pydantic-deep v0.3.4–v0.3.8: 5 Releases in 2 Weeks

TL;DR: In two weeks of radio silence, pydantic-deep shipped five versions: a redesigned TUI with live token tracking, browser automation via Playwright, Docker sandbox and headless CI runner, a one-command curl installer, web search fixes for multi-provider setups, and three new capabilities for self-aware agents (StuckLoopDetection, LimitWarnerCapability, BM25 history search). This post is the overview — detailed deep dives ship every day this week.

by Kacper Włodarczyk — AI Engineering Lead at Vstorm. We build production AI agent systems and open-source the hard parts. Follow along on GitHub or oss.vstorm.co.

We went quiet for two weeks. No tweets, no posts, no “excited to share” announcements.

We were building.

pydantic-deep is our open-source AI coding agent — a Python-first alternative to Claude Code and similar tools, built on top of Pydantic AI. We ship real features from our production deployments: 30+ AI agent systems at Vstorm, real edge cases, real failures, real fixes.

In two weeks, we shipped v0.3.4 through v0.3.8. Five releases. Here’s what each one added.

v0.3.4 — TUI as Default + /improve Redesign

The biggest quality-of-life change in this release: running pydantic-deep now launches the TUI by default. No more flags, no more subcommands.

But the more interesting part is /improve. We redesigned the entire pipeline. Instead of loosely summarizing sessions, it now extracts two distinct types of insights:

UserFactInsight — things the agent learned about the user (preferences, patterns, context)
AgentLearningInsight — things the agent learned about itself (what worked, what failed)

Both feed into MEMORY.md as the primary target. Raw tool traces are now included in synthesis (Meta-Harness architecture: raw traces beat summaries every time — models reason better on actual data than on compressed interpretations of that data).

Other fixes in v0.3.4: /config command in TUI, per-session debug logging in .pydantic-deep/logs/, fixed /improve crash on {} in prompts, fixed silent failures now reporting failed_sessions, fixed API key loading, fixed MEMORY.md path mismatch.

v0.3.5 — Browser Automation, Docker Sandbox, Headless Runner

Three infrastructure features landed together in v0.3.5.

Browser automation via BrowserCapability: Playwright-backed, 9 tools:

from pydantic_deep.capabilities import BrowserCapability

agent = create_deep_agent(
    model="anthropic:claude-opus-4-6",
    capabilities=[BrowserCapability()],
)
# Agent can now: navigate, click, type_text, get_text,
# screenshot, scroll, go_back, go_forward, execute_js

Domain allowlist, auto-screenshot on navigation, browser lifecycle managed by wrap_run. The agent can interact with real web pages.

Docker sandbox (--sandbox docker): file operations and shell commands run inside a container. /workspace is mounted, cleanup is automatic.

pydantic-deep --sandbox docker

Named workspaces: persistent Docker environments that survive between sessions, shareable across threads.

pydantic-deep --workspace my-project  # reuse same container
pydantic-deep sandbox list            # see running workspaces
pydantic-deep sandbox stop my-project

Headless runner (pydantic-deep run): for benchmarks, CI/CD, scripted automation. No TUI, no interaction — just the agent executing a task and returning results.

Also in v0.3.5: DEFAULT_USAGE_LIMITS removes Pydantic AI’s default 50-request cap (which breaks any non-trivial coding task). Per-turn token display (in:X · out:Y · total:Z · reqs:N) in TUI header.

v0.3.6 — One-Command Install

The single most practical improvement for new users.

curl -fsSL https://raw.githubusercontent.com/vstorm-co/pydantic-deep/main/install.sh | bash

That’s it. No virtualenv setup, no pip, no version conflicts.

Two more features shipped with this:

pydantic-deep update — self-update command. Run it, get the latest version.
Startup update notifications — on launch, pydantic-deep checks for newer releases (24h cache, 2s timeout so it doesn’t block startup). If there’s an update, it tells you.

Small release, high leverage. The install experience was a friction point — now it isn’t.

v0.3.7 — Web Search Fixed for Non-Anthropic Models

A focused fix: web search now works correctly for OpenRouter-routed models and any non-Anthropic provider.

Previously, the DuckDuckGo fallback wasn’t reliably bundled. It is now, in both cli and tui extras.

If you run pydantic-deep with GPT-4.1, Gemini, or any model via OpenRouter, web search now works.

v0.3.8 — Three Self-Aware Capabilities

The most feature-dense release. Three new capabilities for agents that understand their own state.

StuckLoopDetection

Agents get stuck. Not in a “model is confused” way — in a mechanical way. The agent calls read_file("config.json"), gets a result, can’t figure out the next step, calls read_file("config.json") again. And again. And again.

StuckLoopDetection catches this at the capability level, before you burn 50 API calls:

from pydantic_deep import create_deep_agent

agent = create_deep_agent(
    model="anthropic:claude-opus-4-6",
    stuck_loop_detection=True,  # default: True
)

Three patterns detected:

Repeated identical calls — same tool, same arguments, N times in a row (default threshold: 3)
A-B-A-B alternating — tool A, tool B, tool A, tool B… locked in a two-call loop
No-op loops — same call, same result, still repeating

Action is configurable: warn (triggers ModelRetry) or error (raises StuckLoopError).

Per-run isolation via for_run() ensures parallel agent.run() calls don’t share stuck-detection state.

Tomorrow: full deep dive on StuckLoopDetection — what causes each pattern, how detection works, production examples.

LimitWarnerCapability

Before this capability, context window usage was only visible in the TUI status bar. The model itself had no awareness — it would happily keep generating output as it approached 90% usage, then hit the auto-compression trigger with no warning.

LimitWarnerCapability fixes this by injecting warnings directly into the conversation as user messages:

At 70% context usage: [URGENT] Context window at 70%. Start wrapping up current task.
At 90% context usage: [CRITICAL] Context window at 90%. Auto-compression imminent.

The model reads these. It can adjust. Auto-enabled when context_manager=True (the default).

(Full deep dive on Tuesday.)

BM25-Ranked History Search

search_conversation_history now uses BM25 scoring instead of naive substring matching. Same formula as Elasticsearch and Lucene. Zero dependencies — pure Python implementation.

Better search results, no new deps.

What’s Coming This Week

Every day this week, one deep dive:

Sunday (today): v0.3.4–v0.3.8 overview ← you’re reading it
Monday: StuckLoopDetection — full walkthrough, 3 patterns, real examples
Tuesday: LimitWarnerCapability — teaching agents to know their limits
Wednesday: Browser automation + Docker sandbox
Thursday: The curl install and what we learned about distribution
Friday: /improve pipeline — how agents get smarter after every session

Key Takeaways

pydantic-deep v0.3.4–v0.3.8 shipped five releases in two weeks of active development
TUI is now default — better token visibility, live thinking stream, session auto-save
Browser + Docker + CI — pydantic-deep now runs in full infrastructure stack
curl | bash install — no more onboarding friction
Three self-aware capabilities — StuckLoopDetection, LimitWarnerCapability, BM25 search — agents that understand their own state
All capabilities are pure Python where possible, zero extra dependencies by default

Get Started

Install pydantic-deep:

curl -fsSL https://raw.githubusercontent.com/vstorm-co/pydantic-deep/main/install.sh | bash

Or with pip:

pip install "pydantic-deep[tui]"

GitHub: github.com/vstorm-co/pydantic-deep
OSS portal: oss.vstorm.co
Vstorm: vstorm.co

pydantic-deep v0.3.4–v0.3.8: What We Built in 2 Weeks of Silence