Merge pull request 'Performance pass: benchmark the deterministic compute path (#32)' (#59) from feat/perf-pass into main

Reviewed-on: #59

apps/desktop/src/renderer/src/components/screens/calibration-screen.tsx

@@ -178,6 +178,9 @@ export function CalibrationScreen({ onBack, data }: { onBack: () => void; data?:
               {c.active
                 ? 'You are not bad at estimating; you are optimistic in a very stable way. Stable, I can work with.'
                 : 'Not enough closed history yet — I’m forecasting from cold-start priors and widening the cone to stay honest. The curve takes over at 20.'}
+              {!c.active && c.excludedSameDay > 0
+                ? ` And ${c.excludedSameDay} closed ${c.excludedSameDay === 1 ? 'issue' : 'issues'} closed the same day they were started — 0 working days can’t calibrate, so they don’t count toward the 20.`
+                : ''}
             </p>
           </Card>
         </div>

apps/desktop/src/renderer/src/components/screens/runway-screen.tsx

@@ -17,7 +17,7 @@ export function RunwayScreen({
 }: {
   onOpenCalibration: () => void
   onOpenMilestone: (id?: number) => void
-  calibration?: { n: number; coldStart: boolean }
+  calibration?: { n: number; coldStart: boolean; excludedSameDay?: number }
   milestones?: RunwayMilestone[]
   capacity?: CapacityMember[]
 }) {
@@ -30,9 +30,11 @@ export function RunwayScreen({
           hours: `${capacityPerWorkday(m).toFixed(2)} pd/day`,
         }))
       : CAPACITY
+  const excluded = calibration?.excludedSameDay ?? 0
   const calibNote = calibration
     ? calibration.coldStart
-      ? `cold-start priors · ${calibration.n}/20 closed issues estimated`
+      ? `cold-start priors · ${calibration.n}/20 closed issues estimated` +
+        (excluded > 0 ? ` · ${excluded} same-day close${excluded === 1 ? '' : 's'} can’t calibrate` : '')
       : `calibrated on ${calibration.n} closed ${calibration.n === 1 ? 'issue' : 'issues'}`
     : 'calibrated on 27 closed issues'
   return (

apps/desktop/src/renderer/src/components/shell/app-shell.tsx

@@ -306,7 +306,15 @@ export function AppShell() {
               setMilestoneId(id ?? null)
               setView('milestone')
             }}
-            calibration={calibration ? { n: calibration.model.n, coldStart: calibration.model.coldStart } : undefined}
+            calibration={
+              calibration
+                ? {
+                    n: calibration.model.n,
+                    coldStart: calibration.model.coldStart,
+                    excludedSameDay: calibration.coverage.excludedSameDay,
+                  }
+                : undefined
+            }
             milestones={runwayMilestones}
             capacity={capacityMembers}
           />

apps/desktop/src/renderer/src/data/fixtures.ts

@@ -361,11 +361,14 @@ export interface CalibrationData {
   scatter: number[][]
   fit: number
   effect: { raw: string; banded: string; p50: string }
+  /** Closed+estimated issues that can't calibrate (same-day / 0-day closes). */
+  excludedSameDay: number
 }
 
 export const CALIBRATION: CalibrationData = {
   n: 27,
   active: true,
+  excludedSameDay: 0,
   labels: [
     { label: 'est/1d', n: 8, median: '1.1d', bias: 8 },
     { label: 'est/2d', n: 9, median: '2.4d', bias: 18 },

apps/desktop/src/renderer/src/lib/backlog.ts

@@ -1,6 +1,8 @@
 import {
+  type CalibrationCoverage,
   type CalibrationModel,
   type CalibrationSample,
+  calibrationCoverage,
   calibrationSamples,
   type CapacityMember,
   capacityPerWorkday,
@@ -171,10 +173,11 @@ export function backlogCalibration(
   issues: GiteaIssue[],
   timelines: Timelines = {},
   asOf: Date = new Date(),
-): { model: CalibrationModel; data: CalibrationData } {
+): { model: CalibrationModel; data: CalibrationData; coverage: CalibrationCoverage } {
   const samples = calibrationSamples(issues, timelines, asOf)
   const model = fitCalibration(samples)
-  return { model, data: calibrationData(model, samples, issues) }
+  const coverage = calibrationCoverage(issues, timelines, asOf)
+  return { model, coverage, data: calibrationData(model, samples, issues, coverage.excludedSameDay) }
 }
 
 const pctFromMu = (mu: number) => Math.round((Math.exp(mu) - 1) * 100)
@@ -188,6 +191,7 @@ export function calibrationData(
   model: CalibrationModel,
   samples: CalibrationSample[],
   openIssues: GiteaIssue[],
+  excludedSameDay = 0,
 ): CalibrationData {
   const labels = PRIOR_BUCKETS.map((b) => {
     const inBucket = samples.filter((s) => s.bucket === b)
@@ -227,6 +231,7 @@ export function calibrationData(
     scatter: samples.map((s) => [s.estimateDays, s.actualWorkingDays]),
     fit: Number(Math.exp(model.global.mu).toFixed(2)),
     effect,
+    excludedSameDay,
   }
 }
 

packages/core/src/agent/agent-tools.ts

@@ -17,10 +17,11 @@ export const QUERY_PROJECT_TOOL: ToolDecl = {
     properties: {
       view: {
         type: 'string',
-        enum: ['focus', 'board', 'calibration', 'issue', 'search'],
+        enum: ['focus', 'board', 'calibration', 'issue', 'search', 'standup'],
         description:
           'focus = Now/Next/Later; board = issues by lifecycle column; calibration = estimate-vs-actual; ' +
-          'issue = one issue (needs filters.issueId); search = issues matching filters.query.',
+          'issue = one issue (needs filters.issueId); search = issues matching filters.query; ' +
+          "standup = today's plan + overnight drift + the stalest blocker.",
       },
       filters: {
         type: 'object',

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/agent/query-project.test.ts

@@ -0,0 +1,74 @@
+import { describe, expect, it } from 'vitest'
+
+import { extractLabelFacts } from '../labels/label-schema.js'
+import type { LifecycleEvent } from '../lifecycle/lifecycle-v0.js'
+import type { GiteaIssue } from '../gitea/types.js'
+import { buildProjectView, type ProjectSnapshot } from './query-project.js'
+
+const asOf = new Date('2026-02-12T00:00:00Z')
+
+function issue(over: Partial<GiteaIssue>): GiteaIssue {
+  const labels = over.labels ?? ['est/2d']
+  return {
+    number: 1,
+    title: '#1',
+    body: '',
+    state: 'open',
+    labels,
+    facts: extractLabelFacts(labels),
+    milestone: null,
+    assignee: null,
+    assignees: [],
+    createdAt: '2026-02-02T09:00:00Z',
+    updatedAt: '2026-02-02T09:00:00Z',
+    closedAt: null,
+    url: '',
+    ...over,
+  }
+}
+
+describe('buildProjectView: standup (#28)', () => {
+  it('surfaces schedule drift + a stale blocker, with a per-person plan', () => {
+    // #7 started work 6 working days ago (steeping) and blocks #8; est is 2d → past estimate.
+    const steeping = issue({ number: 7, labels: ['est/2d', 'p/1'], assignee: 'christian' })
+    const blocked = issue({ number: 8, labels: ['est/3d', 'p/2'], assignee: 'stephen' })
+    const timelines: Record<number, LifecycleEvent[]> = {
+      7: [{ type: 'commit', at: '2026-02-04T09:00:00Z' }],
+    }
+    const snap: ProjectSnapshot = { issues: [steeping, blocked], timelines, deps: [{ issue: 8, dependsOn: 7 }] }
+
+    const v = buildProjectView('standup', undefined, snap, asOf) as {
+      plan: { who: string; issue: number; why: string }[]
+      drift: { issue: number; note: string }[]
+      nag: { issue: number; steepingDays: number; blocks: number[] } | null
+    }
+
+    // a stale blocker is nagged, and it's the steeping one that blocks another
+    expect(v.nag).not.toBeNull()
+    expect(v.nag!.issue).toBe(7)
+    expect(v.nag!.steepingDays).toBeGreaterThan(0)
+    expect(v.nag!.blocks).toContain(8)
+
+    // drift caught the past-estimate steep
+    expect(v.drift.some((d) => d.issue === 7)).toBe(true)
+
+    // plan gives an earliest pick per person (both assignees represented)
+    expect(v.plan.map((p) => p.who)).toEqual(expect.arrayContaining(['christian', 'stephen']))
+    // #7 leads its person's plan on the critical path
+    expect(v.plan.find((p) => p.issue === 7)?.why).toContain('critical')
+  })
+
+  it('is calm when nothing is steeping (no nag, empty drift)', () => {
+    const snap: ProjectSnapshot = {
+      issues: [issue({ number: 1, labels: ['est/2d'] })],
+      timelines: {},
+      deps: [],
+    }
+    const v = buildProjectView('standup', undefined, snap, asOf) as {
+      drift: unknown[]
+      nag: unknown | null
+    }
+    expect(v.nag).toBeNull()
+    expect(v.drift).toEqual([])
+  })
+})

packages/core/src/agent/query-project.ts

@@ -4,9 +4,9 @@
  * deterministic code (scheduler, lifecycle inference, calibration); the model
  * only requests a shape and narrates it — it never computes (decisions.md).
  *
- * v0 serves focus / board / calibration / issue. The remaining views
+ * Serves focus / board / calibration / issue / search / standup. The remaining
- * (milestone / runway / standup / search) return a `notImplemented` marker so
+ * views (milestone / runway) return a `notImplemented` marker so the model
- * the model degrades honestly instead of inventing data.
+ * degrades honestly instead of inventing data.
  */
 
 import { fitCalibration, calibrationSamples } from '../calibration/calibration-v0.js'
@@ -121,6 +121,62 @@ function searchView(snap: ProjectSnapshot, filters: QueryFilters) {
   }
 }
 
+/**
+ * Standup — the morning ritual as a compact, model-narratable payload: today's
+ * plan (the scheduler's earliest pick per person), overnight drift (real
+ * anomalies — issues sitting in review or steeping past their estimate), and the
+ * single stalest blocker to nag about. All from deterministic code; the model
+ * narrates it. Satisfies #28's "surfaces schedule drift + at least one stale blocker".
+ */
+function standupView(snap: ProjectSnapshot, asOf: Date) {
+  const plan = schedule(toSchedulable(snap.issues), snap.deps)
+  const open = snap.issues.filter((i) => i.state === 'open')
+  const infOf = (i: GiteaIssue) => inferLifecycle(i, snap.timelines[i.number] ?? [], asOf)
+
+  // plan: the earliest scheduled pick per assignee (dependency + priority order)
+  const seen = new Set<string>()
+  const planPicks: { who: string; issue: number; title: string; why: string }[] = []
+  for (const item of plan.items) {
+    const who = open.find((i) => i.number === item.number)?.assignee ?? 'unassigned'
+    if (seen.has(who)) continue
+    seen.add(who)
+    planPicks.push({
+      who,
+      issue: item.number,
+      title: item.title,
+      why: item.critical
+        ? 'on the critical path'
+        : item.blocks.length
+          ? `blocks ${item.blocks.map((b) => `#${b}`).join(', ')}`
+          : 'next by dependency + priority',
+    })
+  }
+
+  // drift: real overnight anomalies (review-sitting, steeping past estimate)
+  const drift: { issue: number; note: string }[] = []
+  for (const i of open) {
+    const inf = infOf(i)
+    if (inf.column === 'review') drift.push({ issue: i.number, note: `#${i.number} is sitting in review` })
+    else if (inf.steepingDays != null && inf.steepingDays > (i.facts.estimateDays ?? 2))
+      drift.push({
+        issue: i.number,
+        note: `#${i.number} has steeped ${inf.steepingDays}d past its ${i.facts.estimateDays ?? 2}d estimate`,
+      })
+  }
+
+  // nag: the single longest-steeping open issue + what it blocks
+  let nag: { issue: number; steepingDays: number; blocks: number[] } | null = null
+  for (const i of open) {
+    const inf = infOf(i)
+    if (inf.steepingDays == null) continue
+    if (!nag || inf.steepingDays > nag.steepingDays) {
+      nag = { issue: i.number, steepingDays: inf.steepingDays, blocks: plan.items.find((it) => it.number === i.number)?.blocks ?? [] }
+    }
+  }
+
+  return { date: asOf.toISOString().slice(0, 10), plan: planPicks.slice(0, 5), drift: drift.slice(0, 5), nag }
+}
+
 /** Build the compact payload for one view. Unknown/unbuilt views return a marker. */
 export function buildProjectView(
   view: ProjectView,
@@ -140,6 +196,8 @@ export function buildProjectView(
       return issueView(snap, f, asOf)
     case 'search':
       return searchView(snap, f)
+    case 'standup':
+      return standupView(snap, asOf)
     default:
       return { notImplemented: view }
   }

packages/core/src/cache/cache-v0.test.ts

@@ -0,0 +1,92 @@
+import { DatabaseSync } from 'node:sqlite'
+
+import { describe, expect, it } from 'vitest'
+
+import type { GiteaIssue } from '../gitea/types.js'
+import { extractLabelFacts } from '../labels/label-schema.js'
+import { type CacheDriver, initCache, readIssue, upsertIssue } from './cache-v0.js'
+
+/** Adapt node:sqlite's DatabaseSync to the CacheDriver seam (main uses better-sqlite3). */
+function memoryDriver(): CacheDriver {
+  const db = new DatabaseSync(':memory:')
+  return {
+    exec: (sql) => db.exec(sql),
+    run: (sql, params = []) => {
+      db.prepare(sql).run(...(params as never[]))
+    },
+    get: (sql, params = []) => db.prepare(sql).get(...(params as never[])) as Record<string, unknown> | undefined,
+    all: (sql, params = []) => db.prepare(sql).all(...(params as never[])) as Record<string, unknown>[],
+  }
+}
+
+function issue(over: Partial<GiteaIssue> = {}): GiteaIssue {
+  const labels = over.labels ?? ['est/5d', 'p/1', 'deadline/hard']
+  return {
+    number: 42,
+    title: 'Monte Carlo engine',
+    body: 'percentile bands',
+    state: 'open',
+    labels,
+    facts: extractLabelFacts(labels),
+    milestone: { id: 7, title: 'P2 — Scheduler', dueOn: '2026-09-01T00:00:00Z' },
+    assignee: 'christian',
+    assignees: ['christian'],
+    createdAt: '2026-07-08T00:00:00Z',
+    updatedAt: '2026-07-08T01:00:00Z',
+    closedAt: null,
+    url: 'https://gitea/christian/commitea/issues/42',
+    ...over,
+  }
+}
+
+describe('cache-v0', () => {
+  it('mirrors one issue and reads its facts back through extractLabelFacts (acceptance)', () => {
+    const d = memoryDriver()
+    initCache(d)
+    upsertIssue(d, issue({ labels: ['est/5d', 'p/1', 'deadline/hard'] }))
+
+    const back = readIssue(d, 42)!
+    expect(back.labels).toEqual(['est/5d', 'p/1', 'deadline/hard'])
+    // facts are re-derived on read, not stored
+    expect(back.facts.estimateDays).toBe(5)
+    expect(back.facts.priority).toBe(1)
+    expect(back.facts.hardDeadline).toBe(true)
+    // the rest of the domain shape round-trips
+    expect(back.milestone).toEqual({ id: 7, title: 'P2 — Scheduler', dueOn: '2026-09-01T00:00:00Z' })
+    expect(back.assignee).toBe('christian')
+    expect(back.state).toBe('open')
+  })
+
+  it('re-derives facts from the current labels after a re-reconcile (upsert in place, no dup)', () => {
+    const d = memoryDriver()
+    initCache(d)
+    upsertIssue(d, issue({ labels: ['est/2d', 'p/3'] }))
+    // reconcile again with changed labels + closed
+    upsertIssue(d, issue({ labels: ['est/8d', 'p/1'], state: 'closed', closedAt: '2026-07-09T00:00:00Z' }))
+
+    expect(d.all('SELECT number FROM issues')).toHaveLength(1) // upsert by number, not a second row
+    const back = readIssue(d, 42)!
+    expect(back.facts.estimateDays).toBe(8)
+    expect(back.facts.priority).toBe(1)
+    expect(back.facts.hardDeadline).toBe(false) // deadline/hard dropped
+    expect(back.state).toBe('closed')
+    expect(back.closedAt).toBe('2026-07-09T00:00:00Z')
+  })
+
+  it('reads an issue with no milestone / empty labels', () => {
+    const d = memoryDriver()
+    initCache(d)
+    upsertIssue(d, issue({ number: 9, labels: [], milestone: null, assignee: null, assignees: [] }))
+    const back = readIssue(d, 9)!
+    expect(back.milestone).toBeNull()
+    expect(back.labels).toEqual([])
+    expect(back.facts.estimateDays).toBeNull()
+    expect(back.assignee).toBeNull()
+  })
+
+  it('returns null for an uncached issue', () => {
+    const d = memoryDriver()
+    initCache(d)
+    expect(readIssue(d, 999)).toBeNull()
+  })
+})

packages/core/src/cache/cache-v0.ts

@@ -0,0 +1,159 @@
+/**
+ * SQLite cache, v0 (#3) — a rebuildable local mirror of the reconciled backlog.
+ * It is an index over the durable truth in gitea, never the source of truth (D4):
+ * delete it, resync, lose nothing. This module owns the schema + the pure
+ * row<->domain mappers; the actual SQLite handle is injected as a `CacheDriver`,
+ * so core stays free of any native driver (better-sqlite3 lives in main; tests
+ * use node:sqlite). Facts are never stored — they are re-derived from the label
+ * set on read via `extractLabelFacts`, so the mirror can't drift from the label
+ * semantics.
+ */
+
+import type { GiteaIssue, GiteaMilestoneRef } from '../gitea/types.js'
+import { extractLabelFacts } from '../labels/label-schema.js'
+
+/**
+ * The injected IO boundary: a thin synchronous SQL executor. Core writes the SQL;
+ * the host binds a real driver (better-sqlite3 in the desktop main process,
+ * node:sqlite in tests). Kept minimal on purpose — no ORM, no query builder.
+ */
+export interface CacheDriver {
+  /** Run one or more DDL/utility statements (no params, no result). */
+  exec(sql: string): void
+  /** Execute a single parameterized write. */
+  run(sql: string, params?: readonly unknown[]): void
+  /** First row of a parameterized query, or undefined. */
+  get(sql: string, params?: readonly unknown[]): Record<string, unknown> | undefined
+  /** All rows of a parameterized query. */
+  all(sql: string, params?: readonly unknown[]): Record<string, unknown>[]
+}
+
+/** The cache schema — five tables mirroring gitea's shape. Regenerable; drop and rebuild freely. */
+export const CACHE_SCHEMA = `
+CREATE TABLE IF NOT EXISTS milestones (
+  id       INTEGER PRIMARY KEY,
+  title    TEXT NOT NULL,
+  state    TEXT,
+  due_on   TEXT
+);
+CREATE TABLE IF NOT EXISTS issues (
+  number       INTEGER PRIMARY KEY,
+  title        TEXT NOT NULL,
+  body         TEXT NOT NULL DEFAULT '',
+  state        TEXT NOT NULL,
+  labels       TEXT NOT NULL DEFAULT '[]',  -- JSON array of label names; facts re-derived on read
+  milestone_id INTEGER,
+  assignee     TEXT,
+  assignees    TEXT NOT NULL DEFAULT '[]',  -- JSON array of logins
+  created_at   TEXT,
+  updated_at   TEXT,
+  closed_at    TEXT,
+  url          TEXT,
+  FOREIGN KEY (milestone_id) REFERENCES milestones(id)
+);
+CREATE TABLE IF NOT EXISTS labels (
+  id    INTEGER PRIMARY KEY,
+  name  TEXT NOT NULL
+);
+CREATE TABLE IF NOT EXISTS comments (
+  id            INTEGER PRIMARY KEY,
+  issue_number  INTEGER NOT NULL,
+  author        TEXT,
+  body          TEXT NOT NULL DEFAULT '',
+  created_at    TEXT
+);
+CREATE TABLE IF NOT EXISTS issue_events (
+  id            INTEGER PRIMARY KEY AUTOINCREMENT,
+  issue_number  INTEGER NOT NULL,
+  type          TEXT NOT NULL,
+  at            TEXT NOT NULL
+);
+CREATE INDEX IF NOT EXISTS idx_issue_events_number ON issue_events(issue_number);
+CREATE INDEX IF NOT EXISTS idx_comments_number ON comments(issue_number);
+`
+
+/** Create the schema if absent. Idempotent. */
+export function initCache(driver: CacheDriver): void {
+  driver.exec(CACHE_SCHEMA)
+}
+
+const UPSERT_ISSUE = `
+INSERT INTO issues (number, title, body, state, labels, milestone_id, assignee, assignees, created_at, updated_at, closed_at, url)
+VALUES (?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?, ?)
+ON CONFLICT(number) DO UPDATE SET
+  title = excluded.title, body = excluded.body, state = excluded.state, labels = excluded.labels,
+  milestone_id = excluded.milestone_id, assignee = excluded.assignee, assignees = excluded.assignees,
+  created_at = excluded.created_at, updated_at = excluded.updated_at, closed_at = excluded.closed_at, url = excluded.url
+`
+
+const UPSERT_MILESTONE = `
+INSERT INTO milestones (id, title, state, due_on) VALUES (?, ?, ?, ?)
+ON CONFLICT(id) DO UPDATE SET title = excluded.title, state = excluded.state, due_on = excluded.due_on
+`
+
+/**
+ * Mirror one reconciled issue into the cache (and its milestone, if any). Upsert
+ * by `number`, so re-reconciling the same issue updates in place — never duplicates.
+ */
+export function upsertIssue(driver: CacheDriver, issue: GiteaIssue): void {
+  if (issue.milestone) {
+    driver.run(UPSERT_MILESTONE, [issue.milestone.id, issue.milestone.title, null, issue.milestone.dueOn])
+  }
+  driver.run(UPSERT_ISSUE, [
+    issue.number,
+    issue.title,
+    issue.body,
+    issue.state,
+    JSON.stringify(issue.labels),
+    issue.milestone?.id ?? null,
+    issue.assignee,
+    JSON.stringify(issue.assignees),
+    issue.createdAt,
+    issue.updatedAt,
+    issue.closedAt,
+    issue.url,
+  ])
+}
+
+const READ_ISSUE = `
+SELECT i.*, m.title AS m_title, m.due_on AS m_due
+FROM issues i LEFT JOIN milestones m ON m.id = i.milestone_id
+WHERE i.number = ?
+`
+
+function str(v: unknown): string {
+  return typeof v === 'string' ? v : ''
+}
+
+/**
+ * Read one mirrored issue back as a domain object, re-deriving `facts` from the
+ * stored label set (so the mirror can't disagree with the label semantics).
+ * Returns null when the issue isn't cached.
+ */
+export function readIssue(driver: CacheDriver, number: number): GiteaIssue | null {
+  const row = driver.get(READ_ISSUE, [number])
+  if (!row) return null
+
+  const labels = (JSON.parse(str(row.labels) || '[]') as string[]) ?? []
+  const assignees = (JSON.parse(str(row.assignees) || '[]') as string[]) ?? []
+  const milestone: GiteaMilestoneRef | null =
+    row.milestone_id != null
+      ? { id: Number(row.milestone_id), title: str(row.m_title), dueOn: (row.m_due as string | null) ?? null }
+      : null
+
+  return {
+    number: Number(row.number),
+    title: str(row.title),
+    body: str(row.body),
+    state: row.state === 'closed' ? 'closed' : 'open',
+    labels,
+    facts: extractLabelFacts(labels),
+    milestone,
+    assignee: (row.assignee as string | null) ?? null,
+    assignees,
+    createdAt: str(row.created_at),
+    updatedAt: str(row.updated_at),
+    closedAt: (row.closed_at as string | null) ?? null,
+    url: str(row.url),
+  }
+}

packages/core/src/calibration/calibration-v0.test.ts

@@ -5,6 +5,7 @@ import type { LifecycleEvent } from '../lifecycle/lifecycle-v0.js'
 import type { GiteaIssue } from '../gitea/types.js'
 import {
   CALIBRATION_BUCKET_FLOOR,
+  calibrationCoverage,
   calibrationSamples,
   type CalibrationSample,
   COLD_START_THRESHOLD,
@@ -118,4 +119,26 @@ describe('calibrationSamples', () => {
     const noEst = issue({ number: 9, labels: [] })
     expect(calibrationSamples([open, noEst], { ...events(8), ...events(9) }, asOf)).toEqual([])
   })
+
+  it('coverage counts same-day closes as excluded candidates, not as "more closes needed"', () => {
+    // usable: commit Wed 01-07 → close Mon 01-12 = 3 working days
+    const usable = issue({ number: 7, labels: ['est/2d'] })
+    // same-day close: commit and close on the same day = 0 working days → excluded
+    const sameDay = issue({ number: 10, labels: ['est/2d'], createdAt: '2026-01-12T08:00:00Z' })
+    const sameDayEvents = {
+      10: [
+        { type: 'commit', at: '2026-01-12T09:00:00Z' } as LifecycleEvent,
+        { type: 'close', at: '2026-01-12T17:00:00Z' } as LifecycleEvent,
+      ],
+    }
+    const open = issue({ number: 8, state: 'open', labels: ['est/2d'], closedAt: null })
+    const noEst = issue({ number: 9, labels: [] })
+
+    const cov = calibrationCoverage([usable, sameDay, open, noEst], { ...events(7), ...sameDayEvents }, asOf)
+    expect(cov.candidates).toBe(2) // closed + estimated only (usable + sameDay)
+    expect(cov.usable).toBe(1)
+    expect(cov.excludedSameDay).toBe(1)
+    // the honest denominator: usable matches the model's n
+    expect(cov.usable).toBe(calibrationSamples([usable, sameDay, open, noEst], { ...events(7), ...sameDayEvents }, asOf).length)
+  })
 })

packages/core/src/calibration/calibration-v0.ts

@@ -125,3 +125,40 @@ export function calibrationSamples(
   }
   return out
 }
+
+/** How the closed+estimated backlog splits into usable samples vs. what can't calibrate. */
+export interface CalibrationCoverage {
+  /** Closed issues carrying an estimate — the calibration candidates. */
+  candidates: number
+  /** Candidates that yielded a usable actual (> 0 working days) → become samples. */
+  usable: number
+  /**
+   * Candidates excluded because the issue closed with 0 working days (same-day
+   * close) or no resolvable actual — real closes that structurally can't
+   * calibrate. Counting them keeps `usable/threshold` honest: it's not "N more
+   * closes away" if some of your closes will never count.
+   */
+  excludedSameDay: number
+}
+
+/**
+ * Coverage of the calibration candidates — how many closed+estimated issues are
+ * usable vs. silently unusable (same-day / 0-day closes). {@link calibrationSamples}
+ * drops the latter; this counts them so the UI can say *why* the sample is thin.
+ */
+export function calibrationCoverage(
+  issues: GiteaIssue[],
+  timelines: Record<number, LifecycleEvent[]>,
+  asOf: Date,
+): CalibrationCoverage {
+  let candidates = 0
+  let usable = 0
+  for (const issue of issues) {
+    if (issue.state !== 'closed') continue
+    if (issue.facts.estimateDays == null) continue
+    candidates++
+    const inf = inferLifecycle(issue, timelines[issue.number] ?? [], asOf)
+    if (inf.actualWorkingDays != null && inf.actualWorkingDays > 0) usable++
+  }
+  return { candidates, usable, excludedSameDay: candidates - usable }
+}

packages/core/src/index.ts

@@ -23,6 +23,8 @@ export type {
   GiteaRequestInit,
 } from './gitea/types.js'
 
+export { CACHE_SCHEMA, initCache, readIssue, upsertIssue } from './cache/cache-v0.js'
+export type { CacheDriver } from './cache/cache-v0.js'
 export { describeChange, isLabelChange, planIssueChange, proposalsFor, summarizeChange } from './changes/apply-changes-v0.js'
 export type {
   ChangeProposal,
@@ -80,6 +82,7 @@ export type {
 
 export {
   CALIBRATION_BUCKET_FLOOR,
+  calibrationCoverage,
   calibrationSamples,
   COLD_START_THRESHOLD,
   fitCalibration,
@@ -87,6 +90,7 @@ export {
 } from './calibration/calibration-v0.js'
 export type {
   BucketFit,
+  CalibrationCoverage,
   CalibrationModel,
   CalibrationSample,
   PersonBias,
@@ -107,6 +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
+  })
+})