@10et/cli 1.0.0

package/packages/pi/extensions/training-buffer-tool.ts

@@ -0,0 +1,376 @@
+/**
+ * Training Buffer Tool
+ *
+ * Wraps the RL training buffer for the Pi agent. Two modes:
+ * 1. Automatic: captures (state, action, reward) tuples on agent_end events
+ * 2. Manual: tenet_training_buffer tool lets the agent record tuples explicitly
+ *
+ * Also provides /tuples command for inspecting the buffer and /mine command
+ * for extracting tuples from project history.
+ *
+ * @purpose RL training buffer — automatic + manual tuple capture, mining from history
+ */
+
+import { existsSync, readFileSync } from "fs"
+import { join } from "path"
+import type { PiContext, JflConfig, AgentEndEvent } from "./types.js"
+import { jflImport } from "./jfl-resolve.js"
+import { emitCustomEvent } from "./map-bridge.js"
+
+let projectRoot = ""
+
+async function getTrainingBuffer(): Promise<any> {
+  try {
+    // @ts-ignore — resolved from jfl package at runtime
+    const { TrainingBuffer } = await jflImport("training-buffer")
+    return new TrainingBuffer(projectRoot)
+  } catch {
+    return null
+  }
+}
+
+async function getTupleMiner(): Promise<any> {
+  try {
+    // @ts-ignore — resolved from jfl package at runtime
+    return await jflImport("tuple-miner")
+  } catch {
+    return null
+  }
+}
+
+export async function setupTrainingBufferTool(ctx: PiContext, _config: JflConfig): Promise<void> {
+  projectRoot = ctx.session.projectRoot
+
+  ctx.registerTool({
+    name: "tenet_training_buffer",
+    description: "Record a training tuple (state, action, reward) in the RL training buffer. Use after completing a task to capture the outcome for policy head training.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        action_type: {
+          type: "string",
+          description: "Type of action taken",
+          enum: ["fix", "refactor", "feature", "test", "config", "experiment"],
+        },
+        description: {
+          type: "string",
+          description: "What was done (1-2 sentences)",
+        },
+        files: {
+          type: "string",
+          description: "Comma-separated list of files changed",
+        },
+        scope: {
+          type: "string",
+          description: "Scope of change",
+          enum: ["small", "medium", "large"],
+        },
+        outcome: {
+          type: "string",
+          description: "Did it improve things?",
+          enum: ["improved", "neutral", "regressed"],
+        },
+        delta: {
+          type: "string",
+          description: "Numeric score delta if known (e.g. 0.02 or -0.01)",
+        },
+      },
+      required: ["action_type", "description", "outcome"],
+    },
+    async handler(input) {
+      const {
+        action_type,
+        description,
+        files,
+        scope,
+        outcome,
+        delta: deltaStr,
+      } = input as {
+        action_type: string
+        description: string
+        files?: string
+        scope?: string
+        outcome: string
+        delta?: string
+      }
+
+      const tb = await getTrainingBuffer()
+      if (!tb) {
+        return "Training buffer unavailable. Ensure jfl-cli is built (npm run build)."
+      }
+
+      const delta = deltaStr ? parseFloat(deltaStr) : (
+        outcome === "improved" ? 0.02 :
+        outcome === "regressed" ? -0.01 :
+        0.0
+      )
+
+      const existing: any[] = tb.read()
+      const recentDeltas = existing.slice(-10).map((e: any) => e.reward.composite_delta)
+      const currentScore = existing.length > 0
+        ? existing[existing.length - 1].state.composite_score + existing[existing.length - 1].reward.composite_delta
+        : 0
+
+      const entry = tb.append({
+        agent: "pi-agent",
+        state: {
+          composite_score: currentScore,
+          dimension_scores: {},
+          tests_passing: 0,
+          tests_total: 0,
+          trajectory_length: existing.length,
+          recent_deltas: recentDeltas,
+          agent: "pi-agent",
+        },
+        action: {
+          type: action_type as "fix" | "refactor" | "feature" | "test" | "config" | "experiment",
+          description,
+          files_affected: files ? files.split(",").map(f => f.trim()) : [],
+          scope: (scope || "medium") as "small" | "medium" | "large",
+          branch: ctx.session.branch,
+        },
+        reward: {
+          composite_delta: delta,
+          dimension_deltas: {},
+          tests_added: 0,
+          quality_score: outcome === "improved" ? 1.0 : outcome === "regressed" ? 0.0 : 0.5,
+          improved: outcome === "improved",
+        },
+        metadata: {
+          branch: ctx.session.branch,
+          source: "manual",
+        },
+      })
+
+      await emitCustomEvent(ctx, "training:tuple:added", { id: entry.id, delta })
+
+      return [
+        `Training tuple recorded: ${entry.id}`,
+        `  Action: ${action_type} — ${description.slice(0, 60)}`,
+        `  Delta: ${delta >= 0 ? "+" : ""}${delta.toFixed(4)}`,
+        `  Buffer now has ${existing.length + 1} tuples`,
+      ].join("\n")
+    },
+  })
+
+  ctx.registerTool({
+    name: "tenet_mine_tuples",
+    description: "Mine training tuples from project history (journals, MAP events, eval results, telemetry). Extracts (state, action, reward) data that can train the policy head.",
+    inputSchema: {
+      type: "object",
+      properties: {
+        source: {
+          type: "string",
+          description: "What to mine from",
+          enum: ["journals", "flows", "sessions", "evals", "all"],
+        },
+        write: {
+          type: "string",
+          description: "Write mined tuples to training buffer? (yes/no, default: no)",
+          enum: ["yes", "no"],
+        },
+      },
+      required: ["source"],
+    },
+    async handler(input) {
+      const { source, write: writeFlag } = input as { source: string; write?: string }
+
+      const miner = await getTupleMiner()
+      if (!miner) {
+        return "Tuple miner unavailable. Ensure jfl-cli is built."
+      }
+
+      const tb = await getTrainingBuffer()
+      let tuples: any[] = []
+
+      switch (source) {
+        case "journals":
+          tuples = miner.mineJournalTuples(projectRoot)
+          break
+        case "flows":
+          tuples = miner.mineFlowTuples(projectRoot)
+          break
+        case "sessions":
+          tuples = miner.mineSessionTuples(projectRoot)
+          break
+        case "evals":
+          tuples = miner.mineEvalTuples(projectRoot)
+          break
+        case "all": {
+          const result = miner.mineAll({ dirs: [projectRoot], telemetry: true })
+          tuples = result.tuples
+          break
+        }
+      }
+
+      if (writeFlag === "yes" && tb && tuples.length > 0) {
+        for (const t of tuples) {
+          tb.append(t)
+        }
+        await emitCustomEvent(ctx, "training:mined", { count: tuples.length, source })
+      }
+
+      const improved = tuples.filter((t: any) => t.reward.improved).length
+      const avgDelta = tuples.length > 0
+        ? tuples.reduce((sum: number, t: any) => sum + t.reward.composite_delta, 0) / tuples.length
+        : 0
+
+      const lines = [
+        `Mined ${tuples.length} tuples from ${source}`,
+        `  Improved: ${improved}/${tuples.length} (${tuples.length > 0 ? ((improved / tuples.length) * 100).toFixed(0) : 0}%)`,
+        `  Avg delta: ${avgDelta >= 0 ? "+" : ""}${avgDelta.toFixed(4)}`,
+        `  Agents: ${[...new Set(tuples.map((t: any) => t.agent))].join(", ") || "none"}`,
+      ]
+
+      if (writeFlag === "yes") {
+        lines.push(`  Written to training buffer ✓`)
+      } else {
+        lines.push(`  Dry run — pass write="yes" to persist`)
+      }
+
+      return lines.join("\n")
+    },
+  })
+
+  ctx.registerCommand({
+    name: "tuples",
+    description: "Show training buffer stats and recent tuples",
+    async handler(_args, ctx) {
+      const tb = await getTrainingBuffer()
+      if (!tb) {
+        ctx.ui.notify("Training buffer unavailable.", { level: "warn" })
+        return
+      }
+
+      const stats = tb.stats()
+      const entries: any[] = tb.read()
+      const recent = entries.slice(-5)
+
+      const lines = [
+        `Training Buffer: ${stats.total} tuples`,
+        `  Avg reward: ${stats.avgReward >= 0 ? "+" : ""}${stats.avgReward.toFixed(4)}`,
+        `  Improvement rate: ${(stats.improvedRate * 100).toFixed(1)}%`,
+        `  By agent: ${Object.entries(stats.byAgent).map(([k, v]) => `${k}=${v}`).join(", ") || "none"}`,
+        `  By source: ${Object.entries(stats.bySource).map(([k, v]) => `${k}=${v}`).join(", ") || "none"}`,
+        "",
+        "Recent tuples:",
+      ]
+
+      for (const e of recent as any[]) {
+        const sign = e.reward.composite_delta >= 0 ? "+" : ""
+        lines.push(`  [${sign}${e.reward.composite_delta.toFixed(4)}] ${e.action.type}: ${e.action.description.slice(0, 60)}`)
+      }
+
+      ctx.ui.notify(lines.join("\n"), { level: "info" })
+    },
+  })
+
+  ctx.registerCommand({
+    name: "mine",
+    description: "Mine training tuples from project history and write to buffer",
+    async handler(args, ctx) {
+      const source = args.trim() || "all"
+      const miner = await getTupleMiner()
+      if (!miner) {
+        ctx.ui.notify("Tuple miner unavailable.", { level: "warn" })
+        return
+      }
+
+      const tb = await getTrainingBuffer()
+      if (!tb) {
+        ctx.ui.notify("Training buffer unavailable.", { level: "warn" })
+        return
+      }
+
+      const result = miner.mineAll({
+        dirs: [projectRoot],
+        telemetry: source === "all",
+      })
+
+      for (const t of result.tuples) {
+        tb.append(t)
+      }
+
+      await emitCustomEvent(ctx, "training:mined", {
+        count: result.tuples.length,
+        stats: result.stats,
+      })
+
+      ctx.ui.notify(
+        [
+          `Mined ${result.stats.totalMined} tuples:`,
+          `  Journals: ${result.stats.journalTuples}`,
+          `  Sessions: ${result.stats.sessionTuples}`,
+          `  Flows: ${result.stats.flowTuples}`,
+          `  Evals: ${result.stats.evalTuples}`,
+          `  Telemetry: ${result.stats.telemetryTuples}`,
+          `Written to training buffer ✓`,
+        ].join("\n"),
+        { level: "info" }
+      )
+    },
+  })
+}
+
+export async function onTrainingAgentEnd(
+  ctx: PiContext,
+  event: AgentEndEvent
+): Promise<void> {
+  const turnCount = event.messages?.length ?? event.turnCount ?? 0
+  // Only log tuples when actual work happened (files changed).
+  // Previously this logged every session as "improved" with a fake +0.01 delta,
+  // polluting the training buffer with 187+ garbage noops.
+  const filesChanged = event.filesChanged || []
+  if (turnCount < 2 || filesChanged.length === 0) return
+
+  const tb = await getTrainingBuffer()
+  if (!tb) return
+
+  const existing: any[] = tb.read()
+  // Safe access — reward or composite_delta may be missing on older entries
+  const recentDeltas = existing.slice(-10).map((e: any) => e.reward?.composite_delta ?? 0)
+  const lastEntry = existing.length > 0 ? existing[existing.length - 1] : null
+  const currentScore = lastEntry
+    ? (lastEntry.state?.composite_score ?? 0) + (lastEntry.reward?.composite_delta ?? 0)
+    : 0
+
+  const exitReason = event.exitReason as string | undefined
+  const isSuccess = exitReason !== "error" && exitReason !== "timeout"
+
+  const entry = tb.append({
+    agent: "pi-agent",
+    state: {
+      composite_score: currentScore,
+      dimension_scores: {},
+      tests_passing: 0,
+      tests_total: 0,
+      trajectory_length: existing.length,
+      recent_deltas: recentDeltas,
+      agent: "pi-agent",
+    },
+    action: {
+      type: filesChanged.some(f => f.endsWith(".test.ts") || f.endsWith(".test.js")) ? "test"
+        : filesChanged.some(f => f.includes("config") || f.includes(".json")) ? "config"
+        : "feature",
+      description: `Agent turn: ${turnCount} turns, ${filesChanged.length} files changed`,
+      files_affected: filesChanged.slice(0, 20),
+      scope: filesChanged.length > 10 ? "large" : filesChanged.length > 3 ? "medium" : "small",
+      branch: ctx.session.branch,
+    },
+    reward: {
+      // Don't fabricate deltas — use 0 to indicate "no eval ran".
+      // Real deltas come from the RL agent loop (peter daily), not session logging.
+      composite_delta: 0,
+      dimension_deltas: {},
+      tests_added: 0,
+      quality_score: isSuccess ? 0.5 : 0.3,
+      improved: false,  // Only the RL eval loop can determine improvement
+    },
+    metadata: {
+      branch: ctx.session.branch,
+      source: "ci",
+    },
+  })
+
+  await emitCustomEvent(ctx, "training:tuple:auto", { id: entry.id })
+}

package/packages/pi/extensions/types.ts

@@ -0,0 +1,169 @@
+/**
+ * JFL Pi Extension Types
+ *
+ * Full type definitions for the JFL Pi extension layer.
+ * PiContext wraps Pi's real ExtensionAPI and exposes full TUI capabilities
+ * to all JFL sub-extensions.
+ *
+ * @purpose Type definitions for JFL Pi extension — full TUI + lifecycle access
+ */
+
+// ─── Tool / Command formats for JFL sub-extensions ───────────────────────────
+
+export interface JflToolDef {
+  name: string
+  label?: string
+  description: string
+  promptSnippet?: string
+  promptGuidelines?: string[]
+  inputSchema: {
+    type: "object"
+    properties: Record<string, { type: string; description?: string; enum?: string[] }>
+    required?: string[]
+  }
+  handler(input: Record<string, unknown>): Promise<string>
+  renderCall?(args: Record<string, any>, theme: PiTheme): any
+  renderResult?(result: any, options: { expanded: boolean; isPartial: boolean }, theme: PiTheme): any
+}
+
+export interface JflCommandDef {
+  name: string
+  description?: string
+  handler(args: string, ctx: PiContext): Promise<void> | void
+}
+
+// ─── Pi Theme interface (subset we use) ──────────────────────────────────────
+
+export interface PiTheme {
+  fg(color: string, text: string): string
+  bg(color: string, text: string): string
+  bold(text: string): string
+}
+
+// ─── Widget factory signature ────────────────────────────────────────────────
+
+export type WidgetFactory = (tui: any, theme: PiTheme) => {
+  render: (width: number) => string[]
+  invalidate: () => void
+}
+
+// ─── PiContext shim — passed to all JFL sub-extensions ───────────────────────
+
+export interface PiContext {
+  session: {
+    readonly projectRoot: string
+    readonly id: string
+    readonly branch: string
+  }
+
+  log(msg: string, level?: "debug" | "info" | "warn" | "error"): void
+  emit(event: string, data?: unknown): void
+  on(event: string, handler: (data: unknown) => void | Promise<void>): void
+
+  registerTool(tool: JflToolDef): void
+  registerCommand(cmd: JflCommandDef): void
+  registerShortcut(key: string, opts: { description: string; handler: () => Promise<void> | void }): void
+
+  ui: {
+    notify(msg: string, opts?: { level?: "info" | "warn" | "error" | "success" }): void
+    input(title: string, placeholder?: string): Promise<string | undefined>
+    confirm(title: string, message: string): Promise<boolean>
+    select<T extends string>(title: string, items: Array<{ value: T; label: string; description?: string }>): Promise<T | null>
+    editor(title: string, content: string): Promise<string | undefined>
+    custom<T>(factory: (tui: any, theme: PiTheme, keybindings: any, done: (result: T) => void) => any, opts?: { overlay?: boolean; overlayOptions?: any; onHandle?: (handle: any) => void }): Promise<T>
+    setWidget(id: string, content: string[] | WidgetFactory | undefined, opts?: { placement?: "aboveEditor" | "belowEditor" }): void
+    setStatus(key: string, text: string | undefined): void
+    setFooter(factory: ((tui: any, theme: PiTheme, footerData: any) => any) | undefined): void
+    setHeader?(factory: ((tui: any, theme: PiTheme) => { render: (width: number) => string[]; invalidate: () => void }) | undefined): void
+    setEditorText(text: string): void
+    theme: PiTheme
+    hasUI: boolean
+  }
+
+  pi: {
+    setSessionName(name: string): void
+    appendEntry(type: string, data?: unknown): void
+    setLabel(entryId: string, label: string | undefined): void
+    sendMessage(msg: { customType: string; content: string; display?: boolean; details?: any }, opts?: { triggerTurn?: boolean; deliverAs?: "steer" | "followUp" | "nextTurn" }): void
+    events: { on(event: string, handler: (data: unknown) => void): void; emit(event: string, data?: unknown): void }
+    getActiveTools(): string[]
+    getAllTools(): Array<{ name: string; description: string }>
+    exec(command: string, args: string[], opts?: { signal?: AbortSignal; timeout?: number }): Promise<{ stdout: string; stderr: string; code: number }>
+    sessionManager: any
+    model: any
+  }
+
+  cancel(): { cancel: true }
+}
+
+// ─── Pi event shapes ──────────────────────────────────────────────────────────
+
+export interface AgentStartEvent {
+  prompt?: string
+  model?: string
+}
+
+export interface AgentEndEvent {
+  messages?: unknown[]
+  turnCount?: number
+  model?: string
+  duration?: number
+  filesChanged?: string[]
+  toolsUsed?: string[]
+  exitReason?: "success" | "error" | "cancelled" | "max_turns"
+}
+
+export interface ToolExecutionEvent {
+  toolName?: string
+  tool?: string
+  input?: Record<string, any>
+  args?: Record<string, any>
+  result?: unknown
+  isError?: boolean
+  duration?: number
+}
+
+export interface TurnStartEvent {
+  turnIndex: number
+  timestamp: number
+}
+
+export interface TurnEndEvent {
+  turnIndex: number
+  message?: any
+  toolResults?: any[]
+}
+
+export interface ModelSelectEvent {
+  model: { provider: string; id: string }
+  previousModel?: { provider: string; id: string }
+  source: "set" | "cycle" | "restore"
+}
+
+// ─── JFL config ──────────────────────────────────────────────────────────────
+
+export interface JflConfig {
+  name?: string
+  type?: "gtm" | "service" | "portfolio"
+  setup?: "building-product" | "gtm-only" | "contributor"
+  working_branch?: string
+  portfolio_parent?: string
+  gtm_parent?: string
+  context_scope?: {
+    produces?: string[]
+    consumes?: string[]
+    denied?: string[]
+  }
+  pi?: {
+    max_peter_iterations?: number
+    enable_stratus?: boolean
+    enable_portfolio_sync?: boolean
+    disable_footer?: boolean
+    disable_shortcuts?: boolean
+    disable_notifications?: boolean
+    disable_onboarding?: boolean
+    disable_briefing?: boolean
+    onboarding_variant?: "v1" | "v2" | "v3"
+    auto_start?: boolean
+  }
+}