pantheon-opencode 1.5.0-beta.16 → 1.5.0-beta.18

package/src/pantheon/native-probe.ts

@@ -1,234 +0,0 @@
-/**
- * Native Probe — lazy capability detection for native task() delegation.
- *
- * The probe is invoked on the FIRST native delegation of a session (not at
- * startup). It creates a child session with a minimal "ping" prompt, waits
- * for idle, and classifies the output.
- *
- * Results are cached per probe instance (singleton guard):
- *   - UNAVAILABLE is TERMINAL — no retry.
- *   - AVAILABLE is also cached — no re-probe.
- *
- * Uses the OpenCode V1 session API (session.create + promptAsync) injected via
- * the parent context.
- *
- * @module native-probe
- */
-
-import { createPantheonLogger } from './logger.ts'
-
-const log = createPantheonLogger({ module: 'native-probe' })
-
-// ─── Types ─────────────────────────────────────────────────────────────
-
-/** Possible outcomes of a native capability probe. */
-export type NativeProbeResult = {
-  status: 'AVAILABLE' | 'UNAVAILABLE' | 'UNSUPPORTED'
-}
-
-/** Minimal V1 client shape needed by the probe (session.create + promptAsync). */
-export interface ProbeClient {
-  session: {
-    create: (input: { body: { parentID: string } }) => Promise<{ id: string }>
-    promptAsync: (input: {
-      path: { id: string }
-      body: { agent: string; parts: Array<{ type: 'text'; text: string }> }
-    }) => Promise<unknown>
-  }
-}
-
-/** Context injected into the probe function. */
-export interface NativeProbeContext {
-  client: ProbeClient
-  parentSessionID: string
-}
-
-/** Options for a single probe invocation. */
-export interface NativeProbeOptions {
-  /** Timeout in milliseconds. Default: 30000. */
-  timeoutMs?: number
-}
-
-/** The probe function signature. */
-export type NativeProbeFn = (
-  ctx: NativeProbeContext,
-  options?: NativeProbeOptions,
-) => Promise<NativeProbeResult>
-
-// ─── Helpers ───────────────────────────────────────────────────────────
-
-/** Default timeout for the probe. */
-const DEFAULT_TIMEOUT_MS = 30_000
-
-/** Sentinel value indicating a timeout occurred. */
-const TIMEOUT_SENTINEL = Symbol('timeout')
-
-/**
- * Wrap a promise with a timeout.
- * Returns the value on success, throws the original error on rejection,
- * or returns TIMEOUT_SENTINEL if the timeout fires first.
- */
-async function withTimeout<T>(
-  promise: Promise<T>,
-  ms: number,
-): Promise<T | typeof TIMEOUT_SENTINEL> {
-  return new Promise<T | typeof TIMEOUT_SENTINEL>((resolve, reject) => {
-    const timer = setTimeout(() => resolve(TIMEOUT_SENTINEL), ms)
-    void promise.then(
-      (value) => {
-        clearTimeout(timer)
-        resolve(value)
-      },
-      (err: unknown) => {
-        clearTimeout(timer)
-        reject(err)
-      },
-    )
-  })
-}
-
-/**
- * Extract text content from a promptAsync response.
- * Returns empty string if no recognizable content found.
- */
-function extractOutput(result: unknown): string {
-  if (result == null) return ''
-  if (typeof result === 'string') return result
-  if (typeof result === 'object') {
-    const obj = result as Record<string, unknown>
-    // Check common shapes: { output }, { content }, { text }
-    if (typeof obj.output === 'string') return obj.output
-    if (typeof obj.content === 'string') return obj.content
-    if (typeof obj.text === 'string') return obj.text
-  }
-  // Non-empty object → treat as content (accepted message, etc.)
-  if (typeof result === 'object' && result !== null) {
-    return 'acknowledged'
-  }
-  return ''
-}
-
-function isUnsupportedError(message: string): boolean {
-  const lower = message.toLowerCase()
-  return (
-    lower.includes('not available') ||
-    lower.includes('not supported') ||
-    lower.includes('unsupported') ||
-    lower.includes('method not found')
-  )
-}
-
-// ─── Implementation ────────────────────────────────────────────────────
-
-/**
- * Create a native probe function with per-instance caching.
- *
- * The probe is LAZY: it does not run at startup. Call it on the first
- * native delegation of the session. The result is cached — subsequent
- * calls return the cached value without re-probing.
- *
- * @returns A probe function that caches results per instance.
- */
-export function createNativeProbe(): NativeProbeFn {
-  let cachedResult: NativeProbeResult | null = null
-  let inFlight: Promise<NativeProbeResult> | null = null
-
-  return async (
-    ctx: NativeProbeContext,
-    options?: NativeProbeOptions,
-  ): Promise<NativeProbeResult> => {
-    // Return cached result if available (singleton guard)
-    if (cachedResult !== null) {
-      return cachedResult
-    }
-    // Concurrent first dispatches share one probe instead of creating multiple
-    // child sessions. The promise is cleared only after a terminal result.
-    if (inFlight !== null) return inFlight
-
-    inFlight = probe(ctx, options)
-    try {
-      return await inFlight
-    } finally {
-      inFlight = null
-    }
-  }
-
-  async function probe(
-    ctx: NativeProbeContext,
-    options?: NativeProbeOptions,
-  ): Promise<NativeProbeResult> {
-    const timeoutMs = options?.timeoutMs ?? DEFAULT_TIMEOUT_MS
-
-    // ── Phase 1: create child session ──────────────────────────────
-    let childID: string
-    try {
-      const createResult = await withTimeout(
-        ctx.client.session.create({ body: { parentID: ctx.parentSessionID } }),
-        timeoutMs,
-      )
-      if (createResult === TIMEOUT_SENTINEL) {
-        // Timeout during session.create → UNAVAILABLE
-        log.warn('[Native Probe] session.create timed out after', timeoutMs, 'ms')
-        const result: NativeProbeResult = { status: 'UNAVAILABLE' }
-        cachedResult = result
-        return result
-      }
-      childID = createResult.id
-    } catch (err: unknown) {
-      const msg = err instanceof Error ? err.message : String(err)
-      log.error('[Native Probe] session.create failed:', msg)
-      // Check for unsupported API patterns
-      if (isUnsupportedError(msg)) {
-        const result: NativeProbeResult = { status: 'UNSUPPORTED' }
-        cachedResult = result
-        return result
-      }
-      // Other creation failures → UNAVAILABLE
-      const result: NativeProbeResult = { status: 'UNAVAILABLE' }
-      cachedResult = result
-      return result
-    }
-
-    // ── Phase 2: send probe prompt ─────────────────────────────────
-    let promptResult: unknown
-    try {
-      const raw = await withTimeout(
-        ctx.client.session.promptAsync({
-          path: { id: childID },
-          body: { agent: 'probe', parts: [{ type: 'text', text: 'Reply with PONG.' }] },
-        }),
-        timeoutMs,
-      )
-      if (raw === TIMEOUT_SENTINEL) {
-        // Timeout during promptAsync → UNAVAILABLE
-        log.warn('[Native Probe] promptAsync timed out after', timeoutMs, 'ms')
-        const result: NativeProbeResult = { status: 'UNAVAILABLE' }
-        cachedResult = result
-        return result
-      }
-      promptResult = raw
-    } catch (err: unknown) {
-      const msg = err instanceof Error ? err.message : String(err)
-      log.error('[Native Probe] promptAsync failed:', msg)
-      if (isUnsupportedError(msg)) {
-        const result: NativeProbeResult = { status: 'UNSUPPORTED' }
-        cachedResult = result
-        return result
-      }
-      // promptAsync error → UNAVAILABLE
-      const result: NativeProbeResult = { status: 'UNAVAILABLE' }
-      cachedResult = result
-      return result
-    }
-
-    // ── Phase 3: classify output ───────────────────────────────────
-    const output = extractOutput(promptResult)
-    const hasContent = output.trim() !== ''
-
-    const result: NativeProbeResult = {
-      status: hasContent ? 'AVAILABLE' : 'UNAVAILABLE',
-    }
-    cachedResult = result
-    return result
-  }
-}

package/src/pantheon/step-cap.ts

@@ -1,110 +0,0 @@
-/**
- * R4 — Per-agent step caps (max_steps).
- *
- * Pure library. routing.yml `agents.<name>.max_steps` gives each agent a
- * step budget (e.g. explore: 25, reviewer: 25, oracle: 15). When an agent's
- * step count reaches max_steps it is forced to summarize-and-stop: the
- * delegation is either skipped with a capped summary (already at cap) or
- * the prompt is appended with a stop instruction (cap hit mid-dispatch).
- *
- * The tracker is PERMANENT-PER-PROCESS: counters accumulate for the process
- * lifetime and are intentionally NOT reset on successful delegations or
- * session end — a per-session reset would let an agent exceed its process
- * budget over a long session. `reset()` exists as a tested utility for
- * tests and future explicit reset points; no production caller wires it.
- *
- * @module step-cap
- */
-
-// ─── Types ─────────────────────────────────────────────────────────────
-
-/** Default step budget when routing.yml omits max_steps for an agent. */
-export const DEFAULT_MAX_STEPS = 25
-
-// ─── Stop instruction ──────────────────────────────────────────────────
-
-/**
- * The forced summarize-and-stop instruction injected into a prompt when an
- * agent's step count reaches max_steps.
- */
-export function buildStopInstruction(agent: string, maxSteps: number): string {
-  return (
-    `\n\n[STEP CAP REACHED — ${agent} at ${maxSteps} steps] ` +
-    `You have reached the maximum step budget for this delegation. ` +
-    `STOP working immediately and summarize what was accomplished (and ` +
-    `what remains, if any) as your final message. Do not start new work.`
-  )
-}
-
-/**
- * The text returned when a delegation is skipped because the agent is
- * already at its step cap — marks the delegation as capped and asks for a
- * summary of prior results instead of starting new work.
- */
-export function cappedSummary(agent: string, maxSteps: number): string {
-  return (
-    `[STEP CAP REACHED] ${agent} has reached its max_steps budget (${maxSteps}). ` +
-    `Delegation skipped — no new work was started. ` +
-    `Summarize prior results and stop.`
-  )
-}
-
-// ─── Tracker ───────────────────────────────────────────────────────────
-
-/**
- * Per-agent step counters with a max_steps budget. Keyed by lowercase agent
- * name (matching is case-insensitive). Agents without a configured budget
- * are never capped.
- */
-export class StepCapTracker {
-  private readonly steps = new Map<string, number>()
-  readonly maxStepsByAgent: Readonly<Record<string, number>>
-
-  constructor(maxStepsByAgent: Readonly<Record<string, number>> = {}) {
-    this.maxStepsByAgent = maxStepsByAgent
-  }
-
-  /** The configured budget for an agent, or undefined when uncapped. */
-  maxStepsFor(agent: string): number | undefined {
-    return this.maxStepsByAgent[agent.toLowerCase()]
-  }
-
-  /** Current step count for an agent. */
-  stepCount(agent: string): number {
-    return this.steps.get(agent.toLowerCase()) ?? 0
-  }
-
-  /** Whether the agent has reached (or exceeded) its max_steps budget. */
-  isCapped(agent: string): boolean {
-    const max = this.maxStepsFor(agent)
-    if (max === undefined) return false
-    return this.stepCount(agent) >= max
-  }
-
-  /**
-   * Record one step for an agent. Returns the new count and whether this
-   * step hit the cap (capped === true exactly when steps >= maxSteps).
-   */
-  recordStep(agent: string): {
-    steps: number
-    capped: boolean
-    maxSteps: number | undefined
-  } {
-    const key = agent.toLowerCase()
-    const max = this.maxStepsFor(agent)
-    const steps = (this.steps.get(key) ?? 0) + 1
-    this.steps.set(key, steps)
-    return { steps, capped: max !== undefined && steps >= max, maxSteps: max }
-  }
-
-  /**
-   * Reset an agent's step count.
-   *
-   * Intentionally NOT wired in production: the tracker is permanent-per-
-   * process (see module docstring) so agents cannot exceed their process
-   * budget via resets. Exists for tests and future explicit reset points.
-   */
-  reset(agent: string): void {
-    this.steps.delete(agent.toLowerCase())
-  }
-}