SLOPSHOPPER

compact-handoff

EXPERIMENTAL: run /handoff or /handoffplan, then replace the conversation with the handoff and continue in the same terminal. Needs compact-handoff-answer…

newguardcommandtoasttimer
★ 64v0.1.1MITupdated 2026-10-15REMvisual/claude-handoff/mods/compact-handoff
A shopper browsing a rack in a slop shop
Preview · a replayed session in a sandbox
claude · ~/work/app · compact-handoff
› fix the failing auth test and add an audit log call ⏺ Read(src/auth.ts) ⎿ Read 6 lines ⏺ Update(src/auth.ts) ⎿ Added 2 lines, removed 1 line ⏺ Bash(bun test) ⎿ 3 pass, 1 fail ● Done. refresh now rejects expired claims and logs an audit event. ✻ Worked for 42s · done 4:20 PM › /compact-handoff ⎿ compact-handoff: Running /handoff in compact mode. The conversation is replaced when it finishes. ────────────────────────────────────────────────────────────────────────────────────────────────────────────────────── › ? for shortcuts
README

<img src="https://github.com/REMvisual/claude-handoff/releases/download/v1.6.0/banner.png" alt="claude-handoff" width="900"/>

Never lose context between AI coding sessions.

Two skills for Claude Code that capture decisions, failed approaches, measurements, and next steps — so your next session picks up exactly where you left off. Stop wasting 20-40% of each session rediscovering what was already tried.

Now a Claude Code mod: /compact-handoff writes the handoff, swaps the conversation for it, and keeps working in the same terminal. No new window, no paste. Jump to the mod ↓

Download Latest

Install

git clone https://github.com/REMvisual/claude-handoff.git
cp -r claude-handoff/skills/handoff ~/.claude/skills/
cp -r claude-handoff/skills/handoffplan ~/.claude/skills/

Verify: type /handoff in Claude Code — it should appear in autocomplete.

Usage

/handoff              # Capture session context — next session explores
/handoffplan          # Capture context + write a plan — next session executes

No arguments needed. The skill mines your full conversation, gathers git state, validates the output, and gives you a ready-to-paste resume prompt:

Read `plans/handoffs/HANDOFF_fix-auth-bug_2026-03-19.md` (seq 2, PROJ-abc1)
and continue from "Where We're Going".

Paste that into a fresh Claude Code session. It picks up the chain and starts working.

What you get

Every handoff captures a structured snapshot of your session:

┌─────────────────────────────────────────────────────┐
│  HANDOFF_fix-auth-bug_2026-03-19.md                 │
├─────────────────────────────────────────────────────┤
│  The Goal           — what we're solving and why    │
│  Where We Are       — 15-25 bullets of current state│
│  What We Tried      — every approach, chronological │
│  Key Decisions      — what was chosen AND rejected  │
│  Evidence & Data    — real numbers, not summaries   │
│  User Feedback      — preferences, corrections, tone│
│  Where We're Going  — ordered next steps            │
│  Quick Start        — exact commands for next session│
└─────────────────────────────────────────────────────┘

"What We Tried" is the most valuable section. Failed approaches are the most expensive thing to rediscover. This section captures every attempt — what was tried, what happened, why it was kept or abandoned — so the next session never repeats work.

/handoff vs /handoffplan

Both capture your session. The difference is what the next session does:

/handoff/handoffplan
When to usePausing mid-workDone with research, ready to build
What it writesHandoff fileHandoff + phased plan + tracked tasks
Next sessionReads handoff, onboards, exploresReads plan, claims Phase 1, starts coding
Paste prompt"Continue from Where We're Going""Execute the plan starting at Phase 1"

/handoffplan creates beads (or tasks in your tracker) with dependency chains between phases, anti-goals from failed approaches, rollback strategies, and success criteria tied to real numbers from the session.

How it works

  1. Mines your conversation — 12-item extraction checklist covering goals, approaches, failures, decisions, measurements, code analysis, and user preferences
  2. Gathers external state — git log, diff, uncommitted changes, active tasks (in parallel)
  3. Detects chain continuity — finds prior handoffs in the same work stream, inherits chain tag and sequence number
  4. Validates the output — Phase 1 must hit the pass minimum on first write; Phase 2 then fills gaps toward the ceiling
  5. Adapts to context size — at 500K+ tokens, switches to multi-pass map-reduce extraction to avoid the "lost in the middle" problem

Handoffs link across sessions automatically:

HANDOFF_fix-auth_2026-03-17.md  (seq 1)
    └─ HANDOFF_fix-auth_2026-03-18.md  (seq 2)
        └─ HANDOFF_fix-auth_2026-03-19.md  (seq 3)

Your third session on a feature knows about the first two.

Why this exists

Claude already generates session summaries when context compresses. They look like handoffs — but they lack chain tracking, self-validation, evidence mining, and comprehension gates. In controlled A/B testing (same bug, same codebase, fresh sessions), the skill-based session:

  • Required zero user intervention (the unstructured session needed "don't alter anything yet")
  • Traced the complete failure call chain (the unstructured session proposed a surface-level fix)
  • Found bugs not listed in the handoff through structured adjacent exploration
  • Produced justified fixes instead of correct-but-unexplained ones

Prevent skill shadowing

Without this, Claude may generate freeform "handoffs" when sessions end — documents that look right but skip the process. Add to your CLAUDE.md:

When ending a session or saving progress:
- ALWAYS use the /handoff skill. NEVER generate handoff summaries without it.

Works with

The skills work standalone. They unlock more when paired with:

  • Beads — Issue tracking in your repo. Handoffs auto-detect active beads and use them as chain tags.
  • OpenViking — Persistent AI memory. Handoffs search for prior decisions and context from earlier sessions.
  • Git — git log, diff, status gathered automatically.
  • Any CLI task tracker — The skills use bd (beads) as default — swap in Linear, Jira CLI, or GitHub Issues.

compact-handoff mod (experimental)

Runs the handoff, then replaces the conversation with it and keeps going in the same terminal, with no new session and no paste. It works like /compact, but the summary is your validated handoff.

claude plugin marketplace add REMvisual/claude-handoff
claude plugin install compact-handoff@claude-handoff

This also installs its small companion, compact-handoff-answer (see below). The skills above must be installed too, from v2.1.0 or later.

/compact-handoff [note]       # /handoff, then continue from "Where We're Going"
/compact-handoffplan [note]   # /handoffplan, then start executing Phase 1
/compact-handoff cancel       # drop a run that's stuck

What happens:

  1. The skill runs in compact-mode: it asks nothing, commits the session's work, and writes the handoff (and plan).
  2. When that turn ends, the mod checks the files on disk. If anything is missing, empty, interrupted or refused, the context is left untouched.
  3. The conversation is replaced by one short message naming the files. No summarizer call is made, so it takes milliseconds.
  4. The skill's own paste prompt is sent automatically, so chain tracking continues (seq N+1).

The mod also shows a one-time toast when context passes 70%, suggesting /compact-handoff.

Why two plugins: Claude Code doesn't let a plugin answer a compaction it started itself. compact-handoff-answer is about 20 lines. It answers only compactions marked compact-handoff:v1 and leaves every other compaction alone.

Caveats:

  • Experimental. The mod API is early access; tested on Claude Code 2.1.300.
  • Where an organization manages Claude Code, compaction hooks from user-installed plugins don't run.
  • Your own Stop and SessionStart hooks still run around the compaction, so slow ones add to the wait.
  • Classic PreCompact hooks (including the one below) don't run on these compactions. SessionStart hooks with the compact matcher still do.

PreCompact hook (optional)

Safety net that runs before context compaction, capturing ~50 lines of git state and active tasks.

cp claude-handoff/hooks/precompact-handoff.sh ~/.claude/hooks/
chmod +x ~/.claude/hooks/precompact-handoff.sh

Uninstall

rm -rf ~/.claude/skills/handoff ~/.claude/skills/handoffplan
claude plugin uninstall compact-handoff@claude-handoff
claude plugin prune    # removes compact-handoff-answer, installed as its dependency

License

MIT — If this saved you time, give it a star.

Source 3 files
hooks/register.ts 197 lines
1// compact-handoff: /compact-handoff and /compact-handoffplan run the handoff skills in compact mode,
2// then replace the conversation with the handoff and continue in the same session.
3//
4// Order of events (each step is the one the engine allows; see the README):
5//   command  -> timer -> $.command.run(handoff|handoffplan, "compact-mode ...")   (no prompt from inside a command hook)
6//   tool.call Write/Edit records the handoff/plan files the skill writes
7//   turn.complete -> verify the files on disk -> queue /compact-handoff-go       (no compaction while a turn runs)
8//   compact-handoff-go -> timer -> $.session.compact(marker + stub)              (answered by compact-handoff-answer;
9//                                                                                 a plugin cannot answer its own compaction)
10//   -> $.prompt.submit(the skill's own paste prompt)
11import { atom, read, update } from 'claude-code'
12import type { EngineInterface, Register } from 'claude-code'
13
14import type { Mode, Run } from '../types'
15import {
16  buildInstructions,
17  classifyPath,
18  extractPastePrompt,
19  fallbackPrompt,
20  parseSentinel,
21  resolvePath,
22  samePath,
23  stubText,
24} from './logic'
25
26const runAtom = atom({ plugin: 'compact-handoff', key: 'run' } as const, null as Run | null)
27const nudgedAtom = atom({ plugin: 'compact-handoff', key: 'nudged' } as const, false)
28
29const GO = 'compact-handoff-go'
30const NUDGE_PERCENT = 70
31const MAX_COMPACT_TRIES = 40
32
33function say($: EngineInterface, text: string) {
34  $.ui.toast(`compact-handoff: ${text}`)
35}
36
37async function abort($: EngineInterface, why: string) {
38  await update($, runAtom, () => null)
39  say($, `${why}. Context unchanged.`)
40}
41
42async function start($: EngineInterface, mode: Mode, args: string) {
43  args = (args ?? '').trim()
44  if (args === 'cancel') {
45    await update($, runAtom, () => null)
46    return { text: 'cancelled.' }
47  }
48  const running = await read($, runAtom)
49  if (running) {
50    return { text: 'a run is already in progress (`/compact-handoff cancel` to drop it).' }
51  }
52  await update($, runAtom, () => ({ mode, stage: 'handoff' as const, startedAt: Date.now(), files: [] }))
53  const skill = mode === 'plan' ? 'handoffplan' : 'handoff'
54  const skillArgs = `compact-mode ${args}`.trim()
55  // A prompt or command raised from inside this hook would wait on the turn the hook holds.
56  $.clock.after(100, () => runSkill($, skill, skillArgs))
57  return { text: `Running /${skill} in compact mode. The conversation is replaced when it finishes.` }
58}
59
60async function runSkill($: EngineInterface, skill: string, args: string) {
61  try {
62    await $.command.run({ command: skill, args })
63  } catch (err) {
64    try {
65      await $.prompt.submit({ text: `Run the ${skill} skill with these arguments: ${args}` })
66    } catch (err2) {
67      await abort($, `could not start /${skill} (${String(err2 ?? err)})`)
68    }
69  }
70}
71
72async function finishHandoff($: EngineInterface, reason: string, text: string) {
73  const run = await read($, runAtom)
74  if (!run || run.stage !== 'handoff') return
75  if (reason !== 'answer') return abort($, `the handoff turn ended early (${reason})`)
76
77  const sentinel = parseSentinel(text)
78  if (sentinel && !sentinel.ok) return abort($, `the skill reported: ${sentinel.reason}`)
79
80  const reported = sentinel?.ok ? sentinel.files : []
81  const pick = (kind: 'HANDOFF' | 'PLAN') =>
82    reported.find(f => classifyPath(f) === kind) ?? [...run.files].reverse().find(f => classifyPath(f) === kind)
83  const handoff = pick('HANDOFF')
84  const plan = run.mode === 'plan' ? pick('PLAN') : undefined
85  if (!handoff) return abort($, 'no handoff file was written')
86  if (run.mode === 'plan' && !plan) return abort($, 'no plan file was written')
87  if (sentinel?.ok && run.files.length && !run.files.some(f => samePath(f, handoff))) {
88    say($, `the skill named ${handoff}, which no Write/Edit in this run touched; checking it on disk anyway`)
89  }
90
91  const cwd = await $.session.cwd()
92  for (const file of [handoff, plan].filter((f): f is string => !!f)) {
93    let content = ''
94    try {
95      content = await $.fs.read(resolvePath(cwd, file))
96    } catch {
97      return abort($, `cannot read ${file}`)
98    }
99    if (content.trim().length < 200) return abort($, `${file} looks empty`)
100  }
101
102  const pastePrompt = extractPastePrompt(text) ?? fallbackPrompt(run.mode, handoff, plan)
103  await update($, runAtom, r => (r ? { ...r, stage: 'compacting' as const, handoff, plan, pastePrompt } : r))
104  // Compaction is refused while a turn runs; a queued command runs once the session is idle.
105  $.command.run({ command: GO }).catch(err => abort($, `could not queue the compaction (${String(err)})`))
106}
107
108function compactLoop($: EngineInterface, attempt: number) {
109  void (async () => {
110    const run = await read($, runAtom)
111    if (!run || run.stage !== 'compacting' || !run.handoff) return
112    const instructions = buildInstructions(stubText(run.mode, run.handoff, run.plan))
113    let result
114    try {
115      result = await $.session.compact({ instructions })
116    } catch (err) {
117      if (attempt >= MAX_COMPACT_TRIES) return abort($, `the compaction kept being refused (${String(err)})`)
118      $.clock.after(500, () => compactLoop($, attempt + 1))
119      return
120    }
121    if (result.skip !== undefined) return abort($, `the compaction was skipped: ${result.skip}`)
122    if (result.usage) say($, 'compact-handoff-answer is not installed, so the engine summarized instead')
123    await update($, runAtom, () => null)
124    await update($, nudgedAtom, () => false)
125    try {
126      await $.prompt.submit({ text: run.pastePrompt ?? fallbackPrompt(run.mode, run.handoff, run.plan), asUser: true })
127    } catch (err) {
128      say($, `context replaced, but the continue-prompt could not be sent (${String(err)}). Paste it from the handoff.`)
129    }
130  })()
131}
132
133async function nudge($: EngineInterface) {
134  if (await read($, nudgedAtom)) return
135  const usage = await $.session.usage()
136  if ((usage.context.percent ?? 0) < NUDGE_PERCENT) return
137  await update($, nudgedAtom, () => true)
138  say($, `context at ${Math.round(usage.context.percent ?? 0)}%. /compact-handoff is recommended.`)
139}
140
141export const register: Register = on => {
142  on('session.start', async ($, e, next) => {
143    await $.command.register({
144      name: 'compact-handoff',
145      description: 'Run /handoff, then replace the conversation with it and continue (experimental)',
146    })
147    await $.command.register({
148      name: 'compact-handoffplan',
149      description: 'Run /handoffplan, then replace the conversation with it and start Phase 1 (experimental)',
150    })
151    await $.command.register({ name: GO, description: 'Internal step of /compact-handoff' })
152    return next(e)
153  })
154
155  // The internal step stays runnable but out of the typeahead and /help.
156  on('command.describe', { command: GO }, async ($, e, next) => ({ ...(await next(e)), isHidden: true }))
157
158  on('command.run', { command: 'compact-handoff' }, async ($, e) => start($, 'resume', e.args))
159  on('command.run', { command: 'compact-handoffplan' }, async ($, e) => start($, 'plan', e.args))
160  on('command.run', { command: GO }, async $ => {
161    const run = await read($, runAtom)
162    if (!run || run.stage !== 'compacting') return { text: 'nothing to compact.' }
163    // Compacting from inside this hook is refused too; a timer runs after it returns.
164    $.clock.after(300, () => compactLoop($, 0))
165    return { text: 'replacing the conversation with the handoff.' }
166  }).catch(() => ({ text: 'internal step failed. Context unchanged.' }))
167
168  on('tool.call', { tool: 'Write' }, async ($, e, next) => {
169    const res = await next(e)
170    if (!e.agentId && !res.deny && !res.isError && classifyPath(e.file_path)) {
171      await update($, runAtom, r =>
172        r && r.stage === 'handoff' && !r.files.includes(e.file_path) ? { ...r, files: [...r.files, e.file_path] } : r,
173      )
174    }
175    return res
176  }).catch(($, e, next) => next(e))
177
178  on('tool.call', { tool: 'Edit' }, async ($, e, next) => {
179    const res = await next(e)
180    if (!e.agentId && !res.deny && !res.isError && classifyPath(e.file_path)) {
181      await update($, runAtom, r =>
182        r && r.stage === 'handoff' && !r.files.includes(e.file_path) ? { ...r, files: [...r.files, e.file_path] } : r,
183      )
184    }
185    return res
186  }).catch(($, e, next) => next(e))
187
188  on('turn.complete', async ($, e, next) => {
189    const res = await next(e)
190    if (e.agentId) return res
191    const run = await read($, runAtom)
192    if (run) await finishHandoff($, e.reason, res.text)
193    else await nudge($)
194    return res
195  }).catch(($, e, next) => next(e))
196}
197
hooks/logic.ts 95 lines
1// Pure helpers for the compact-handoff mod: no `$`, so the tests can call them directly.
2import type { Mode } from '../types'
3
4/**
5 * First line of the compaction instructions. compact-handoff-answer answers only a
6 * compaction whose instructions start with it, and returns the text after it.
7 */
8export const MARKER = 'compact-handoff:v1'
9
10const HANDOFF_FILE = /(?:^|[\\/])(?:plans|\.claude)[\\/]handoffs[\\/](HANDOFF|PLAN)_[^\\/]*\.md$/
11
12/** `HANDOFF` or `PLAN` for a file the handoff skills write, else null. */
13export function classifyPath(path: string): 'HANDOFF' | 'PLAN' | null {
14  const m = HANDOFF_FILE.exec(path)
15  if (!m) return null
16  return m[1] === 'PLAN' ? 'PLAN' : 'HANDOFF'
17}
18
19export type Sentinel = { ok: true; files: string[] } | { ok: false; reason: string }
20
21/** Reads the skill's last `COMPACT-HANDOFF-READY:` / `COMPACT-HANDOFF-FAILED:` line; null when there is none. */
22export function parseSentinel(text: string): Sentinel | null {
23  const lines = text.split(/\r?\n/).map(l => l.trim().replace(/^`+|`+$/g, '').trim())
24  for (let i = lines.length - 1; i >= 0; i--) {
25    const line = lines[i] ?? ''
26    const ready = /^COMPACT-HANDOFF-READY:\s*(.+)$/.exec(line)
27    if (ready) {
28      const files = (ready[1] ?? '')
29        .split('|')
30        .map(f => f.trim().replace(/^`+|`+$/g, ''))
31        .filter(Boolean)
32      return { ok: true, files }
33    }
34    const failed = /^COMPACT-HANDOFF-FAILED:\s*(.*)$/.exec(line)
35    if (failed) return { ok: false, reason: failed[1] || 'no reason given' }
36  }
37  return null
38}
39
40/** The prompt between the dashes after `PASTE THIS INTO YOUR NEXT SESSION:`, or null. */
41export function extractPastePrompt(text: string): string | null {
42  const lines = text.split(/\r?\n/)
43  const head = lines.findIndex(l => l.includes('PASTE THIS INTO YOUR NEXT SESSION:'))
44  if (head < 0) return null
45  const isRule = (l: string) => /^\s*-{10,}\s*$/.test(l)
46  let start = head + 1
47  if (isRule(lines[start] ?? '')) start++
48  const end = lines.findIndex((l, i) => i >= start && isRule(l))
49  if (end < 0) return null
50  const body = lines
51    .slice(start, end)
52    .map(l => l.replace(/^ {3}/, ''))
53    .join('\n')
54    .trim()
55  return body || null
56}
57
58/** Used only when the skill's own paste prompt could not be extracted. */
59export function fallbackPrompt(mode: Mode, handoff: string, plan?: string): string {
60  if (mode === 'plan' && plan) {
61    return `Read \`${plan}\` and \`${handoff}\`. Execute the plan starting at Phase 1: claim its bead, read the Quick Start files, then start coding Phase 1's first concrete action.`
62  }
63  return `Read \`${handoff}\` and continue from "Where We're Going". Summarize what you understand, then wait for my go-ahead before executing.`
64}
65
66/** The one message the conversation is replaced with. */
67export function stubText(mode: Mode, handoff: string, plan?: string): string {
68  const files = plan ? `Handoff: \`${handoff}\`. Plan: \`${plan}\`.` : `Handoff: \`${handoff}\`.`
69  return (
70    `The conversation was replaced by compact-handoff (${mode} mode) after a full handoff. ${files} ` +
71    'Everything earlier in this session is in those files; the next message says what to do.'
72  )
73}
74
75/**
76 * The compaction's instructions: the marker line, then the stub. Without the answering
77 * plugin the engine's summarizer reads them instead, so they must make sense to it too.
78 */
79export function buildInstructions(stub: string): string {
80  return `${MARKER}\n${stub}`
81}
82
83/** Makes a skill-reported relative path absolute against the session's directory. */
84export function resolvePath(cwd: string, path: string): string {
85  if (/^([a-zA-Z]:[\\/]|[\\/])/.test(path)) return path
86  const sep = cwd.includes('\\') ? '\\' : '/'
87  return `${cwd.replace(/[\\/]+$/, '')}${sep}${path.replace(/^\.[\\/]/, '')}`
88}
89
90/** Same file, whatever the separators or case of a Windows path. */
91export function samePath(a: string, b: string): boolean {
92  const norm = (p: string) => p.replace(/\\/g, '/').toLowerCase()
93  return norm(a) === norm(b) || norm(a).endsWith('/' + norm(b)) || norm(b).endsWith('/' + norm(a))
94}
95
types/index.d.ts 22 lines
1export type Mode = 'resume' | 'plan'
2
3/** One /compact-handoff run, from the command until the continue-prompt is sent. */
4export type Run = {
5  mode: Mode
6  /** `handoff`: the skill's turn is running. `compacting`: waiting for the compaction. */
7  stage: 'handoff' | 'compacting'
8  startedAt: number
9  /** Handoff/plan files written during the run, as the tools reported them. */
10  files: string[]
11  /** Set once the turn ended with verified files. */
12  handoff?: string
13  plan?: string
14  pastePrompt?: string
15}
16
17declare module 'claude-code' {
18  interface PluginState {
19    'compact-handoff': { run: Run | null; nudged: boolean }
20  }
21}
22