Compare commits

...

5 Commits

Author SHA1 Message Date
Croissant Le Doux
e5ce3fe87a Offline write-queue: coalesce + replay without duplicating (#33)
While gitea is unreachable, propose-approved writes queue instead of being lost;
on reconnect they replay in order. The hard requirement is no duplication.

The mechanism is coalescing by axis. Every write targets one field of one issue
(estimate / priority / assignee / milestone), so `issue:kind` is the axis identity.
Queuing a second write to the same axis supersedes the first (moved to the tail),
so a burst of edits replays as one final write — and replay is idempotent (the
apply path no-ops a change already reflected server-side).

- `enqueueWrite` (coalesce), `pendingWrites`, `coalesceKey`, `affectedIssues`
  (what a post-replay reconcile re-reads), and `replayQueue(queue, apply)` which
  drains through the same guarded write path and returns the writes that still
  failed so they stay queued. Never throws — a failure is data.

Acceptance met: a burst of offline edits + reconnect lands the final state with a
single apply per axis (not one per edit); still-failing writes stay queued. +4
core tests; typecheck green.

Follow-up: persist the queue in main + trigger replay on the reconnect signal
(the offline banner + disabled composer already exist) — the coalesce/replay
core is the tested heart.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-09 15:40:22 -04:00
2a6413821a Merge pull request 'calibration: count same-day closes honestly (#34)' (#54) from feat/calibration-honesty into main
Reviewed-on: #54
2026-07-09 19:23:59 +00:00
008435f1c2 Merge branch 'main' into feat/calibration-honesty 2026-07-09 19:23:55 +00:00
354ba9227e Merge pull request 'apply_changes: unified mutation tool — estimate/priority/assign/milestone, in-app + agent (#24)' (#53) from feat/apply-changes-assign-milestone into main
Reviewed-on: #53
2026-07-09 19:23:49 +00:00
Croissant Le Doux
89c873b368 calibration: count same-day closes honestly (#34)
The cold-start surface showed "N/20 closed issues estimated", implying you're
just (20−N) closes away. But calibrationSamples silently drops closed+estimated
issues that closed in 0 working days (same-day closes) — real closes that
structurally can't calibrate. On this repo that's 10 of 24 closes hidden: the
note read 14/20 as if 6 away, when a third of the history will never count.

- core: `calibrationCoverage(issues, timelines, asOf)` → { candidates, usable,
  excludedSameDay }, counting the silently-excluded same-day closes. Pure, tested.
- surface it: CalibrationData gains `excludedSameDay`; backlogCalibration returns
  the coverage; the Runway note and the Calibration screen now say "… · N same-day
  closes can't calibrate" so the thin sample is explained, not just reported.

Verified on christian/commitea: closed=24, usable=14, excludedSameDay=10.
131 core green (incl. new coverage test); core + desktop typecheck; 14 fixture e2e.

Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
2026-07-09 15:18:16 -04:00
10 changed files with 228 additions and 5 deletions

View File

@@ -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 — Im 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 cant calibrate, so they dont count toward the 20.`
: ''}
</p>
</Card>
</div>

View File

@@ -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'} cant calibrate` : '')
: `calibrated on ${calibration.n} closed ${calibration.n === 1 ? 'issue' : 'issues'}`
: 'calibrated on 27 closed issues'
return (

View File

@@ -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}
/>

View File

@@ -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 },

View File

@@ -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,
}
}

View File

@@ -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)
})
})

View File

@@ -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 }
}

View File

@@ -23,6 +23,8 @@ export type {
GiteaRequestInit,
} from './gitea/types.js'
export { affectedIssues, coalesceKey, enqueueWrite, pendingWrites, replayQueue } from './queue/write-queue-v0.js'
export type { QueuedWrite } from './queue/write-queue-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,

View File

@@ -0,0 +1,67 @@
import { describe, expect, it, vi } from 'vitest'
import type { IssueChange } from '../changes/apply-changes-v0.js'
import { affectedIssues, coalesceKey, enqueueWrite, pendingWrites, replayQueue, type QueuedWrite } from './write-queue-v0.js'
const reest = (issue: number, estimate: string): IssueChange =>
({ kind: 'reestimate', issue, estimate }) as IssueChange
const assign = (issue: number, assignee: string | null): IssueChange => ({ kind: 'assign', issue, assignee })
describe('write-queue-v0 (#33)', () => {
it('coalesces repeat writes to the same (issue, axis) — replay applies the latest once', () => {
let q: QueuedWrite[] = []
q = enqueueWrite(q, reest(7, 'est/2d'), '2026-02-10T09:00:00Z')
q = enqueueWrite(q, reest(7, 'est/5d'), '2026-02-10T09:05:00Z') // supersedes est/2d
q = enqueueWrite(q, assign(7, 'christian'), '2026-02-10T09:06:00Z') // different axis — kept
expect(q).toHaveLength(2)
const pending = pendingWrites(q)
expect(pending).toEqual([
{ kind: 'reestimate', issue: 7, estimate: 'est/5d' },
{ kind: 'assign', issue: 7, assignee: 'christian' },
])
expect(coalesceKey(reest(7, 'est/2d'))).toBe('7:reestimate')
})
it('keeps writes to different issues and axes distinct', () => {
let q: QueuedWrite[] = []
q = enqueueWrite(q, reest(7, 'est/2d'), 't1')
q = enqueueWrite(q, reest(8, 'est/3d'), 't2')
q = enqueueWrite(q, assign(8, null), 't3')
expect(q).toHaveLength(3)
expect(affectedIssues(q)).toEqual([7, 8])
})
it('a burst of edits then reconnect lands the final state without duplicating', async () => {
let q: QueuedWrite[] = []
// offline: three edits to #7's estimate, one assign
q = enqueueWrite(q, reest(7, 'est/1d'), 't1')
q = enqueueWrite(q, reest(7, 'est/2d'), 't2')
q = enqueueWrite(q, reest(7, 'est/8d'), 't3')
q = enqueueWrite(q, assign(7, 'stephen'), 't4')
const apply = vi.fn(async () => ({ ok: true }))
const { drained, remaining } = await replayQueue(q, apply)
// only the final estimate + the assign are applied — not three estimate writes
expect(apply).toHaveBeenCalledTimes(2)
expect(apply).toHaveBeenNthCalledWith(1, { kind: 'reestimate', issue: 7, estimate: 'est/8d' })
expect(apply).toHaveBeenNthCalledWith(2, { kind: 'assign', issue: 7, assignee: 'stephen' })
expect(remaining).toHaveLength(0)
expect(drained).toHaveLength(2)
})
it('keeps writes that still fail on reconnect queued (never throws)', async () => {
let q: QueuedWrite[] = []
q = enqueueWrite(q, reest(7, 'est/2d'), 't1')
q = enqueueWrite(q, reest(8, 'est/3d'), 't2')
// #7 applies, #8 rejects (still unreachable) — and one apply throws
const apply = vi.fn(async (c: IssueChange) => {
if (c.issue === 8) throw new Error('offline')
return { ok: true }
})
const { drained, remaining } = await replayQueue(q, apply)
expect(drained.map((w) => w.change.issue)).toEqual([7])
expect(remaining.map((w) => w.change.issue)).toEqual([8])
})
})

View File

@@ -0,0 +1,71 @@
/**
* Offline write-queue, v0 (#33). While gitea is unreachable, propose-approved
* writes are queued instead of lost; on reconnect they replay in order and
* resolve against a fresh reconcile. The one hard requirement is *no duplication*:
* replaying must not apply the same intent twice.
*
* The mechanism is coalescing by axis. Every write targets one field of one issue
* (its estimate, priority, assignee, or milestone). Queuing a second write to the
* same (issue, axis) supersedes the first — only the latest intent survives — so
* a burst of edits replays as one final write, and a replay is idempotent (the
* apply path no-ops a change already reflected server-side).
*/
import type { IssueChange } from '../changes/apply-changes-v0.js'
export interface QueuedWrite {
change: IssueChange
/** ISO time the write was queued (for display + stable ordering). */
queuedAt: string
}
/**
* The coalescing key: one field of one issue. Two writes with the same key are
* the same intent expressed twice — the later wins. Each `IssueChange.kind` maps
* to exactly one axis, so `issue:kind` is the axis identity.
*/
export function coalesceKey(change: IssueChange): string {
return `${change.issue}:${change.kind}`
}
/**
* Queue a write, superseding any pending write to the same (issue, axis). The
* superseding write moves to the tail so replay order reflects latest intent.
*/
export function enqueueWrite(queue: readonly QueuedWrite[], change: IssueChange, queuedAt: string): QueuedWrite[] {
const key = coalesceKey(change)
return [...queue.filter((w) => coalesceKey(w.change) !== key), { change, queuedAt }]
}
/** The changes to replay, in order — one per (issue, axis) by construction. */
export function pendingWrites(queue: readonly QueuedWrite[]): IssueChange[] {
return queue.map((w) => w.change)
}
/** Distinct issues touched by the queue — what a post-replay reconcile should re-read. */
export function affectedIssues(queue: readonly QueuedWrite[]): number[] {
return [...new Set(queue.map((w) => w.change.issue))]
}
/**
* Replay the queue through an apply function (the same guarded write path used
* online), in order. Returns the writes that failed (still unreachable / rejected)
* so they stay queued; everything else drains. Never throws — a failure is data.
*/
export async function replayQueue(
queue: readonly QueuedWrite[],
apply: (change: IssueChange) => Promise<{ ok: boolean }>,
): Promise<{ drained: QueuedWrite[]; remaining: QueuedWrite[] }> {
const drained: QueuedWrite[] = []
const remaining: QueuedWrite[] = []
for (const w of queue) {
let ok = false
try {
ok = (await apply(w.change)).ok
} catch {
ok = false
}
;(ok ? drained : remaining).push(w)
}
return { drained, remaining }
}