On the record
Every problem Valo solves — with receipts.
These are not marketing claims. Each one below is a real, publicly filed issue on Claude Code’s and Codex’s own issue trackers — the same walls hundreds of developers keep hitting. Valo exists because we hit them too.
94 documented issues · 66 still unresolved · across both engines Valo runs (Claude Code and Codex). Every one links to the source — verify it yourself.
Your work just vanishes
Close the wrong window, hit a context limit, or install an update — and the whole conversation is gone. Resume brings back a fraction, or nothing. It happens on both engines.
What Valo does · Forever History + self-repair
Valo writes every chat to durable storage the moment it happens, and repairs broken sessions on its own. Chats survive restarts, reboots, and reinstalls — forever, and searchable by meaning, not just exact wording. The same search covers your code, your docs and every past conversation in one query, and it runs about twice as fast as the engine’s own file tools — measured on video, not asserted. Watch it run side by side →
The receipts · 21 issues, 10 unresolved
Claude Code
- #23800 Chat history loss after restart, /resume recovers only 20-40% — Closed, never fixed
- #27349 Desktop upgrade erases local chat history every time — Fixed
- #24172 CRITICAL: Conversations disappear when closing VSCode — Open
- #33130 Desktop chats lost after restart — Closed, never fixed
- #29017 Conversation history lost in VSCode extension — Open
- #42890 ALL session histories truncated (~1967 sessions) — Closed, never fixed
- #31105 Remote Control: history lost when connection drops — Fixed
- #9606 Conversation gone after exit and relaunch — Fixed
- #28180 Silent crash — entire conversation disappeared — Closed, never fixed
- #43303 Closed terminal window, all history gone — Fixed
- #26123 /resume broken since v2.1.31 — 3 root causes — Fixed
- #24304 Conversation history missing on resume — broken parent chains — Open
- #3138 --resume fails after context limits — complete session loss — Open
- #25920 --resume fails for sessions with first message > ~15KB — Fixed
- #13480 Oversized image breaks conversation permanently — Fixed
Codex
- #26849 Too easy to lose a session and all its content — Open
- #30107 Previous Codex chats not restored after reinstalling the VS Code extension — Open
- #3088 Restarting extension deletes chat history — Fixed
- #2935 History of local chats in VS Code is not saved — Fixed
- #8256 codex resume fails with "stream disconnected before completion" — Fixed
- #16278 Remote compact task timeout — session becomes unrecoverable — Fixed
Hit stop, and a blank stranger takes over
This is the one almost nobody names correctly. Interrupt a turn and the CLI is killed before it writes what just happened to disk — so the next thing you type resumes a session that is now empty or broken, and a completely fresh AI spawns in its place with none of your context. It is trained to be helpful, so it reads your recent commits and open files and confidently continues as if nothing happened. You do notice something is off — it "hallucinates," it "stops listening," it does the opposite of what you just asked. You just blame the AI, never knowing a silent reset underneath you was the real cause.
What Valo does · Pause + session auto-heal
Valo treats a stop as sacred. It captures the turns the CLI never flushed to disk, and if the session cannot cleanly resume, it rebuilds your entire conversation into the new one — so you keep talking to the AI that actually has your context, never a blank stranger guessing from your git history. Pause halts cleanly instead of hard-killing to prevent it; silent auto-heal repairs it when it slips through.
The receipts · 22 issues, 16 unresolved
The twist: the reset is invisible, but the damage is loud. So much of what people call "the AI hallucinating," "it ignored me," or "it got dumber" is really this — reported constantly, under the wrong name. The issues below are just the interrupts that error out loudly enough to file a bug.
Claude Code
- #6836 150+ reports: interrupting mid-tool-call leaves the conversation in a broken state — Open
- #3003 Interrupting during a tool call permanently corrupts the conversation — Closed, never fixed
- #12162 Interrupting during tool use breaks the entire conversation — Closed, never fixed
- #21041 Orphaned tool calls corrupt the conversation after an interrupt — Closed, never fixed
- #14229 Confirms it stopped — then silently keeps running and spending tokens — Closed, never fixed
- #21895 [CRITICAL] Interrupt button does not stop processing — Fixed
- #49309 Escape & Ctrl+C do not interrupt prompts on macOS — Fixed
- #3455 Escape & Ctrl+C show feedback but don't stop the agent — Fixed
- #19340 Claude won't stop — keeps going, ignores every command — Fixed
- #32679 Stop keeps going — wasting credits until session is burned — Closed, never fixed
- #11803 VS Code: Esc / stop button ignored most of the time — Closed, never fixed
- #23350 Abort state persists within session after manual Escape — Closed, never fixed
- #7455 Freezing at session resume after Ctrl+C interruption — Closed, never fixed
- #34476 No way to stop agents without killing the session — Closed, never fixed
Codex
- #32653 Desktop crashes the entire application on a missing tool-call result — Open
- #24571 Desktop stuck thinking after an interrupted/resumed turn — Open
- #8479 Parallel tool calls cause a "tool_call_id missing response" error — Closed, never fixed
- #7275 Interrupted history is not protocol-compliant — breaks on strict providers — Closed, never fixed
- #12382 CLI resume stuck on "Working" after an unclean disconnect — Open
- #9428 Moving the Codex panel cancels the active chat and loses progress — Closed, never fixed
- #11852 Resume after reconnecting shows a stale "working" state — Fixed
- #3208 Resume a CLI session after premature Ctrl+C exit — Fixed
It deletes things you never approved
Agents have wiped thousands of untracked files, emptied the trash, and destroyed documents with no explicit instruction — and the reports keep coming.
What Valo does · Safety Guards
Valo’s Safety Guards inspect dangerous actions before they run and stop the ones that cross a line — destructive shell commands, reads of secret files like .env, writes into protected folders, and freshly published packages that are how supply-chain attacks arrive. You can add your own in plain language.
The receipts · 6 issues, 5 unresolved
- #23913 Agent deleted 2,229 untracked files — Closed, never fixed
- #30988 Claude batch-deleted 50 files uninstructed — Closed, never fixed
- #24210 Destructive deletion without explicit instruction — Closed, never fixed
- #32637 Cowork destroyed iCloud-offloaded legal documents — Closed, never fixed
- #46970 "Move to trash" empties entire Trash — Fixed
- #49539 Deleted entries from ~/.git-credentials — Closed, never fixed
Your secrets leak into the chat
Credentials get printed into tool output, committed to GitHub, and read straight out of .env files — with no redaction.
What Valo does · Password vault + Streamer Mode
Valo keeps secrets in an encrypted vault and injects them at run time, so the AI uses them without ever seeing the values. Streamer Mode scrubs them off your screen too.
The receipts · 11 issues, 9 unresolved
- #32523 Claude reveals secret keys in terminal output — Closed, never fixed
- #29910 Built-in secrets management needed — Open
- #11271 Claude exposes env vars when confused — Closed, never fixed
- #20966 Secrets/tokens in tool output, no redaction — Fixed
- #59094 Agents read .env files without redaction — Open
- #44868 .env/.dev.vars exposed via grep/Read — Open
- #34819 Full credential file contents displayed — Closed, never fixed
- #13106 Sensitive tokens in .claude/ not gitignored — Fixed
- #2142 Claude commits API keys to GitHub — Open
- #32733 Secure secrets injection for web Claude Code — Open
- #25053 Mark sensitive env vars to prevent exposure — Closed, never fixed
Permission fatigue
Hundreds of identical approval prompts, allow-rules that get ignored anyway, and a best case of "approve everything for this session" — which just moves the same reflex click to the top of every session.
What Valo does · Per-tool and per-secret Ask?
You decide once per tool, not once per popup. Run in Full trust and the AI just works — no prompt on every read, edit or command. What you pinned by hand still holds: any tool group you set to Ask? keeps asking in every mode, Full trust included, and so does any individual password you mark that way. That is the combination nothing else offers — full speed on the ordinary work, with sending mail or spending an API key still stopping to ask, because you said so and nothing can quietly override it.
The receipts · 9 issues, 5 unresolved
Count your own clicks honestly: you approve a hundred times and revert maybe ten. A prompt you approve on reflex is not a safety check, it is a toll booth — it buys interruption, not protection. Real protection is a Guard that blocks the genuinely destructive before it runs, and work you can always revert. So Valo spends your attention only on the handful of tools that act as you, and gets out of the way everywhere else.
- #32973 "Always allow" option needed — Fixed
- #26796 Auto-approve settings ignored for basic commands — Fixed
- #27333 Edit/Write allow rules don't auto-approve — Fixed
- #25966 Can't auto-approve read-only MCP tools — Closed, never fixed
- #10801 No way to bypass MCP approval in VSCode — Closed, never fixed
- #16561 Compound bash commands not parsed individually — Open
- #55124 Hundreds of identical permission prompts — Open
- #29500 Gap between "allow once" and "always" — Closed, never fixed
- #15772 VSCode ignores ALL permission settings — Fixed
It fills up and gets dumber
The longer you talk, the worse it gets — and right at the limit, the tools meant to compact the conversation fail exactly when you need them most.
What Valo does · Context Trim + orchestrator
Context Trim cleans up the conversation you are already in. It keeps everything that matters, clears out the bulk that does not, and the chat carries on — you never have to start over and re-explain yourself. The AI can do it on its own as things fill up, or you can trim by hand with a button. When a job is genuinely too big for one chat, the orchestrator splits it across several sessions instead of cramming it into a context that is already tired, and Handoff opens a fresh chat carrying the full brief. You can also see exactly what is in there, and hide the tools, skills or messages Claude does not need.
The receipts · 9 issues, 5 unresolved
This is the one that matters most if you are new. Everyone starts out wanting to stay in the chat they are already in — that is where all the context lives, everything you already explained, every preference you already stated. Every other tool answers that with "start a new chat and explain it again." Context Trim is the answer that lets you stay.
- #23047 /compact fails with "too long" at context limit — Fixed
- #26317 Same compact catch-22, forced to /clear — Closed, never fixed
- #25620 Can't manual /compact at context limit (199K tokens) — Closed, never fixed
- #18360 Context blocking too aggressive at ~80% — Fixed
- #34650 No --max-context flag, 1M burns quota 5x faster — Fixed
- #41037 Configurable context window limit needed — Fixed
- #17428 Enhanced /compact with file-backed summaries — Open
- #23463 Subagent results silently overflow context — no truncation — Closed, never fixed
- #16209 Context full after subagents, /compact stops working — Closed, never fixed
It forgets between sessions
Memory that never actually gets applied, instructions dropped after compaction, and auto-memory that corrupts from one session to the next.
What Valo does · Valo MemoryCode
Valo MemoryCode works the way human memory does. A lesson stays dormant until you touch the thing it relates to, then it surfaces on its own — so it never clogs the context the way a file of standing instructions does, and a problem you already solved together does not come back. Write it once, after the hard fix, and it arrives at exactly the moment it matters instead of sitting in a document hoping to be found.
The receipts · 6 issues, 6 unresolved
- #14227 Persistent memory between sessions needed — Closed, never fixed
- #37314 Memory stored but never applied, same mistakes recur — Open
- #31409 Memory instructions forgotten after compaction — Open
- #27298 Layered memory system needed — Open
- #23769 Auto-memory corrupts across sessions — Open
- #2954 Context persistence across sessions — Open
Two sessions break each other
Run more than one session at once and they race on the same credentials file — quietly invalidating each other’s login.
What Valo does · Multi-project workspace
Valo is built to run many chats and many projects at once, each held in one window you flip between like browser tabs — without sessions stepping on each other.
The receipts · 4 issues, 4 unresolved
- #48786 Multiple concurrent CLI sessions cause auth failure on token refresh — Open
- #56339 Multiple CLI sessions race on ~/.claude/.credentials.json — Open
- #54179 Opening multiple sessions invalidates each other's auth token — Open
- #24990 Issue running multiple claude code terminal sessions simultaneously — Open
No way to reach it from your phone
Long-standing requests to start, steer, or just check on a session from a phone — still unaddressed.
What Valo does · On the Go
Valo’s On the Go lets you talk to it from your phone over Telegram — text, voice, photos, even approving permissions, from anywhere.
The receipts · 6 issues, 6 unresolved
- #28420 Start sessions from phone, Siri integration — Open
- #32779 Interact with desktop sessions from mobile — Open
- #25115 Mobile companion with voice input — Open
- #29399 Voice input for Remote Control — Open
- #34255 Remote Control silently drops, no reconnect — Open
- #47926 Resume sessions across devices — Open
None of this is a knock on the engines.
Claude and Codex are the best AI models in the world — Valo runs on them. But the app wrapped around them decides whether your work survives, your secrets stay yours, and a stop button is safe to press. That layer is what Valo is.
See everything Valo does →