Merge branch 'main' into feat/perf-pass

Reviewed-on: #57

packages/core/src/agent/memory-v0.test.ts

@@ -0,0 +1,71 @@
+import { describe, expect, it } from 'vitest'
+
+import type { DirectiveEntry } from '../directives/record-directive-v0.js'
+import type { Focus, ScheduledItem } from '../scheduler/scheduler-v0.js'
+import {
+  activeDirectives,
+  assembleHotContext,
+  estimateTokens,
+  HOT_CONTEXT_BUDGET_TOKENS,
+} from './memory-v0.js'
+
+const item = (number: number, title: string): ScheduledItem =>
+  ({ number, title, labels: [], order: 0, startDay: 0, endDay: 1, durationDays: 1, blockedBy: [], blocks: [], critical: false, rationale: '' })
+
+const focus: Focus = { now: item(7, 'Fix lifecycle inference'), next: item(8, 'Webhook listener'), later: null }
+
+const directive = (over: Partial<DirectiveEntry>): DirectiveEntry => ({
+  id: 'd1',
+  ts: '2026-02-10T09:00:00Z',
+  status: 'accepted',
+  kind: 'note',
+  quote: 'pilots come first',
+  ...over,
+})
+
+describe('memory-v0 (#27)', () => {
+  it('estimateTokens is a slight over-estimate (~4 chars/token)', () => {
+    expect(estimateTokens('')).toBe(0)
+    expect(estimateTokens('abcd')).toBe(1)
+    expect(estimateTokens('a'.repeat(4001))).toBe(1001)
+  })
+
+  it('activeDirectives keeps accepted/amended, most-recent-first', () => {
+    const ds = [
+      directive({ id: 'a', ts: '2026-02-01T00:00:00Z', status: 'accepted', quote: 'old' }),
+      directive({ id: 'b', ts: '2026-02-11T00:00:00Z', status: 'amended', quote: 'new' }),
+      directive({ id: 'c', ts: '2026-02-12T00:00:00Z', status: 'withdrawn', quote: 'gone' }),
+      directive({ id: 'd', ts: '2026-02-09T00:00:00Z', status: 'proposed', quote: 'maybe' }),
+    ]
+    expect(activeDirectives(ds).map((d) => d.quote)).toEqual(['new', 'old'])
+  })
+
+  it('assembles hot context under the 2k budget even with a huge charter', () => {
+    const huge = 'Charter line that goes on and on. '.repeat(2000) // ~34k tokens
+    const ds = Array.from({ length: 50 }, (_, i) =>
+      directive({ id: `d${i}`, ts: `2026-02-${String((i % 27) + 1).padStart(2, '0')}T00:00:00Z`, quote: `directive number ${i}` }),
+    )
+    const out = assembleHotContext({ charter: huge, directives: ds, focus })
+
+    expect(estimateTokens(out)).toBeLessThanOrEqual(HOT_CONTEXT_BUDGET_TOKENS)
+    // focus (tiny) is always kept; the charter is the part that gets truncated
+    expect(out).toContain('## Focus')
+    expect(out).toContain('#7 Fix lifecycle inference')
+    expect(out).toContain('…') // charter was clamped
+    // at least some recent directives survived
+    expect(out).toContain('## Active directives')
+  })
+
+  it('never inlines ticket bodies — only numbers + titles appear for focus', () => {
+    // The assembler takes no issue bodies by construction; focus shows #number title only.
+    const out = assembleHotContext({ charter: 'Ship the beta.', directives: [directive({})], focus })
+    expect(out).toContain('Now: #7 Fix lifecycle inference')
+    expect(out).toContain('[note] pilots come first')
+    expect(out).not.toMatch(/body|description|comment/i)
+  })
+
+  it('degrades to just focus when there is no charter or directives', () => {
+    const out = assembleHotContext({ charter: '', directives: [], focus })
+    expect(out).toBe(['## Focus', 'Now: #7 Fix lifecycle inference', 'Next: #8 Webhook listener', 'Later: —'].join('\n'))
+  })
+})

packages/core/src/agent/memory-v0.ts

@@ -0,0 +1,102 @@
+/**
+ * Memory layers, v0 (#27). Reginald's context is tiered so the model always sees
+ * what matters without ever copying ticket data into the prompt:
+ *
+ * - HOT (this module) — charter + active directives + the focus snapshot, packed
+ *   under a hard token budget. Assembled fresh each turn; it's the system-prompt seed.
+ * - WARM — the append-only directive/event ledger + periodic digest, in pm-state.
+ *   Not inlined; summarized on demand.
+ * - COLD — gitea + the sidecar, reached through `query_project` tools. Ticket bodies,
+ *   comments, and per-issue detail live here and are NEVER copied into memory —
+ *   the model fetches them by number when it needs them.
+ *
+ * The invariant: HOT stays under budget, and nothing ticket-shaped is inlined.
+ */
+
+import type { DirectiveEntry } from '../directives/record-directive-v0.js'
+import type { Focus } from '../scheduler/scheduler-v0.js'
+
+/** The hot layer's hard ceiling (#27: hot context assembles under 2k tokens). */
+export const HOT_CONTEXT_BUDGET_TOKENS = 2000
+
+/**
+ * Tokenizer-free estimate (~4 chars/token). Deliberately a slight over-estimate so
+ * a real tokenizer never exceeds what this predicts — the budget stays safe.
+ */
+export function estimateTokens(text: string): number {
+  return Math.ceil(text.length / 4)
+}
+
+/** Directives that still bind: accepted or amended, most-recent-first. */
+export function activeDirectives(all: DirectiveEntry[]): DirectiveEntry[] {
+  return all
+    .filter((d) => d.status === 'accepted' || d.status === 'amended')
+    .slice()
+    .sort((a, b) => (a.ts < b.ts ? 1 : a.ts > b.ts ? -1 : 0))
+}
+
+export interface HotContextInputs {
+  /** The project charter markdown (hot-memory seed). */
+  charter: string
+  /** The directive ledger (any status — filtered to active here). */
+  directives: DirectiveEntry[]
+  /** The current Now/Next/Later focus, or null when nothing is scheduled. */
+  focus: Focus | null
+}
+
+function focusBlock(focus: Focus | null): string {
+  if (!focus) return ''
+  const slot = (label: string, item: Focus['now']) => (item ? `${label}: #${item.number} ${item.title}` : `${label}: —`)
+  return ['## Focus', slot('Now', focus.now), slot('Next', focus.next), slot('Later', focus.later)].join('\n')
+}
+
+function directivesBlock(directives: DirectiveEntry[]): string[] {
+  // one compact line each; the verbatim quote is the payload, kind is the tag
+  return directives.map((d) => `- [${d.kind}] ${d.quote}`)
+}
+
+/** Truncate to a token budget on a whitespace boundary, with an ellipsis marker. */
+function clampToTokens(text: string, budgetTokens: number): string {
+  if (estimateTokens(text) <= budgetTokens) return text
+  const maxChars = Math.max(0, budgetTokens * 4 - 1)
+  const cut = text.slice(0, maxChars)
+  const lastBreak = cut.lastIndexOf('\n')
+  return `${(lastBreak > maxChars * 0.6 ? cut.slice(0, lastBreak) : cut).trimEnd()}\n…`
+}
+
+/**
+ * Assemble the HOT context under `budget` tokens. Priority when space is tight:
+ * the focus snapshot (tiny, always kept) → the most recent active directives
+ * (each while they fit) → the charter fills whatever budget remains (truncated).
+ * Never inlines ticket bodies — only charter text, directive quotes, and focus
+ * titles, all authored/short. Returns a single prompt-ready block.
+ */
+export function assembleHotContext(inputs: HotContextInputs, budget = HOT_CONTEXT_BUDGET_TOKENS): string {
+  const focus = focusBlock(inputs.focus)
+  const focusTokens = focus ? estimateTokens(focus) : 0
+
+  // fit the most recent active directives into ~⅔ of what's left after focus
+  const active = activeDirectives(inputs.directives)
+  const directiveCap = Math.max(0, Math.floor((budget - focusTokens) * (2 / 3)))
+  const keptDirectives: string[] = []
+  let directiveTokens = 0
+  for (const line of directivesBlock(active)) {
+    const t = estimateTokens(line) + 1
+    if (directiveTokens + t > directiveCap) break
+    keptDirectives.push(line)
+    directiveTokens += t
+  }
+  const directives = keptDirectives.length ? ['## Active directives', ...keptDirectives].join('\n') : ''
+
+  // Measure the fixed tail (directives + focus, with their joiner) exactly, then
+  // give the charter the true remainder — reserving for the "## Charter" header,
+  // the block joiner, and the truncation ellipsis so the total never exceeds budget.
+  const tail = [directives, focus].filter(Boolean).join('\n\n')
+  const tailTokens = tail ? estimateTokens(tail) : 0
+  const reserve = estimateTokens(`## Charter\n${tail ? '\n\n' : ''}\n…`)
+  const charterBudget = Math.max(0, budget - tailTokens - reserve)
+  const charterBody = inputs.charter.trim() ? clampToTokens(inputs.charter.trim(), charterBudget) : ''
+  const charter = charterBody ? `## Charter\n${charterBody}` : ''
+
+  return [charter, tail].filter(Boolean).join('\n\n')
+}

packages/core/src/index.ts

@@ -111,6 +111,8 @@ export {
 } from './agent/agent-tools.js'
 export { buildProjectView } from './agent/query-project.js'
 export type { ProjectSnapshot, ProjectView, QueryFilters } from './agent/query-project.js'
+export { activeDirectives, assembleHotContext, estimateTokens, HOT_CONTEXT_BUDGET_TOKENS } from './agent/memory-v0.js'
+export type { HotContextInputs } from './agent/memory-v0.js'
 export { CAPTURE_SYSTEM, captureWork, parseCaptureArgs, PROPOSE_ISSUES_TOOL } from './agent/capture-work.js'
 export type { CaptureProposal, ProposedIssue } from './agent/capture-work.js'
 

packages/core/src/perf/perf.test.ts

@@ -0,0 +1,75 @@
+/**
+ * Performance pass (#32). The deterministic compute path must stay well under the
+ * PLAN.md targets on representative fixtures:
+ *   - scheduler + Monte Carlo forecast < 1s @ 200 open issues.
+ *   - scaling stays roughly linear (no accidental O(n²) in the hot path).
+ *
+ * Reconcile-<5s@500 is network-bound (~2N gitea calls) and is covered by the live
+ * reconcile, not here — this file benchmarks the pure compute the app runs each
+ * turn. Bounds are the actual targets with comfortable headroom so timing jitter
+ * can't flake the suite; actuals are logged.
+ */
+import { describe, expect, it } from 'vitest'
+
+import { forecast } from '../forecast/forecast-v0.js'
+import { type DependencyEdge, schedule, type SchedulableIssue } from '../scheduler/scheduler-v0.js'
+import { scheduleWithCapacity, type Worker } from '../scheduler/scheduler-capacity-v0.js'
+
+const EST = [1, 2, 3, 5, 8]
+const WORKERS: Worker[] = [
+  { person: 'a', speed: 0.8 },
+  { person: 'b', speed: 0.6 },
+  { person: 'c', speed: 1.0 },
+]
+
+/** A representative open backlog: varied estimates/priorities/assignees + a light dependency web. */
+function backlog(n: number): { issues: SchedulableIssue[]; edges: DependencyEdge[] } {
+  const issues: SchedulableIssue[] = Array.from({ length: n }, (_, i) => ({
+    number: i + 1,
+    title: `Issue ${i + 1} with a representative title of some length`,
+    labels: [`est/${EST[i % EST.length]}d`, `p/${(i % 4) + 1}`],
+    estimateDays: EST[i % EST.length],
+    priority: (i % 4) + 1,
+    assignee: WORKERS[i % WORKERS.length].person,
+  }))
+  // ~1 dependency per 3 issues, always on a lower-numbered issue (acyclic)
+  const edges: DependencyEdge[] = []
+  for (let i = 3; i < n; i += 3) edges.push({ issue: i + 1, dependsOn: i - 1 })
+  return { issues, edges }
+}
+
+function ms(fn: () => void): number {
+  const t0 = performance.now()
+  fn()
+  return performance.now() - t0
+}
+
+describe('perf (#32)', () => {
+  it('scheduler + Monte Carlo forecast < 1s @ 200 open issues', () => {
+    const { issues, edges } = backlog(200)
+    const elapsed = ms(() => {
+      schedule(issues, edges)
+      scheduleWithCapacity(issues, edges, WORKERS)
+      forecast(issues, edges, { workers: WORKERS }) // 2000 trials (default)
+    })
+    // eslint-disable-next-line no-console
+    console.log(`[perf] schedule+capacity+forecast @200 = ${elapsed.toFixed(1)}ms`)
+    expect(elapsed).toBeLessThan(1000)
+  })
+
+  it('scales roughly linearly — 400 issues is well under 4x the 100-issue time', () => {
+    const small = backlog(100)
+    const big = backlog(400)
+    const run = (b: typeof small) => () => {
+      schedule(b.issues, b.edges)
+      forecast(b.issues, b.edges, { workers: WORKERS })
+    }
+    // warm up (JIT) so the ratio reflects steady state
+    run(small)()
+    const t100 = Math.max(ms(run(small)), 0.1)
+    const t400 = ms(run(big))
+    // eslint-disable-next-line no-console
+    console.log(`[perf] @100 = ${t100.toFixed(1)}ms · @400 = ${t400.toFixed(1)}ms · ratio ${(t400 / t100).toFixed(1)}x`)
+    expect(t400).toBeLessThan(t100 * 8) // generous: rules out O(n²), tolerant of jitter
+  })
+})