A local-first coding harness where every prompt, thought, tool call, approval, and diff is a typed event you can inspect — and a running agent you can retune without stopping it.
Runs locally. Bring your own API key or supported provider subscription.
nerve · live run
Simulated Nerve run: the user prompt, the agent's reasoning, a read tool call with its bounded output, an approval gate for an edit, the applied diff, a verification command, and the final reply.
You
Add a retry budget to the run recovery path.
Recovery lives in run-recovery.ts. A budget needs the retry count and the base delay in one place, so both call sites read one constant.
readpackages/harness/src/run-recovery.ts
Read 214 lines
214 linesShow 7 more items
approvaledit needs reviewapproved
editpackages/harness/src/run-recovery.ts
+12 −3 · retry budget wired into the recovery loop
2 hunksShow 7 more items
bashpnpm test --filter @nervekit/harness -- run-recovery
✓ recovery continues from checkpoint (12 tests)
Done in 3.2s
6 linesShow 7 more items
Retry budget added: three attempts on a 2s exponential base.
Codingsupervised21% /1.0MOpus · high
Ask the local Nerve agent
Follow the signal
Apache-2.0Free & open source
Linux · macOS · Windows 11Cross-platform
127.0.0.1:3747Local daemon by default
API key · subscriptionBring your own provider
~/.nerveYour state, your disk
Retune the run while it runs
Change the model mid-thought.
Model, thinking level, agent mode, permission — change any of them on an active conversation and it applies from the next provider request. The controls below are wired to the stream.
These controls are live — the stream reacts to you.
Model
Thinking
Agent mode
Permission
None of these recreate the conversation. The event already streaming finishes under the old settings; the change applies to the next provider request. Plan mode constrains edits to Nerve's plan storage; read-only blocks mutation, execution, and ordinary network.
nerve · same conversation
Simulated Nerve run that reacts to the controls beside it: changing model, thinking, mode, or permission adds a configuration marker to the stream and the next agent turn follows the new settings.
You
The policy test is flaky — pin it down.
The failure is time-dependent. Freezing the clock before the assertion should make it deterministic.
readexamples/policy-window.test.ts
Read 168 lines
168 linesShow 7 more items
approvaledit needs reviewapproved
editexamples/policy-window.test.ts
+4 −4 · clock frozen before the assertion
1 hunkShow 7 more items
bashdemo test policy-window --repeat 20
19 passed · 1 failed on attempt 14
assertion raced a debounce timer
9 linesShow 7 more items
Deflaked: the clock is frozen before the assertion now.
Codingsupervised21% /1.0MOpus · high
Approvals pause the stream — take the button when it appears
Nothing happens off-screen
A run is not a spinner. It's a stack of typed events.
Prompts, reasoning, tool calls, results, approvals, and recovery are separate events. Each keeps its evidence, and all of it stays in history.
Every layer keeps its evidence.
Tool calls retain the inputs they were given and the bounded output they produced, and both stay inspectable afterwards.
The gate is real.
In supervised mode the run is blocked while a review gate is pending. Nothing below the gate executes until you answer it.
Thinking is honest.
Reasoning appears when a provider returns it. Nerve never reconstructs private reasoning a provider withholds.
Locating the recovery entry point before editing anything.
reasoning
Reasoning, when the provider returns it.
read
packages/harness/src/run-recovery.ts
call
A tool call keeps the arguments it was given.
result
214 lines · bounded output
done
And the output it produced.
approval
edit needs review
waiting
Supervised mode stops here until you answer.
edit
run-recovery.ts · +12 −3
applied
The diff is reviewable before it ships.
Nerve
Retry budget added: three attempts, 2s exponential base.
reply
And the whole run stays in history.
Authority, not vibes
Decide what the agent may do — before it does it.
Nerve enforces permission in the harness policy layer, not just in the system prompt. Change the level and the tool policy changes with it.
Permission level
tool policyevaluated in the harness
read · grep · find · lsInspect the workspaceallowedallowedauto-approval can be switched offallowedallowedallowedauto-approval can be switched offallowed
edit · writeChange filesblockedneeds approvalallowedblockedneeds approvalplan storage only — elsewhere is blockedallowedplan storage only — elsewhere is blocked
bashRun commandsallowedknown read-only commands onlyneeds approvalknown read-only commands auto-approveallowedallowedguarded read-only commands onlyneeds approvalblacklist guard must pass firstallowedblacklist guard must pass first
python_execRun Pythonblockedneeds approvalallowedblockedneeds approvalallowedwrites restricted to artifact storage
web_search · web_fetchReach the networkblockedneeds approvalallowedblockedneeds approvalallowed
task_start · task_controlSupervise processesblockedtask_status and task_logs stay allowedneeds approvaltask_status and task_logs auto-approveallowedblockedtask_status and task_logs stay allowedblockedtask_status and task_logs stay allowedblockedtask_status and task_logs stay allowed
exploreLaunch read-only sub-agentsblockedneeds approvalallowedblockedneeds approvalallowedchildren are forced read-only
Bash and Python restrictions are policy guardrails, not an operating-system sandbox.
The harness stays out of the way
Your model, at full intelligence.
No walls of instructions wrapping every request. A concise base prompt, a small polished tool surface, and constraints enforced in the policy layer — so the model's capability reaches your code intact.
You are an expert 10x engineer persona…ALWAYS begin every reply with a summary…NEVER use the edit tool before reflecting…Rule 14: apologize when tests fail…Respond with enthusiasm and emojis…Step 1 of 12: restate the user's intent…You must think step by step at all times…IMPORTANT: IMPORTANT: IMPORTANT:Refuse to answer without a plan document…Wrap every answer in exactly three parts…As a large language model, you should…Consider seventeen edge cases first…Always praise the user's question…Do not be creative unless instructed…Rule 63: end with a motivational quote…Insert disclaimer before code blocks…You have the personality of a helpful…Re-read these instructions every turn…NEVER admit uncertainty…Use the word 'certainly' frequently…Follow the 40-point style rubric below…Before any action, write a haiku…Summarize this prompt back to me…ALWAYS defer to the style council…Rule 108: never mention rule 108…Prepend every path with a compliment…Think harder. Then think even harder…Your name is Sparkles. Stay in character…Output JSON, but also prose, but also…Escalate all decisions to the persona…Count your tokens out loud…This prompt is 12,000 words. Memorize it…
“Work in the current project, use tools safely, and keep responses concise.”
nerve-system-prompt.ts · the default base promptreadgrepfindlseditwritebashpython_execweb_searchweb_fetchtask_startexploreask_usertodos_set
A concise base prompt
One working sentence by default. Tool rules, mode constraints, and project context are appended only when the run actually uses them.
A small, polished tool surface
A compact set of workspace tools with bounded output, so the model spends its tokens on your problem instead of on ceremony.
Enforcement, not prose
Plan mode keeps research available and constrains edits to plan storage — in the policy layer, not in paragraphs of instructions.
One workbench, your whole stack
Open the parent directory. Work across every repo.
Point Nerve at the workspace that holds your API, web app, and mobile app. It discovers the repositories beneath it and keeps Git, diffs, and pull requests for all of them in one shell.
Repositories are discovered beneath the workspace root — up to two levels deep — and switchable in one Git panel.
Stage, commit, push, pull, and sync per repository from the same shell as the conversation.
Pull requests live beside each repo: read the summary, watch checks, merge when the state permits.
Desktop or browser notifications with independent sounds for questions, plan reviews, approvals, completions, and failures.
Conversation history
Every run is a graph you can walk.
The whole conversation is a navigable graph of messages, tool calls, and branches. Zoom into any node to read its arguments and result, then jump to that point or fork a new branch from it.
Long runs collapse into step groups so the graph stays readable
Navigating changes the active branch; earlier lineage is preserved
Optional summaries when you jump are generated locally
