aicippy

AiCippy

An autonomous development agent that streams from Amazon Bedrock, on the model its backend names. It inspects, reasons, plans, edits, runs, debugs, manually verifies, deploys, remembers, and reports real evidence from your development environment — from a terminal, a TUI, or a browser side panel, all driving the same runtime.

pip install aicippy
aicippy task run "add retry handling to the payment webhook and verify it manually"

---

One central runtime with explicit provider choice

Everything that reaches the model goes through a single path:

The CLI selects a provider once per session and never switches it silently. The WebSocket Lambda and browser extension use the main Responses route from the deployment-generated model.runtime.json.

Documentation

Quick start

# 1. AWS credentials that can call bedrock:InvokeModelWithResponseStream
export AWS_PROFILE=your-profile

# 2. Check the live route end to end
aicippy doctor

# 3. Run an autonomous task
aicippy task run "repair the broken auth flow"

# 4. Or dispatch the specialist fleet for larger work
aicippy task run --multi-agent "migrate the billing service to the new API"

# 5. Resume anything that was interrupted
aicippy task resume

aicippy opens the terminal workspace — a Go Bubble Tea surface that speaks to this package over newline-delimited JSON. aicippy --pipe reaches the same runtime and the same command router as plain text, for scripting.

On a normal release launch, AiCippy checks PyPI, refreshes runtime dependencies with pip's eager strategy or uv --upgrade, verifies the installed release, and re-executes a fresh process before opening the TUI. A development version ahead of PyPI is reported without being downgraded. The fresh process then refreshes the workspace knowledge file and its relevant name-only file map. It excludes VCS, dependencies, caches, build output, virtual environments, IDE state, runtime internals, and symlinks. The full map remains local; task prompts receive a compact summary so repository context does not grow with the tree.

Terminal workspace

The default workspace is the Go Bubble Tea surface. The fixed header shows the selected provider and session; workspace, machine, configured domain, and deployment declarations appear only when the workspace fact source marks them as verified or configured. The footer shows task state, session token use, and the configured context capacity. Streamed tool output and code diffs stay in the same conversation instead of opening a second interactive implementation.

The surface accepts both slash and backslash commands. Useful commands include:

/subagent gemini|codex|claude <task> is accepted as well. Delegation uses only the provider the owner named, keeps the workspace as its working directory, streams its shell output, and still honors the configured tool approval policy. It never silently changes to another subscription CLI.

Prompting examples

Use concrete outcomes, constraints, and evidence you expect to see:

Trace the checkout timeout to its canonical owner, make the smallest connected fix,
show the unified diff, and manually exercise the terminal flow that proves it.

Research the current AWS CLI memory-record permissions from primary AWS sources.
Do not change workspace files; cite the result and call out uncertainty.

\subagent codex inspect the migration plan, update only the affected workspace files,
and report changed paths plus manual verification evidence.

The design policy is to keep the main interaction obvious: controls are rounded, their hover state uses an inverse color treatment, and the default dependency floors follow the latest verified stable releases. Dependency resolution can still select newer compatible transitive packages when an upstream package constrains them.

Durable knowledge and resource policy

Workspace facts and proven, evidence-backed task outcomes live in the single workspace-local .aicippy/knowledge.json file. User-owned notes and learned facts are preserved across refreshes; cloud knowledge remains a separate corpus.

The Skills Hub freshness writer has a bounded scope of 30 official, version-pinned instruction-body sources. EventBridge starts the same serialized CodeBuild refresh every day at 03:15 UTC, with a concurrency limit of one. /kb sync starts a one-off refresh and /kb status reports its manifest and freshness result.

The shared shell manager permits at most four live sessions. Each has an owner-selected deadline from 1 to 900 seconds, bounded live previews, process cleanup, and local stdout/stderr evidence logs. This limits active process and UI memory use while preserving task evidence on disk.

Model access and honest recovery

AiCippy does not disguise an unavailable model as a working result. A message that the model is unavailable for the AWS account is a Bedrock model-entitlement blocker; request model access through AWS, then retry. A credential or IAM failure reports its own category, so run aicippy doctor and inspect the exact response. There is no second model to fall back to: the session reports exactly why the turn could not run and /retry re-attempts it once the route is available.

What a task actually does

understand → inspect repo state → plan → select tools → execute → inspect results
     → repair failures → verify → report evidence → preserve state for resume

Every step emits a live event. Nothing is reported as done that was not observed: if a test failed, you get the output; if a step was skipped, you are told.

License

Proprietary — AiVibe Software Services Pvt Ltd. See LICENSE.