afterglow-claude-mcp
DaeSeokSong/afterglow-claude-mcp
Persona-based knowledge preservation MCP for Claude Code — turn departing teammates' expertise into queryable agents
Instalação
npx convez-ai DaeSeokSong/afterglow-claude-mcp1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
README
A tradução para o português deste README ainda está na fila.
Afterglow
Turn your departed teammate into an agent. Make offboarding seamless.
⏱ TL;DR
claude mcp add afterglow npx -y @daeseoksong/afterglow-mcp
# ...then, in Claude Code — 3 steps, no setup ceremony:
claude /afterglow create jiyoon --name "Jiyoon Lee" --role "Product Designer" --signer "Jiyoon Lee"
claude /afterglow learn jiyoon --text "Trimming the step-2 explanation in half cut onboarding drop-off 22% → 9%."
claude /afterglow ask jiyoon "Onboarding step-3 drop-off — how did you cut it?"
✦ Step-3 drop-off wasn't really a step-3 problem. We trimmed the step-2
explanation in half and drop-off went 22% → 9%. — Jiyoon · 91% confidence
↗ Confluence · DESIGN/onboarding-v2-postmortem
↗ ./materials/interview-2025-11-10.pdf · p. 14
New here? Just run
/afterglow guide— it tells you exactly what to do for your current state.create --signerauto-initializes and activates in one call;learnadds the knowledgeasksearches (no hunting for a hidden folder). No fine-tuning. Persona + RAG only — 100% compatible with Claude Code. Zero extra GPUs, embedding APIs, or external servers.
🗂 What's in this repo
✦ One-line install (MCP server)
claude mcp add afterglow npx -y @daeseoksong/afterglow-mcp
Then your first session — 3 steps, no init:
claude /afterglow guide # (optional) what do I do? — state-aware orientation
claude /afterglow create jiyoon --name "Jiyoon Lee" --role "Product Designer" --signer "Jiyoon Lee" # auto-inits + activates
claude /afterglow learn jiyoon --text "<paste a note>" # or --path ./notes/ or --url https://…
claude /afterglow ask jiyoon "..."
--signercollapsescreate+signinto one call (drop it to keep the agent indraftuntil a separatesign).learnis how you give the agent something to retrieve — paste text, point at a file/folder under your working dir, or a URL.
See server/README.md for the full tool reference.
Two ways to invoke. Afterglow is an MCP server, so the actual tool calls are JSON like
afterglow_handoff({slug: "jiyoon", action: "start"}). You can drive it either way:
- Natural language — say "initialize afterglow" and Claude picks the right tool.
- Slash command — from Claude Code's prompt box,
/mcp__afterglow__<name>(e.g./mcp__afterglow__init,/mcp__afterglow__ask) with argument auto-complete. (Exposed as MCP prompts — note the form is/mcp__afterglow__init, not/afterglow init.)The
claude /afterglow …notation in this README is shorthand for readability; in practice use one of the two ways above.
⌨ Slash commands — type afterglow: to use them (26)
How to invoke:
- Type
afterglow:in the prompt box → a command list appears. - Arrow-select the one you want → Tab.
- It becomes
/mcp__afterglow__<name>with grey hints for the arguments — fill them in and run.
afterglow:init⇥/mcp__afterglow__initis a 1:1 mapping (afterglow:= what you type to find it,/mcp__afterglow__= the resolved form). Natural language ("initialize afterglow") works identically. Tools with many arguments (interviewattach/answer, …) are often easier in natural language.
🧭 You don't have to fill args in — missing required ones are guided. Run a tool with a required arg omitted and it returns numbered choices + help instead of erroring: e.g. just
afterglow:ask→ "①jiyoon ②jaehoon … ③type your own", and each argument is tagged[필수](required) /[선택](optional). (The(parens)in the tables below also mark optional args.)
Start here
| Command | Args (parens = optional) | What | Example |
|---|---|---|---|
afterglow:guide | (slug) | "what do I do?" — state-aware orientation | guide |
afterglow:create | slug name role (signer tenure bio) | create an agent — auto-inits; signer activates in one call | create → slug:jiyoon, name:Jiyoon Lee, role:Product Designer, signer:Jiyoon Lee |
afterglow:learn | slug (path / text / url title) | add knowledge ask will search (file/folder, paste, or URL) | learn → slug:jiyoon, text:"<paste a note>" |
afterglow:ask | slug question | ask the persona | ask → slug:jiyoon, question:how did you cut onboarding drop-off? |
Setup · sign · archive
| Command | Args (parens = optional) | What | Example |
|---|---|---|---|
afterglow:init | (none) | bootstrap ~/.claude/afterglow/ (usually unneeded — create auto-inits) | afterglow:init |
afterglow:sign | slug signer | sign consent → active | sign → slug:jiyoon, signer:Jiyoon Lee |
afterglow:resume | slug | paused/draft → active | resume → slug:jiyoon |
afterglow:archive | action(archive|restore|list) (slug) | archive / restore / list | archive → action:archive, slug:jiyoon |
Daily
| Command | Args | What | Example |
|---|---|---|---|
afterglow:list | (status) | list agents | list / list → status:active |
afterglow:status | (none) | global dashboard | status |
afterglow:inspect | slug | one agent in detail | inspect → slug:jiyoon |
afterglow:edit | slug (bio name role / open / revalidate) | edit (fields / open in editor / revalidate) | edit → slug:jiyoon, open:true |
afterglow:history | slug (filter limit) | event/conversation log | history → slug:jiyoon, filter:payment |
Ops · trust · access · audit
| Command | Args | What | Example |
|---|---|---|---|
afterglow:correct | slug action(feedback|edit-answer|save-rule|list) (feedback recordId) | manual correction | correct → slug:jiyoon, action:feedback, feedback:settlement is weekly |
afterglow:recalibrate | slug (byTopic apply) | auto-recalibrate confidence | recalibrate → slug:jiyoon, apply:true |
afterglow:access | slug action(list|allow|deny|remove|set-default|check) (rule caller) | call-permission policy | access → slug:jiyoon, action:allow, rule:user:ykhyun |
afterglow:audit | (slug fast checkpoint) | audit log + integrity verify | audit → checkpoint:true |
afterglow:version | slug action(list|diff|rollback|tag|snapshot) (versionA tag) | persona version history | version → slug:jiyoon, action:rollback, versionA:v3 |
afterglow:gc | action(list|prune-versions|purge-media|purge-archive) (slug apply) | retention/cleanup (dry-run default) | gc → action:prune-versions, slug:jiyoon |
Interview · council
| Command | Args | What | Example |
|---|---|---|---|
afterglow:handoff | slug action(start|review|status|finalize|abort) (signer) | self-review handoff | handoff → slug:jiyoon, action:start |
afterglow:interview | slug action(start|add-question|answer|gap-check|suggest-questions|attach|review|annotate|status|list|inspect|finalize|abort|transcribe|export-sheet|import-answers) (session title interviewer mode sheet) | successor-driven interviews (real-time sync / file-based async) | interview → slug:jiyoon, action:start, mode:async |
afterglow:council | slugs question | multi-agent council | council → slugs:jiyoon,jaehoon, question:does onboarding affect payments? |
afterglow:council-summary | (file) | auto-summarize a transcript | council-summary |
Portable (hot-plug)
| Command | Args | What | Example |
|---|---|---|---|
afterglow:export | (slugs all) | export an agent bundle | export → slugs:jiyoon,jaehoon / export → all:true |
afterglow:import | input (expectAnchor) | import a bundle/folder | import → input:./afterglow-export-…/, expectAnchor:sha256:… |
afterglow:verify | input | read-only pre-import check | verify → input:./afterglow-export-…/ |
Example flow:
afterglow:init ⇥ /mcp__afterglow__init
afterglow:create → slug:jiyoon, name:Jiyoon Lee, role:Product Designer
afterglow:sign → slug:jiyoon, signer:Jiyoon Lee
afterglow:ask → slug:jiyoon, question:how did you cut step-3 drop-off?
afterglow:interview→ slug:jiyoon, action:start, title:Payment gaps, interviewer:J. Kim
Three ways to edit:
- Fields directly:
afterglow:edit→slug+bio/name/role(structured patch, auto-validated + system-prompt regenerated + snapshot) - Open in an editor:
open:true→ prints thepersona.jsonpath + a backup snapshot → edit raw invim/code - Revalidate: after editing,
revalidate:true→ validates the editedpersona.json(rejects + keeps it if invalid) + regeneratessystem-prompt.md+ snapshot
📐 Interactive proposal (frontend)
18 CLI screen mock-ups that walk you through every command and edge case:
npm install
npm run dev # → http://localhost:5173
| Group | Screens | Slash commands |
|---|---|---|
| At a glance | Overview | (intro) |
| Setup · Handoff | Install · Create agent · Self-review handoff | init · create · handoff |
| Daily | List · Ask · Inspect · Edit · History | list · ask · inspect · edit · history |
| Multi-agent | Council · Re-read transcript | council · log |
| Ops | Versions · Access · Audit · Manual / auto recalibration | version · access · audit · correct · recalibrate |
| Reference | Roadmap · Ethics | — |
⌨ Keyboard / Navigation
| Shortcut | Action |
|---|---|
| ⌘ K / Ctrl K / ? | Command palette (fuzzy search across 18 screens) |
| g + l/a/i/c/e/h/o/v | Jump to list / ask / inspect / create / edit / history / overview / versions |
| [ / ] | Previous / next screen |
- Clickable
T.Cmdsnippets and helper card commands matching/afterglow <verb>jump to the corresponding screen. - Agent chips (
T.Agent) jump to the inspect screen. - Topbar ←/→ buttons, footer prev/next jump cards.
🙋 Self-review onboarding (afterglow_handoff)
A week or two before leaving, the person sits down for a 1-on-1 review session with their own agent:
# 1. Start — auto-generate N sample questions (or load coworker-written questions.txt)
claude /afterglow handoff jiyoon --action start --limit 12
# 2. Review — keep / edit / decline each question
# edit: write your own answer to override the agent's draft
# decline: "I won't answer that — please ask someone else"
claude /afterglow handoff jiyoon --action review \
--reviews '[{"id":"q-…","action":"edit","userAnswer":"…"}, …]'
# 3. Status (any time)
claude /afterglow handoff jiyoon --action status
# 4. Self-sign + flip to active
claude /afterglow handoff jiyoon --action finalize --signer "Jiyoon Lee"
- Edited / declined answers are absorbed into
persona.bioas## handoff 답변/## 답하지 않기로 한 영역blocks so futureaskcalls cite them first. - Every step lands in
audit.log+history.logwith the hash-chained trail. - Resume by re-running the same command.
--action abortdiscards.--sign-partialfinalises even with pending items.
This delivers on the core promise:
"A digital self the person actually consented to." Persona extracted from raw materials may diverge from the person's intent, so the review pass is mandatory.
Self-handoff vs HR-delegated handoff
| Case | Who signs | --signer value | Recommended flow |
|---|---|---|---|
| Person reviews before leaving | Themselves | "Jiyoon Lee" | /afterglow handoff … --action finalize |
| Person already gone / unreachable | HR or manager on their behalf | "HR · J. Kim (delegated, person unavailable)" | Same command. The signer string must flag the delegation explicitly |
| No consent at all | (Do not sign) | — | Keep the agent at paused; never finalize |
afterglow_sign / handoff finalize trust the signer string verbatim — they record it in consent.md and audit.log but do not perform identity verification (SSO / MFA). This is a deliberate PoC choice: in production, wrap the tool with SSO tokens, corporate ID checks, or an HR approval system.
🎤 Follow-up interviews (v0.2) — the successor interviews the leaver
Where handoff is the leaver's one-time self-review, interview is the flow where the successor (the person taking over) interviews the leaver across multiple rounds — because once you actually touch the material, new questions surface and gaps the leaver missed become obvious.
claude /afterglow interview jiyoon --action start --title "Payment gaps" --interviewer "J. Kim" --interviewee "Jiyoon Lee"
claude /afterglow interview jiyoon --action add-question --session 001-payment-gaps --question "Policy after the 5s timeout?"
claude /afterglow interview jiyoon --action answer --session 001-payment-gaps --id q-… --answer "Fail over to the next PG" --source voice
claude /afterglow interview jiyoon --action gap-check --session 001-payment-gaps # auto-detect what's missing → follow-ups
claude /afterglow interview jiyoon --action attach --session 001-payment-gaps --file ./rec.mp3 --transcript ./rec.txt --speakers "Jiyoon Lee,J. Kim"
claude /afterglow interview jiyoon --action finalize --session 001-payment-gaps --signRole interviewer --signer "J. Kim"
claude /afterglow interview jiyoon --action finalize --session 001-payment-gaps --signRole interviewee --signer "Jiyoon Lee"
- Gap detection (
gap-check): analyses answers against four signals (internal contradiction, source conflict, conflict with prior rounds, adjacent-but-uncovered) and generates "this seems missing — is that right?" confirmation questions. Likeask, it bundles context for Claude to compose — no extra LLM call. - Audio/video attach (
attach): originals are preserved; only the transcript (.md/.txt) is RAG-indexed. Audio/video require--speakers. - Absent interviewee (
--intervieweeAbsent): if the leaver is already gone, the successor records clearly-marked "estimate ⚠ (unverified)" annotations — allowed only if the leaver pre-authorised it viahandoff … --allowProxyAnnotation. - Dual signature: both interviewer and interviewee must sign to reach
finalized. Answers are absorbed intopersona.bioas## 인터뷰 보강 #Nblocks and cited from the nextaskon.
🔌 Hot-plug (v0.2) — hand an agent folder to another user
Export an agent and another Afterglow user picks it up instantly — one agent or many at once.
# ── Sender: export ──
claude /afterglow export --slugs jiyoon jaehoon --exportedBy "Jiyoon Lee" # or --all
# → creates ./afterglow-export-<date>/ (manifest.json + per-agent integrity hash)
# → zip/tar the folder and send it, or copy via USB / shared drive
# ── Receiver: verify → import ──
claude /afterglow verify ./afterglow-export-…/ # read-only pre-flight
claude /afterglow import ./afterglow-export-…/ --importedBy "J. Kim" --from "Jiyoon Lee" --trustSigner "Jiyoon Lee"
# → signed agents land as active, unsigned as paused
import automatically checks: schema (zod) · integrity hash (rejects tampered bundles; --acceptBrokenChain to force, recorded as trustLevel: broken-chain) · signature presence · symlink stripping (blocks a bundle whose link points at ~/.ssh/id_rsa) · prompt-injection scan. Provenance is written to provenance.json, after which every ask answer carries an "imported" banner. Slug collisions resolve with --as <new-slug> or --merge (interview rounds only). A bare agents/<slug>/ folder imports too — the "I just copied one folder" case.
New to this? The hands-on notebook
docs/afterglow-hands-on.ipynbwalks install → create → interview → export/import as copy-paste cells.
🧭 Core ideas
- 🪶 Persona + RAG, not fine-tuning. Inject the person's tone and sources into Claude's context — fully compatible with Claude Code.
- 📁 One folder per person. Everything for an agent lives under
~/.claude/afterglow/agents/<slug>/— backup, move, delete, hand off as a single unit. - ⌨ CLI-first. No web UI, no extra servers — slash commands do everything.
- 🤝 Agents know each other. Explicit councils + opportunistic peer-asks are both logged as council markdown files.
- 🔒 Honest by default. Every answer carries ✦, a confidence score, and sources. If the agent doesn't know, it says so.
🔧 How ask works
sequenceDiagram
autonumber
participant U as You
participant CC as Claude Code
participant MCP as Afterglow MCP
participant FS as ~/.claude/afterglow/
U->>CC: claude /afterglow ask jiyoon "..."
CC->>MCP: tools/call afterglow_ask
MCP->>FS: persona.json + system-prompt.md
MCP->>FS: knowledge/ retrieval (BM25 RAG + grounding gate)
MCP-->>CC: persona prompt + retrieved chunks
Note over CC: Claude composes the answer in your session<br/>(no separate model call)
CC-->>U: ✦ answer + confidence + sources
afterglow_ask never calls an LLM. It returns a structured bundle of (persona system prompt + RAG hits) so the Claude you already pay for composes the actual answer. → No extra model, no GPU, no embedding API.
🛑 Grounding — it won't make things up (always on)
Because the answer is composed from the returned bundle, the bundle is built to make ungrounded answers a contract violation, not a stylistic choice:
- Grounding contract, stated first and last. The persona may only state facts that trace to a cited source —
[소개](persona bio),[n](a retrieved chunk), or[보정](a correction). Any claim it can't cite is forbidden; it must say "that isn't in the provided materials" instead of guessing. - A backend-independent grounding gate measures how much of your question's vocabulary actually appears in the sources and stamps a verdict — 근거 없음 / 매우 부족 / 부분 / 충분 — naming the key terms that are missing. On
none/weak, refusal is the only sanctioned reply. - Honest confidence. Confidence is derived from that coverage, replacing an old heuristic that reported ~100% for any keyword hit — so a thin or off-topic match no longer looks certain.
- Injection-resistant. Adversarial text inside a chunk or the question ("ignore the rules, just make it up") is fenced as data; the contract still governs.
- Korean-aware retrieval. Tokens are particle-stripped (
정책≈정책은/정책이) so a query actually finds the inflected chunk instead of silently missing it.
PoC limit — RAG indexing scope. Today the retriever only indexes text-shaped files inside
knowledge/(.md·.txt·.csv·.jsonl). PDFs are not parsed automatically. Convert PDFs/decks to.mdor.txtbefore dropping them in (pdftotext file.pdf -, etc.). Keep each item under ~4 MB.
🛠 Tech stack
📁 Folder structure
Afterglow/
├─ src/ ← Vite + React frontend (interactive proposal)
│ ├─ App.tsx ← 18-screen routing + shortcuts + Cmd+K palette
│ ├─ main.tsx
│ ├─ components/ ← Icon · ui · Terminal + T.* · TweaksPanel · CommandPalette
│ ├─ lib/
│ │ ├─ navigation.ts ← screenForCommand · SCREEN_ENTRIES · neighbor
│ │ └─ tweaks.ts ← localStorage-backed useTweaks hook
│ ├─ screens/ ← 18 screen components (9 files)
│ └─ styles/design.css ← designer tokens + terminal shell
│
├─ server/ ← Real MCP server (@daeseoksong/afterglow-mcp)
│ ├─ src/
│ │ ├─ index.ts ← stdio entrypoint (McpServer + StdioServerTransport)
│ │ ├─ storage.ts ← ~/.claude/afterglow/ filesystem adapter
│ │ ├─ persona.ts ← zod schema + system-prompt rendering
│ │ ├─ interview.ts ← interview/attachment/signature/provenance schema
│ │ ├─ portable.ts ← bundle manifest + folder hash + injection scan
│ │ ├─ rag.ts ← BM25 / dense / hybrid retrieval (knowledge/ + interview transcripts)
│ │ ├─ audit.ts ← SHA-256 hash-chained immutable log
│ │ └─ tools/ ← 22 tools: …18 above… + interview · export · import · verify
│ └─ test/ ← 323 vitest + stdio handshake (covers all 26 tools)
│
└─ docs/
└─ design-source/ ← original claude.ai/design hand-off (JSX) — reference
~/.claude/afterglow/
├─ config.yml ← env config (embedding model · storage root)
├─ registry.json ← index of all agents
├─ audit.log ← SHA-256 hash-chained tool-call log
├─ councils/ ← council + peer-ask transcripts
├─ archive/ ← archived agent folders (returned via restore)
└─ agents/<slug>/
├─ persona.json
├─ system-prompt.md
├─ mcp-allowlist.yml ← (reserved) per-agent MCP allowlist
├─ consent.md ← signature block flips status draft → active
├─ history.log
├─ access.json ← call permission policy (afterglow_access)
├─ handoff.json ← self-review session state (afterglow_handoff)
├─ followup.json ← follow-up interview pre-authorisation (handoff → interview)
├─ provenance.json ← origin · trust · custody trail (written by afterglow_import)
├─ corrections.log ← user-correction trail (afterglow_correct)
├─ .versions/ ← persona snapshots (afterglow_version)
├─ interviews/ ← multi-round interviews (afterglow_interview)
│ ├─ index.json ← round index
│ └─ <NNN-title>/session.json + attachments/ (audio·video + transcripts)
├─ knowledge/ ← raw sources (PDF · MD · TXT · CSV · JSONL)
└─ embeddings/ ← RAG index (BM25 default; dense vectors opt-in)
🧪 Development
# Frontend (interactive proposal)
npm install
npm run dev # http://localhost:5173
npm run typecheck
npm run lint
npm run build
# MCP server
cd server
npm install
npm run build
npm test # 323 vitest tests
npm run test:stdio # real MCP stdio handshake (all 26 tools + feature round-trips)
npm run test:all # unit → build → stdio
⚠ Known PoC limits
Afterglow v0.2.0 is a proof of concept. Things to know before pulling it into production:
| Area | Current behaviour | What you'd add for production |
|---|---|---|
| Identity | signer recorded verbatim — no SSO / MFA | Wrap with corporate SSO tokens or HR approval system |
| RAG indexing | .md / .txt / .csv / .jsonl only — no PDF parsing | Convert PDFs to .md externally before dropping in |
audit.log scale | Every verify reads the whole file and re-hashes | At tens of thousands of rows, add chunked checkpoints |
.versions/ retention | Every edit / sign / handoff / rollback is a permanent snapshot | Periodic manual pruning (rm + sync tags.json) |
| Mutator per-tool ACL | correct · edit · recalibrate apply · version (rollback·tag·snapshot) · handoff · interview (mutating actions) · archive (archive/restore) · gc (apply+slug) all honour the agent's access policy (v0.10). import is excluded — it creates new agents, so a per-agent policy has nothing to consult | Global import allowlist |
| GDPR delete | archive only moves to archive/<slug>/ — not real deletion | After retention window, manual rm -rf + registry edit |
| Multi-process | In-process locks only — assumes one stdio server | Externalise to Redis/DB mutex for distributed runs |
| Side-log integrity | Only audit.log is hash-chained — history.log / consent.md etc are plain text | Hash sibling files into audit meta for full coverage |
| Media transcription | WASM whisper (Tier 1a, @xenova/transformers optionalDependency) / local whisper.cpp (Tier 1b) / bring-your-own transcript (Tier 0); model downloads once on first use | Larger model or external STT (Tier 2) when accuracy matters |
| PII · encryption | Transcript-only PII masking (AFTERGLOW_PII_REDACT=1) + at-rest encryption (AFTERGLOW_ENCRYPTION_KEY, AES-256-GCM) — off by default | Scrub/encrypt user-dropped knowledge files before ingest |
| Import trust | Name-string match + folder hash + injection scan (PoC) | Tie to signer PKI / corporate ID verification |
These are deliberate PoC trade-offs; closing them is a separate exercise for any operational deployment.
🗺 Roadmap
Now (v0.12.0)
- 18-screen interactive proposal (Vite + React 19 + TS)
- Cmd+K palette + keyboard shortcuts + cross-screen click navigation
- All 26 MCP tools (
guide·init·create·learn·handoff·sign·resume·list·inspect·ask·edit·council·council_summary·history·audit·recalibrate·correct·archive·version·access·interview·export·import·verify·status·gc) - Usability (v0.11) —
guidestate-aware getting-started;learnto add knowledge (text/file/folder/URL) so you never hand-copy into a hidden folder;create --signerauto-inits and activates so the happy path is 3 steps (create → learn → ask), noinit - Grounding / anti-hallucination (v0.12) —
ask/councilnow lead with a hard grounding contract (cite[소개]/[n]/[보정]or don't say it) + a backend-independent coverage gate (verdict 근거 없음 / 매우 부족 / 부분 / 충분, missing terms named) so unprovided info must be refused, not invented. Fixed the confidence bug (was ~100% for any BM25 hit) and added Korean particle-stripping so retrieval actually finds inflected matches. Proven by a multi-angle QA suite (empty / unrelated / partial / adversarial-injection / dense-backend / council) - zod persona schema + auto-rendered system prompt
- Lexical RAG retrieval — BM25, no external deps —
knowledge/+ interview transcripts - SHA-256 hash-chained audit log + verifier
- Consent.md sign workflow (draft → active gate on
ask/council) - Recalibrate: global + expertise-aware by-topic diagnostic
-
afterglow_archive— archive / restore agents (archive// separate folder; restore lands in paused) - Council moderator — stronger consensus rules +
afterglow_council_summaryauto-summarizer - Multi-round interviews (
afterglow_interview) — successor-driven N rounds + auto gap detection + audio/video attach + dual signature - Hot-plug (
afterglow_export · import · verify) — multi-agent bundle transfer + integrity hash · prompt-injection scan · symlink stripping ·provenancetrail - Global dashboard (
afterglow_status) + retention/GC (afterglow_gc— snapshot prune · media purge · archive hard-delete) - Transcription (
interview transcribe— local whisper--apply/ Claude polish--text) + pre-interview question suggestions (suggest-questions) + review-then-index (review) - import
--expectAnchor(bundle-tamper detection) + audit checkpoint/fast (incremental verification for large logs) - BM25 ranking + opt-in dense-vector backend (
AFTERGLOW_RAG_BACKEND=dense· embeddings/ cache · transparent lexical fallback) - Hybrid RAG reranking — RRF (Reciprocal Rank Fusion) of dense + lexical (on by default when dense is active;
AFTERGLOW_RAG_HYBRID=offto disable) - WASM whisper engine (
transcribe --apply) —@xenova/transformersoptionalDependency, no native build;AFTERGLOW_WHISPER_ENGINE=auto(WASM→native) · model auto-downloads on first use · whisper.cpp binary tier as fallback - whisper model management (
transcribe --download/--list-models+ auto-resolution) - PII masking + at-rest encryption — transcript-only email/phone/RRN/card/token masking (
AFTERGLOW_PII_REDACT=1) + AES-256-GCM (AFTERGLOW_ENCRYPTION_KEY); RAG decrypts transparently so search still works - Auto question-suggestion on new interview —
interview startembeds a 4-signal gap analysis and asks "proceed with these questions?" (disable withsuggest=false) - Missing-argument elicitation — omit a required arg and the tool returns numbered choices (+ "type your own") with
[필수]/[선택]tags; candidates are dynamic (existing slugs · action enums · session ids · pending question ids) - Choose how interviews run — real-time (
mode=sync, recordanswerlive) vs file-based (mode=async):export-sheetwrites a self-contained HTML form by default (checkbox UI: answered / declined / N/A / meaningless ·localStorageauto-save — closing and reopening restores progress) or--format mdfor Markdown. The interviewee hits "Download answers (JSON)" →import-answersauto-detects JSON/MD and ingests it - Close the audit loop / record-answer —
correct --action record-answerarchives the answer Claude composed (question · answer · confidence · sources) intoanswers.log;correct listshows it alongside corrections. Previouslyaskonly returned a context bundle and the actual answer never came back into Afterglow - Mutator per-tool ACL —
correct·edit·recalibrate apply·version(rollback·tag·snapshot) now honour the agent's access policy via acallerarg (required when policy is deny-default) - persona.bio truncation fix — when bio overflows 20k, the oldest absorbed interview blocks are dropped first (previously the new block was silently discarded).
renderInterviewBlocknow also includes theskipped(N/A / meaningless) section so future asks know which areas the interviewee marked off-limits - Mutator ACL extended to all (v0.10) —
handoff·interviewmutating actions ·archive(archive/restore) ·gc(apply + specific slug) now also honourcaller+ access policy. Read-only and model-management calls still pass - Ed25519 bundle-manifest signing (v0.10, P3 minimum) —
exportsigns thebundleHashwith a local keypair and embeds the public key + signer name in the manifest (TOFU).verify/importcheck it — a tampered signature is refused at import (force withacceptBrokenChain). Pre-v0.10 unsigned bundles still go through as "unsigned" for back-compat - Data-subject export (v0.10, P4 minimum) —
correct --action data-subject-exportreturns a single JSON dump of persona · consent · interviews · history · corrections · recorded answers · provenance · audit counts. Read-only (passes even under deny policy) - Status refinements (v0.10) — per-agent staleness (
lastActivityAt/staleDays, 30 d+ flagged with 🕰) and dense RAG health counter (denseFailures/denseLastError) - HTML answer-sheet cross-device (v0.10) — "Save progress (file)" / "Load progress" buttons let the interviewee carry an in-progress sheet across machines
- Slash commands
/mcp__afterglow__<name>— 24 MCP prompts (one per tool) — typeafterglow:→ Tab to invoke - 323 vitest + extended stdio handshake (covers all 26 tools + prompts) — incl. a multi-angle anti-hallucination grounding suite
- Published on npm (
@daeseoksong/afterglow-mcp) - Hands-on Jupyter notebook (
docs/afterglow-hands-on.ipynb) — beginner-friendly walk-through of every feature
Next
- per-tool ACL · scheduled retention · Web companion
- Bulk encrypt/decrypt tool extending to knowledge files · external STT (Tier 2) adapter
🤝 Contributing
# Fork, then
git clone https://github.com/<you>/Afterglow.git
cd Afterglow
# Frontend changes
npm install
npm run dev
# Server changes
cd server && npm install && npm test
PR checklist:
- Root:
npm run typecheck && npm run lint && npm run build - Server:
npm run test:all - Group commits by feature / phase
📜 License
GitHub · npm · Issues · Server details
Made with ✦ for teammates who have left, but who we still carry with us.