one ship: web + functions + npm + release (5 tools)
  • kbot analytics — npm downloads, GitHub traffic, user growth, revenue (5 tools)
  • kbot env — secrets management, sync, rotation guides (5 tools)
  • kbot db — backup, inspect, SQL, migrations, health check (6 tools)
  • Fixed pre-existing duplicate 'sessions' command bug
  • GitHub release: https://github.com/isaacsight/kernel/releases/tag/v3.56.0

Key Builds

  • 9 M4L devices, DJ set builder, Serum 2 preset tool
  • M4L bridge working on TCP 9999
  • 30-min premixed trap set (664 bars, 16,778 notes, F minor, 144 BPM)
  • 7 drum patterns, Roland Cloud instruments

Previous Sessions

2026-03-26: UNIVERSITY SESSION

  • 4 npm publishes. v3.42.0 → v3.45.0
  • 114 science tools across 11 lab files
  • See git history for full details

2026-03-24: MEGA SESSION

  • 13 npm publishes. v3.26.0 → v3.31.2
  • Finance stack, cybersecurity, self-defense, cognitive systems
  • ~10,000 lines, 350+ tools, 26 agents

2026-03-22 → 2026-03-23: SYNTH Game Build

  • 60+ source files, 45K+ lines at kernel.chat/#/play

Prior

See git history.