CHANGELOG
What's new in Power Claude.
Shipped weekly. Tagged by impact — throughput, stability, analytics, UX.
Follow every token down the whole orchestration tree — and find the leak.
When Claude fans work out to sub-agents that spawn their own sub-agents, a session's token cost vanishes into a tree you can't see. The Orchestration Token Tree reconstructs that whole tree from your local session transcript — every prompt, every recursively nested sub-agent — and rolls input / output / cache-read / cache-write tokens and an estimated dollar cost up each branch.
A flame chart in the dashboard makes the heavy branches obvious at a glance: bar width is proportional to tokens, so the one runaway agent eating half your context is the widest bar on the row. That is how you find the leak — the retry storm, the over-eager fan-out, the agent re-reading the same file ten times — and cut it. It runs in the terminal too (pc tree, with a cost view and the heaviest-branch list), updates live as a session runs, and exports the exact same tree to any OpenTelemetry backend (Grafana Tempo, Honeycomb, SigNoz…) with pc tree --otlp — no API key, all reconstructed from data already on your disk.
whole-session tree — every prompt + recursively nested sub-agent · input/output/cache tokens + ≈$ rolled up per branch · flame chart finds the heavy branches at a glance · live as the session runs · pc tree CLI (cost view + heaviest branches) · OTLP export to any OpenTelemetry backend · reconstructed from local transcripts, no API key
Up and running in five minutes — in the editor or the terminal.
Power Claude now has a real front door. The first time the extension loads it opens a native VS Code “Get Started” walkthrough that takes you from zero to your first automatic rotation — install, activate your license, add your Claude account, start rotating, and the everyday safety controls — checking each step off as you go. Prefer the terminal? pc onboard prints the exact same quickstart, and pc install now sends you straight to it.
The point is that it is one source, everywhere: the in-editor walkthrough, the welcome panel, the pc onboard CLI, and the online docs at /docs/power-claude all render from a single content definition, so the steps can never drift apart — and every step links back to the docs so you can return any time. Free for everyone, no license required.
native VS Code “Get Started” walkthrough on first run · pc onboard / welcome / getting-started in the terminal · empty-panel Get Started button + Command Palette commands · one shared source for editor + CLI + docs · /docs/power-claude
A resettable token meter for every client you bill.
If you bill clients by usage, the lifetime and 7-day totals can't tell you what one project actually cost. The Token Trip Odometer fixes that with a car-style trip meter for tokens.
Start a named trip per client or project and it counts tokens from that moment — independent of every other counter. Reset zeroes it for a fresh billing window, Close freezes the total, and you can run several at once and resume a closed one later. Each trip shows input / output / total tokens, elapsed time, and an estimated dollar value for invoicing reference. It costs nothing to run — the odometer is derived from the daily-token ledger Power Claude already keeps, so there is zero added overhead on the request path.
named trip per client/project · counts from reset · reset / close / resume · tokens + ≈$ estimate per trip · survives editor reloads · zero proxy overhead
It finishes the safe work on its own — and asks you only for the parts that are actually yours.
Power Claude already resumed Claude when it stopped on an obvious next step. This release makes that decision smart, local, and honest.
A tier classifier now reads every stop on your machine first: a clear continuation ("Now let me wire X", "I'll finish the remaining steps") resumes immediately with no API call at all; a genuine stop — a real question, a "done — standing by" wrap-up, an open approval gate, "what would you like to work on?" — stays parked. Across a large sample of real stops, only about 4% ever needed a model.
The breakthrough is what happens on a mixed stop, where some of the remaining work is safe and some needs you. Instead of parking everything or charging ahead, Auto-Resume does item-level triage: it finishes the in-scope items now and defers only the ones that are yours — a commit, a deferred decision, another session's files — surfacing those when you return. You get far more work done per session while still making every call that matters.
It also works on subscription plans (the triage step uses the same connection the agent already uses — no separate API key), every auto-resumed turn carries a ↻ marker that states why it continued, and Auto-Resume Audit opens a report of every session and turn that was resumed, parked, or escalated. Tune how eager it is with a conservative / balanced / aggressive dial, and prove behavior on your own history with pc calibrate-nudges before you change it.
decides locally first — ~96% of stops need no model · item-level triage: does the safe subset, defers the rest to you · works on subscription plans (no API key) · ↻ in-chat "why" marker + Auto-Resume Audit · conservative/balanced/aggressive dial · pc calibrate-nudges
Find the session by what was said inside it — not by guessing at titles.
When you keep a dozen Claude sessions open, finding the one that discussed a particular thing means clicking through tab after tab — because the title almost never matches what you actually remember, which is something said inside the conversation. Search Open Session Tabs searches the content instead.
Type a phrase you remember and Power Claude streams each session's transcript and matches it against the real conversation — your messages, Claude's replies, the tool calls and their output — then ranks the hits and shows you the matching turn with your query highlighted. (Power Claude's existing session search only ever looked at titles, first prompts, and file paths; this reads what was actually said.)
There are three ways in: a 🔎 Search Sessions button in the status bar that opens it in one click, a keybindable QuickPick (Ctrl/Cmd+Alt+Shift+S) that searches the tabs you have open right now, and a content-search box at the top of the dashboard Sessions list with a toggle to widen from open tabs to every indexed session. Find the match, click it, and that session's Claude tab jumps to the front. It's also in the CLI — pc session search <query> — running the same engine, so the editor and terminal always agree.
searches conversation CONTENT, not just titles · matched turn previewed with your query highlighted · QuickPick (Ctrl/Cmd+Alt+Shift+S) + dashboard box · open-tabs by default, toggle to all sessions · one click reveals the tab · pc session search parity
Pre-summarizing the conversation in the background — so the compaction freeze can be skipped.
Claude Code freezes your session for minutes when the context window fills and it has to summarize the whole conversation inline. Warm Compacting is how Power Claude removes that wait: while your session is still approaching the limit, it pre-summarizes the earlier ~80% of the conversation in the background — kept sidecar-only, never in your live context and never touching the transcript.
This first release ships in a calibration phase, on by default: it builds that pre-summary and measures the exact token reduction it would deliver on your real sessions — without altering a single request yet. That measurement is how we prove the approach skips the freeze before it ever changes what Claude sees. The live step — substituting the pre-built compact near the wall (recent tail always kept verbatim) so the window never crosses the freeze threshold — turns on next, once the calibration data confirms it, and is available today as an opt-in via powerClaude.warmCompacting.applyRewrite.
No regression by design: in calibration it changes nothing; with the live step on, any unsafe case (an edited session, an unsafe cut point) falls back to normal compaction and the recent tail is always kept verbatim.
And you can watch it: a brand-stamped ⚡ Warm Compact indicator shows in the Claude Code statusline and the dashboard, and pc warm-compact reports the projected token drop any time.
pre-summarizes the early ~80% in the background · sidecar-only — never in your live context, never touches the transcript · v1 CALIBRATES (measures the would-be token drop, alters nothing) · live splice opt-in today via applyRewrite, default-on next release · keeps the recent tail verbatim · no regression · ⚡ statusline chip + pc warm-compact
The pool view now shows how loaded your whole pool is — not one account mistaken for the total.
The Account & Usage popup used to headline the best (least-used) account in your rotation pool. With a fresh reserve account at 0% sitting above accounts at 80%+, that read as broken data. It now leads with a true pool utilization figure — a capacity-weighted average across every account, each weighted by its plan-tier token capacity (Pro / Max-5x / Max-20x), so a Pro account and a Max-20x account are never averaged as equals. The lowest-utilization account you route to next is still shown, now clearly labelled Most-available with its name, so 0% there reads as ready headroom rather than a pool total. The same number appears in pc status, so the popup and CLI always agree.
capacity-weighted pool average · weighted by plan-tier token capacity · most-available account shown and named separately · busiest account called out · pc status parity
Open twenty sessions — and never click through them to load each one.
Power Claude users keep many Claude Code sessions open at once — but Claude Code only loads a session's conversation when its tab is first revealed. Reopen a window with twenty sessions and you pay a tax: click every tab, wait for each to render, just to get back to where you were. Tab Warming removes that tax.
The moment VS Code finishes restoring, Power Claude reveals each open session once — using Claude Code's own open command, which reveals the existing tab (never a duplicate) — to force the conversation to render, then puts you right back on the tab you were on. Click any tab afterward and it is instant.
As unobtrusive as the platform allows. Loading a webview means revealing it — VS Code offers no way to load a hidden tab — so the sweep is briefly visible. But it runs once, only after you've been idle a moment, speeds up adaptively when your machine keeps up, and aborts and restores your tab the instant you touch anything. On by default — reopen your windows and everything is just ready.
reveals each open session once after restore to load it · then returns you to your tab · idle-gated + adaptive + aborts on your activity · click any tab and it is instant · on by default
Know exactly why your session is paused — and trust it to resume itself.
When every Claude account in your pool is rate-limited at once, Power Claude doesn't give up: its proxy keeps each session's retry loop alive (short synthetic-429 waits) so a session can auto-resume the instant a window clears. The problem was that Claude Code only showed its generic "thinking" spinner — no reason, no ETA — so a paused session looked exactly like a hung one. Power Claude now surfaces the real state in one consistent line across four surfaces: a VS Code status-bar banner, the Claude Code statusline, a one-time onset toast, and pc status — reading, for example, All Claude accounts rate-limited · auto-resume ~32m (22:30).
The line names the actual condition (the whole pool is rate-limited, not a crash) and the soonest recovery time across the pool, then clears itself automatically the moment an account frees up — nothing to dismiss.
It pairs with what Power Claude already does: it resumes the session on its own, even hours later. Even when the quota is fully spent and the long 5-hour or weekly window won't clear for hours, Power Claude keeps the retry loop alive and picks the work back up the instant that window opens. You never babysit it or manually restart — walk away and it continues.
one message on 4 surfaces (status-bar banner / statusline / onset toast / pc status) · names the condition + soonest pool reset · auto-clears on recovery · resumes itself even hours later when the 5h/weekly window opens — no babysitting
Get more out of the tokens you already pay for.
Multi-account rotation gets you more Claude for less money. The new Token Saver makes sure you waste less of it. Power Claude can now trim the low-signal noise out of what Claude reads — the bulky tool results in every request (command output, build logs, diffs, JSON dumps) — before they reach the model, so leaner context leaves more room for real work and you hit context-limit walls less often.
It is reversible and audited: lossless where it can be (JSON is whitespace-minified), and it never fabricates — only low-signal noise is dropped or collapsed, while the lines you actually need (errors, stack frames, diff changes) are always preserved. On the heaviest command output the savings reach up to 90%; a typical block trims less, so rather than quote an average we show your real, live savings on your own traffic — every token counted, not estimated.
You decide how far it goes: Conservative / Standard / Aggressive, with per-kind toggles for command output, logs, diffs, and JSON. It is on by default — dial it down or turn it off in Settings.
And you can watch it pay off: the Power Mode dashboard card now shows your realized savings (today / 7 days / lifetime) alongside a projected monthly run-rate, each with a ≈$ estimate — and it stays visible even when the proxy is idle.
trims tool_result noise (command output / logs / diffs / JSON) before it reaches the model · reversible + audited · high-signal lines always preserved · up to 90% on the most repetitive output · live measured savings on your own traffic · Conservative/Standard/Aggressive + per-kind toggles · on by default
Run ten VS Code windows at once — your settings and rotation state never corrupt.
Power Claude is built for developers who keep many VS Code windows open against one Claude account pool. Every window read-modify-writes the same shared files — ~/.claude/settings.json and the live account-rotation state — so without coordination they race: two windows flip the active account at the same instant and one update vanishes, or a half-written settings file lands corrupt. That is classic last-write-wins, and at ten concurrent windows it is not rare.
Power Claude now serializes every write to that shared state behind an atomic cross-process lock (O_CREAT | O_EXCL): exactly one window writes at a time and the rest queue for a few milliseconds, so nothing is torn or lost. If a window crashes mid-write, stale-lock recovery detects the dead holder and reclaims the lock automatically — no wedged state, no manual cleanup. It is a tiny synchronous primitive with zero added bundle dependencies.
This is the companion to the double-launch guard: that guard protects a single session transcript from two drivers, while this protects the shared state files from concurrent writers. Together they make heavy multi-window use safe.
atomic O_CREAT|O_EXCL lock · serializes settings.json + rotation-state writes · stale-lock recovery (PID + start-time) · zero added dependencies · companion to the double-launch guard
See exactly what Claude's /rewind can't undo — before you lean on it.
Claude Code's built-in /rewind is a safety net with a quiet hole: it only tracks files Claude edited through its own editing tools. Anything changed by a bash command (mv, sed -i, a build step), edited by hand, or touched by another session is invisible to it — so a rewind can leave those changes stranded without telling you.
Power Claude now closes the loop. For any session it reports the checkpoint layers — Claude's native /rewind history, plus Power Claude's own durable session-snapshot ref — and computes the bash-gap: the exact list of files that changed since the session started but that native /rewind can't restore. Power Claude's snapshot can restore them, so nothing is silently unrecoverable.
Today it runs on the command line: pc session checkpoints <id> and pc session bash-gap <id>.
L0 native /rewind history + L1 durable snapshot ref · bash-gap = changed-since-start minus Claude-tool-edits · names files /rewind can't restore · CLI: pc session checkpoints / bash-gap · shared consumer-blind core
Power Claude keeps going when the AI says it will — then stops.
Auto-Resume gained a detector for the most common avoidable stop: the assistant announces it will keep working — "I'll continue the cutover automatically once the install finishes", "let me wrap this up" — and then ends its turn anyway, leaving the session parked until you type continue. The new promised-continuation rule catches that and picks the work back up.
The remaining-work detector got broader too: it now recognises "remaining sequence / plan / checklist", "the rest of the …", and "left to do", not just the literal "remaining tasks". Both ship on by default and are confirmed by a fast Haiku check before resuming, so genuine stopping points, finished work, and anything actually waiting on you are left alone — the hard-stop vetoes (approval gates, destructive actions, "once you've verified") still apply.
One reliability fix rides along: new nudge rules now merge into an existing install instead of being written only once, so this detector — and every future one — reaches machines set up by an earlier version, without disturbing your own on/off customisations.
New promised-continuation rule · broadened remaining-work vocab (sequence / plan / checklist / the rest / left to do) · both on by default · Haiku-confirmed · nudge catalog now merges into existing installs (no more seed-once)
Resume an active session and Power Claude warns you — even when it's running in a terminal.
The double-launch guard that stops two processes from driving the same Claude Code session at once now actually fires. It used to depend only on a heartbeat sidecar that a VS Code window writes when it claims a session — so a session you were driving from a terminal (claude --resume) or the pc CLI was invisible, and resuming it again silently raced the transcript. Detection now also scans for a live claude --resume process, so a session being worked on anywhere on the machine is caught. Resuming a busy session pops an explicit modal — Take over / Open anyway / Open read-only — from both the Session Explorer and the dashboard Resume button, and the live-running dot now lights for terminal sessions too.
Everything the editor knows is now on the command line: pc session live lists every session being worked on right now, and pc session recover refuses a session that's already in use unless you pass --force.
Process-aware detection (sidecar + claude --resume scan) · modal: Take over / Open anyway / Open read-only · live dot covers terminal sessions · CLI: pc session live + guarded pc session recover --force
Every dashboard badge, chip, and stat tile now self-explains on hover.
The dashboard had a class of small indicators ("API", "EST", colour dots, the routing strip's request counter, the per-account fairness KPIs, the cooling/healthy metric tiles) where the meaning was implicit — obvious to whoever built the feature, opaque to a new user staring at the panel. We swept every one and attached a verbose markdown tooltip that explains what the indicator represents, what data drives it, and what action (if any) you should take when it changes.
The biggest legibility win is the calibration-source pill on the "Next account available in" card. It used to read API or EST — labels that, in isolation, read like "this is an API account vs an estimate account." That's not what the badge means: it tells you whether the recovery countdown was calibrated from a real Anthropic rate-limit reset header (we know exactly when the window closes) or projected from the published 5h-window heuristic (we're estimating). The pill is now labelled ✓ EXACT / ≈ EST with rich tooltips, and the same helper is reused across the soonest-available card, the upcoming-recovery rows, the rank chip, the status legend, and the fairness KPI tiles. Same wording lives in this changelog and in the in-app tooltips, so the docs don't drift.
Per-element walkthrough, grouped by panel: Reading the Power Claude dashboard — every badge decoded.
Calibration-source pill renamed ✓ EXACT / ≈ EST · routing strip + soonest panel + status legend + pool-count tiles + fairness KPIs all gain verbose data-pc-tip tooltips · tooltip engine unchanged · no settings touched
Your original ask pinned to every session — injected by Power Claude, not written by the AI.
Your Request Banner reads the first message from the session transcript and injects a slim banner into Claude Code's chat DOM — zero context tokens consumed, no model instructions. Power Claude polls its own proxy for the message text and patches it straight into the webview; the bolt badge is tooltip-only and nothing appears as AI output. Built for users running many concurrent sessions: ten windows open, one glance to know which is which.
This release also formalises Power Claude's zero-context-token UI architecture: Your Request Banner, Message Timestamps, Session List Badges, Tab State Icons, Chip Default Off, Preserve Focus Fix, and Conversation Sync all inject at the DOM layer — none add instructions to the model, none consume context window, none appear in the transcript. They work for free, regardless of how large your context already is.
DOM injection · reads JSONL directly · 0 context tokens · bolt badge tooltip · self-heals on update · 7 zero-token features
Sidebar dashboard becomes a hub you act from.
The Overview tab grows a 60-minute rotation handoff lane graph that draws every recent rotation as a curved bezier between account lanes, with state-encoded dots (●/◌/×/◇) so the signal reads independently of colour. The accounts table groups into Needs action / Throttled / Healthy / Idle collapsible buckets, each row carries inline decorator icons (🔔 for recent events — click to filter Events to that account, ! for error-grade events in the last 24h, 🔒 for currently throttled), and clicking a row slides a 320 px drill-down drawer in from the right with a 24-hour activity sparkline and the last 10 rotations involving that account. Events gain a chip-filtered Account column, hover any account row for a quick-action card, ctrl-click two events to open a side-by-side JSON diff modal, right-click any row for state-aware menus (Activate / Re-login / Refresh / Open log / Open claude.ai / Remove). Hand-rolled SVG + CSS using VS Code theme tokens — no chart or UI libraries, no new tabs.
Lane graph · drill-down drawer · chip-filtered events · hover + right-click affordances · state-as-shape glyphs · 7-tab IA unchanged
Five Power Claude features surfaced on the product page.
Five capabilities that already ship in Power Claude now have dedicated tiles on the product and home pages: Auto-Rewake (systemd-run timers wake Claude Code at the soonest reset across your pool, with seven per-rate-limit-type context strings); Auto-Resume (continues your session in a healthy account after a 429 — silent or notified, 60-second snapshots, cross-IDE optional); preserveFocusFix as a standalone tile (the upstream-#32726 patch that stops Claude Code from yanking your cursor on every reveal); LLM Usage Recommender (tells you when to upgrade your Claude plan from your own cap-hit data, and when Anthropic served Haiku at Opus prices); and Log Export (sanitized bundle ready for a support ticket — API keys, OAuth tokens, and session secrets redacted before write).
5 new product-page tiles · no shipped behavior changed · marketing-only · pre-existing config keys, now visible
Throughput Visualizer — Feel the Concurrency Math
A new interactive explainer on the home page. Drag a slider from 1 to 20 concurrent Claude Code sessions and watch two bars race: a single account, which serializes the work, against a 5-account Balanced pool, which spreads it. At ten concurrent sessions a single account finishes a round in roughly the time the Balanced pool needs for five — about 5× faster per round. The widget shows the speedup live as you drag, with the underlying math and its assumptions stated in a methodology footnote.
Interactive slider (1–20 sessions) · single vs 5-account Balanced · live speedup readout · methodology footnote
Premium Trial Callouts — Lock Founder Pricing From Anywhere
If you installed Power Claude from the marketplace and started the no-card 7-day trial, you might never have known the website offers a longer deal. Now you do. Free-trial users get a consistent nudge toward the 14-day Premium Trial (card-on-file, $0 today) that locks founder pricing for 2 years: a "Lock founder pricing" button in the extension dashboard, a suppressible banner on `pc license info` / `pc license status`, and the same callout on the marketplace listing. Every surface reads one shared source for the URL and copy so they never drift, and the nudge is gated on trial state — paid and Premium-Trial users never see it.
Dashboard button + CLI banner + marketplace callout · one shared URL/copy source · trial-gated · silence the CLI banner with POWER_CLAUDE_NO_TRIAL_BANNER=1
"Connection Refused" Now Self-Heals — No More Editor Restarts
If you ever saw "API Error: Connection refused" and had to restart VS Code to make Power Claude work again, this is the fix. The rotation proxy used to run as a child of the editor's extension host, so when VS Code or code-server recycled a stale host it killed the whole process group and took the proxy down with it — the port went silent and nothing brought it back until you manually reloaded. The proxy now runs as its own background service in a separate control group, fully decoupled from the editor: it survives window reloads and crashes, restarts within seconds on its own if it ever does crash, honors the emergency-off switch, and never spawns a duplicate. Installing Power Claude now wires this self-healing into your sessions automatically — there is nothing to hand-configure, and re-installing never duplicates the wiring. Hosts without a service manager fall back to the previous behavior automatically.
Proxy decoupled into its own service + cgroup · survives editor restart/crash · auto-restart with backoff · emergency-off honored · self-wires on install (idempotent) · graceful fallback on minimal hosts
Session Recovery (Pro) — Rescue Lost Work
A first-class recovery system for the failure modes that previously meant losing your session entirely. Nine failure scopes (compact-thrash, tool-loop, context-overflow, crashed, rate-limit-stuck, and four refined existing detectors) classify every session as it runs. Six recovery strategies map to each scope — the headline is context-trim-resume, which exports a trimmed summary and opens a fresh Claude session pre-seeded with it. Every recovery is snapshot-protected (one-click undo), per-session history prevents retry loops, and the dollar-cost-vs-tokens-rescuable math is surfaced inline on the Recover button.
9 failure scopes · 6 recovery strategies · git-ref snapshots · 7 new `pc session` CLI subcommands · $X-rescuable telemetry · Pro tier
Session Awareness — Know Every Session's State at a Glance
Every message now carries a timestamp, and a lifecycle status badge shows exactly where each session stands — running, complete, idle, needs-attention, or archived — as an at-a-glance glyph. A live-running overlay marks any session that is actively working right now, and a double-launch guard stops you from spawning a second run over a session that's already going.
Per-message timestamps · lifecycle status badge (🟢 running · ✅ done · 🟡 idle · ⚠️ attention · 📦 archived) · live-running overlay · double-launch guard
Preserve-Focus Fix — Stop Claude Code From Yanking Your Cursor
Power Claude can now patch the installed Claude Code VS Code extension to pass preserveFocus:true on every autonomous reveal/show call — the upstream fix Anthropic has not yet shipped (issue #32726). Eight surgical patches, idempotent and reversible, with self-healing upgrade detection: when an auto-update overwrites the patched file, the patcher re-runs automatically. Per-extension records keyed by absolute path, so Remote / Insiders / native installs are covered independently. Includes the terminal-show sites that fired on every Bash tool call — window stayed put, but the keyboard cursor jumped to the terminal — now silenced. Opt-in by default — Power Claude does not touch another vendor's bytes without explicit consent.
8 surgical patches · covers terminal show on Bash calls · self-healing upgrade detection · per-install records · idempotent + reversible · opt-in
Auto-Open Tabs Disabled by Default
Power Claude no longer opens Claude Code session tabs automatically. The auto-open paths shipped buggy — tabs appeared in the wrong workspace context, and the rate-limit recovery loop kept rewriting `pending-recovery.json` and re-popping tabs. All three auto-open code paths (startup, rate-limit recovery, watchdog stall recovery) are now gated behind a single dev-mode setting and off in production. Manual resume from the dashboard, Session Explorer, and command palette is unchanged.
Single master gate (`powerClaude.devMode.autoOpenSessions`, default OFF) · 3 subordinate defaults flipped off · detection + UI signals unchanged · manual resume unchanged
1.0 — General Availability
The first stable public release. Multi-window safety lands across the board: proxy refcounting keeps a single shared proxy alive across every VS Code window and tears it down only when the last one closes, while a cross-window lock serializes settings writes so concurrent windows can't clobber each other. A guided first-run setup, a single version source of truth, and a one-command “Uninstall Everything” round out a clean install-to-removal lifecycle.
Multi-window proxy refcounting · cross-window settings lock · guided first-run setup · Uninstall Everything command · 9 CLI entry points
Proxy Core, Parallel Mode & Subscription Enforcement
The biggest release since launch. A zero-dependency Node proxy sits between Claude Code and api.anthropic.com, reads unified rate-limit headers on every response, and pre-emptively rotates accounts before any limit is hit. Parallel account dispatch multiplies effective throughput across every account you own.
Zero-dep HTTP proxy · pre-emptive rotation · parallel dispatch (5 MAX ≈ 10M Opus ITPM) · Ed25519-signed JWT licensing
Watchdog Hotfix
Fixes a false-positive in the session-resilience watchdog where healthy idle sessions were being flagged as stuck two minutes after every user message. Auto-resume is now silent — no QuickPick prompt interrupts the automation.
Heartbeat-stale suppression on end_turn states · silent auto-resume · new sessionResumeWithPrompt command
Continuous Automation — Session-Resilience Watchdog
Polls every 30 seconds. Detects four classes of session death: heartbeat-stale, tool-truncation, ended-with-pending-tasks, and rate-limit. Surfaces a one-click resume QuickPick with a reason-aware prompt.
30-second polling · 4 detection rules · reason-aware resume prompts
Session Explorer — Searchable Session History
A new activity-bar view lists every Claude Code session in the workspace, grouped by Pinned / Today / Yesterday / This week / Older / Archived. Each session is bidirectionally linked to the files the agent read, wrote, or edited — extracted directly from the JSONL transcript with no extra instrumentation. Multi-tab Session Detail (Files, Timeline, Search, Notes, Metadata), a bipartite force-directed Session Graph, and a pre-commit review webview with per-file Stage / Unstage / Revert actions round it out.
Activity-bar view · per-session Files Touched sidebar · Session Detail multi-tab · Session Graph · pre-commit review · 12 auto-tags · `Ctrl+Alt+S` search
Insight Graph — Daily Story of Your Workflow
A timeline view of every Claude Code session you ran — rotations, recoveries, watchdog interventions, model swaps. Click any moment for the full event chain. Filters by tier, by failure, by elapsed cost.
Daily timeline · per-event drill-down · cost overlay · failure-mode filters
Pre-emptive Rotation v2 — Token-Window Awareness
The rotation engine now reads Anthropic's 5-hour and 7-day token-window headers and rotates before either approaches its threshold. Eliminates the late-window stalls that v0.9 still occasionally surfaced.
Dual-window awareness · header-driven thresholds · zero late-window stalls (measured n=120 days)
Watchdog v1 — Heartbeat Detection
First version of the session-resilience watchdog. Polls every 30 seconds, detects heartbeat-stale and tool-truncation states, prompts a manual resume. Foundation for v0.13's reason-aware auto-resume.
30-second polling · 2 detection rules (heartbeat-stale, tool-truncation) · manual resume
Session Trail — Per-File Edit Provenance
Editor gutter shows which session touched which line. Hover for the prompt that produced the change. Backed by the same event log that powers the Insight Graph.
Gutter marks · prompt-on-hover · per-session colorization
Parallel Account Dispatch (Preview)
Opt-in preview of routing concurrent requests to different accounts simultaneously. Reduces wall-clock time on multi-step automations. Enabled via experimental flag; GA in v0.15.0.
Concurrent-dispatch preview · per-account quota tracking · feature-flag opt-in
Telemetry Dashboard — First Pass
Read-only dashboard inside VS Code: rotations per day, account utilization, watchdog interventions, session length. All metrics derived from the local event log — none of it is sent to us.
Local-only metrics · zero telemetry egress · 8 panels (rotations, utilization, watchdog, length, …)
Auto-Nudge — Stop Claude Asking "Should I Continue?"
Detects the "should I continue?" confirmation pattern and auto-responds with a configured nudge. Three modes: aggressive, balanced, conservative. Default balanced. Optional Haiku fallback for ambiguous stops.
3 modes · pattern + optional LLM fallback · configurable nudge text
Status Bar Item — At-a-Glance State
New status bar item shows current account, tier, remaining budget, and rotation health. Hover expands to a detail card. Click jumps to the dashboard panel.
Live state in status bar · hover detail · click-through to dashboard
Multi-Account Pool — Round-Robin Rotation
Configure multiple Claude accounts and the extension rotates between them as quotas drain. Replaces the manual paste-key-and-restart cycle that earlier preview builds required.
Pool size unbounded · round-robin selection · automatic on-quota-exhaustion rotation
Ed25519-Signed Activation Tokens
Paste your subscription key once; the server validates it and issues an Ed25519-signed activation token the extension verifies locally. After the first activation, every launch is an offline cryptographic check — no network round-trip, no spinner.
One-time online activation · offline verification thereafter · token latency < 50ms
First Public Build
Marketplace listing goes live. Core feature: detect rate-limit responses from Claude Code and prompt the user to swap accounts. No rotation engine yet — that lands in v0.3.0.
Marketplace launch · rate-limit detection · manual-swap UX
Closed Preview — Internal Dogfooding
First end-to-end build shared with a handful of design partners. No license enforcement, no rotation, no dashboard — just rate-limit detection plumbed into VS Code. The kernel of what shipped at v0.1.0.
Design-partner preview · detect-only build · no UI surface beyond a status bar dot
First Commit
Project bootstrap. Repo scaffold, VS Code extension manifest, the first lines of the rotation contract that every later release builds on. The original sin: an afternoon lost to a rate-limit hit while trying to ship something else.
Repo bootstrap · extension manifest · the idea
Stop losing afternoons to rate limits.
Premium Trial: card-on-file, $0 today — locks today's founder rate at $12/mo for 2 years. Free: no card, 7-day trial inside the extension.
Neural-LLM is an independent software company. We are not Anthropic, and we are not affiliated with Anthropic. Power Claude is third-party tooling. Claude™ is a trademark of Anthropic, PBC.