Skip to content

capability-comparison.md

Latest commit

 

History

History
166 lines (145 loc) · 29 KB

capability-comparison.md

File metadata and controls

166 lines (145 loc) · 29 KB
Raw

Capability Alignment: Pi TypeScript vs Pi Java

This document is intentionally conservative. It tracks the Java/Spring implementation against the current upstream Pi project without claiming full UX or ecosystem parity.

Upstream reference: earendil-works/pi Latest checked release: v0.81.1 (2026-07-21); refreshed 2026-07-23. The current upstream is now earendil-works/pi; the former badlogic/pi-mono URL redirects there. This document uses the current project name while retaining the Java source package com.pi.mono.* for compatibility.

Latest Upstream Release Delta

v0.81.1 adds deterministic, checksummed release source archives and makes compaction/branch summary retries follow the configured retry policy, emitting retry lifecycle events to interactive, JSON, RPC, and SDK consumers. Pi Java does not yet have standalone-release packaging, streaming lifecycle events, or an equivalent retry-aware compaction path; these are additional upstream gaps beyond the v0.80.3 snapshot below.

The v0.80.3 baseline added Claude Sonnet 5 provider catalog support, configurable outputPad, external editor settings, richer RPC session tree access (get_entries, get_tree), extension session metadata notifications, provider HTTP error response bodies, reasoning token usage counts, deterministic --no-session --session-id behavior, invalid session-file overwrite protection, BMP image handling, visible incomplete-response errors when assistant output stops at the token limit, retry for provider stream errors that explicitly ask callers to retry, and modern Azure Foundry endpoint support. Post-release main also enables Bedrock prompt caching for Claude 5, includes AWS profile-aware Bedrock fixes, exposes model-resolution helpers, skips unauthenticated default models, marks question-tool execution sequentially, adds entry renderers for session entries, derives short session entry ids from generated id random tails, serializes split-turn compaction summaries, delays Copilot device-code token polling, honors server-provided slow_down intervals in device-code polling, clamps OpenAI Responses max-output-token values below the provider minimum, improves Cloudflare 524 retry classification, rotates long-running OpenAI Codex WebSocket connections, detects DS4 context overflow errors, fails tool calls from length-truncated assistant messages with error tool results, uses (no tool output) placeholders for empty tool result payloads, normalizes null message content at ingestion boundaries, tolerates extra replacement fields in edit calls, adds before_provider_headers extension hook behavior upstream, and updates Codex SSE transport behavior. The Java implementation now covers the session-tree inspection idea through in-process APIs, CLI tree/export commands, and a minimal pi-cli --rpc JSONL adapter for get_entries/get_tree with rendered entry summaries plus tail-derived shortId fields; it also normalizes null AgentMessage content to an empty string and null metadata to an empty map at construction time, serializes split-turn compaction summary requests before merging history and turn-prefix summaries, resolves preferred LLM providers by provider id or model id while skipping unavailable providers, exposes available-model catalog and exact model-to-provider helpers, includes an OAuth device-code polling helper and minimal GitHub Copilot OAuth device-flow client that can wait before the first token poll, honor server-provided slow_down intervals, exchange the GitHub OAuth access token for a Copilot API token, persist the Copilot token plus GitHub refresh material to a local api_key credential store, and expose a credential-backed Copilot provider that can refresh expired Copilot API tokens from persisted GitHub refresh material, discover account-specific models from the Copilot /models endpoint, select the token-derived Copilot API base URL, add Copilot dynamic headers, merge ProviderHeaderContributor headers before request headers, parse a tested non-streaming OpenAI-compatible chat response, parse returned Copilot tool_calls into unified Java metadata.toolCalls, and serialize assistant tool_calls history for continuation requests when a persisted token exists, aligns the OpenAI default model to gpt-5.5, honors request-scoped OpenAI API keys, OPENAI_API_KEY env overrides, model/temperature/max-token options, request-scoped provider headers, and a Java ProviderHeaderContributor hook boundary for OpenAI/Anthropic/Bedrock/GitHub Copilot non-streaming HTTP requests, clamps OpenAI request output-token values below 16, retries OpenAI-compatible HTTP 524 timeout responses, classifies OpenAI-compatible DS4 context-overflow response bodies, parses OpenAI response content instead of returning placeholders, maps OpenAI token usage into usage.inputTokens, usage.outputTokens, usage.totalTokens, and usage.reasoningTokens, surfaces finish_reason=length as an incomplete-response marker with metadata, normalizes Azure Foundry/OpenAI base URLs to /openai/v1, preserves OpenAI HTTP error response bodies, retries OpenAI, Anthropic, and Bedrock non-streaming provider errors whose response body explicitly asks callers to retry, converts CLI image attachments into OpenAI non-streaming image_url content parts, sends OpenAI non-streaming tool schemas, parses tool_calls into unified metadata.toolCalls, serializes assistant tool_calls history for continuation requests, sends empty OpenAI tool-result payloads as (no tool output), sends non-streaming Anthropic Messages API requests with request-scoped apiKey/ANTHROPIC_API_KEY auth, Claude 5 thinking payloads, image content blocks for CLI image attachments, non-streaming tool schemas, parsed text/usage/thinking responses, parsed tool_use metadata, and provider-native tool_result blocks with empty-result placeholders, exports Java tools with sequential executionMode metadata, executes provider-returned toolCalls through Java tools, appends provider-native TOOL_RESULT messages, fails length-truncated assistant tool calls with error TOOL_RESULT messages instead of executing potentially partial arguments, and performs bounded multi-round non-streaming continuation turns, sends SigV4-signed non-streaming Bedrock Anthropic invoke payloads with Claude 5 thinking, prompt-cache blocks, image content blocks for CLI image attachments, non-streaming tool schemas, parsed text/usage/thinking responses, parsed tool_use metadata, provider-native tool_result blocks with empty-result placeholders, and provider-header contributors that cannot override SigV4-managed headers, resolves Bedrock runtime credentials from explicit properties, selected AWS shared credentials profiles, AWS environment variables, request-scoped ChatOptions.env() AWS keys, or request-scoped AWS_PROFILE plus optional AWS_SHARED_CREDENTIALS_FILE, preserves nested usage.reasoningTokens metadata across JSONL save/load, applies loaded outputPad to CLI user, assistant, and thinking lines, composes messages with configured externalEditor through /edit, supports deterministic startup session ids and ephemeral CLI runs, rejects overwriting non-empty invalid JSONL session files, applies edit replacements as literal text while tolerating model-supplied extra replacement metadata when required fields are valid, converts BMP files to PNG data URLs in the read tool, expands CLI @file references into attached file blocks before sending messages, exposes Anthropic-compatible plus Bedrock Claude Sonnet 5 catalog entries, emits a Java-side session_info_changed metadata event when a session is renamed, and rejects non-positive or oversized bash timeouts in line with the latest checked main-branch hardening. It does not yet implement real-account GitHub Copilot auth validation, streaming transport, full automatic compaction and compaction-trigger/session rewriting, Anthropic streaming tool-use, Bedrock streaming transports, Codex WebSocket connection rotation, the broader upstream RPC surface, provider-native multimodal request parts beyond current OpenAI, Anthropic, and Bedrock image attachments, upstream's exact external-editor keybinding/TUI behavior, Codex SSE transport compression, full extension-runtime before_provider_headers event delivery, or custom/streaming provider header parity.

Snapshot

2026-07-23 conclusion

Do not turn this into a percentage: the upstream packages combine runtime, terminal UX, extension ecosystem, release tooling, and SDK/RPC integration, while Pi Java deliberately adds Spring embedding. Across the twelve product-level rows below, Pi Java has one Java-specific advantage (Spring embedding), session fundamentals are mostly aligned, eight areas are partial or minimal, and the terminal UI plus executable extension/package ecosystem are not aligned. The highest-value parity work is therefore streaming/evented agent execution, then the extension/package runtime—not more built-in file tools.

Capability Area Upstream TS package(s) Java module(s) Alignment Notes
Unified LLM API and providers packages/ai pi-core, pi-llm Partial Java has typed provider contracts, AgentMessage null content normalization, OpenAI/mock providers, model-id provider resolution with unavailable-provider skipping, available-model catalog and exact model-to-provider helpers, upstream gpt-5.5 default-model alignment, OAuth device-code polling interval semantics, minimal GitHub Copilot OAuth device-flow client foundation with Copilot API token exchange, local token persistence, credential-backed provider availability, automatic expired Copilot API token refresh from persisted GitHub refresh material, account-specific model discovery through the Copilot /models endpoint, a non-streaming OpenAI-compatible Copilot chat transport foundation, non-streaming Copilot tool_calls metadata parsing, OpenAI-compatible assistant tool_calls history serialization, and provider-header contributors for Copilot chat requests, request-scoped OpenAI API key/env/model/temperature/max-token/header options plus provider-header contributors, OpenAI request output-token floor clamping, OpenAI-compatible HTTP 524 retry classification, OpenAI-compatible DS4 context-overflow classification, request-scoped Anthropic API key/env/header auth plus provider-header contributors, request-scoped Bedrock AWS key/profile env auth plus provider-header contributors that cannot override SigV4-managed headers, Azure Foundry/OpenAI base URL normalization, OpenAI response content parsing, OpenAI usage/reasoning token metadata, incomplete output-length markers, OpenAI HTTP error response-body preservation, explicit retry-instruction handling for OpenAI/Anthropic/Bedrock non-streaming client errors, OpenAI non-streaming image attachment content parts, non-streaming OpenAI tool schema and tool_calls metadata support, assistant tool_calls history serialization, empty OpenAI tool-result placeholder payloads, non-streaming Anthropic Messages API requests/responses with Claude 5 thinking payloads, image attachment content blocks, tool-use metadata, parsed thinking content metadata, and provider-native tool-result blocks with empty-result placeholders, SigV4-signed non-streaming Bedrock Anthropic invoke payloads with Claude Sonnet 5 thinking, prompt-cache blocks, image attachment content blocks, tool-use metadata, parsed thinking content metadata, provider-native tool-result blocks with empty-result placeholders, and Bedrock credentials from explicit config, shared profiles, global AWS env vars, request-scoped AWS key env vars, or request-scoped AWS profile env vars. It does not yet include upstream's broad provider/OAuth/model registry, Anthropic streaming tool-use, streaming tool-result orchestration, full extension-runtime before_provider_headers delivery, or real Bedrock streaming transport parity.
Agent runtime and state packages/agent pi-core, pi-session Mostly aligned for session basics Java supports session trees, JSONL persistence, nested usage/reasoning token metadata, invalid non-empty JSONL overwrite protection, restore, resume, fork, import, export, deterministic startup session ids, session rename metadata, a subscribable session_info_changed event source, a tested bounded non-streaming tool-call/result continuation path, length-truncated assistant tool-call failure handling, and serialized split-turn compaction summary orchestration.
Tool execution packages/coding-agent built-ins pi-tools Partial Java has read/write/edit/bash/find/grep/ls, permission categories, BMP-to-PNG read payloads, literal first/all edit replacements with extra replacement metadata tolerated, strict 1-60 second bash timeout validation, sequential executionMode metadata on exported tool schemas, and session-level multi-round execution of provider-returned toolCalls into TOOL_RESULT messages. It does not yet mirror extension-defined tools.
CLI workflow packages/coding-agent pi-cli Partial, improved Java CLI now exposes --no-session, --session-id, @file attachment expansion, /session, /rename, /resume, /tree, /fork, /export, /import, /models, /resources, /prompts, /skills, and /edit.
Context files packages/coding-agent resource loader pi-cli Partial Java loads AGENTS.md and CLAUDE.md from parent directories into startup/session metadata.
Skills and prompt templates packages/coding-agent resource loader pi-cli Partial Java discovers user-global prompts/skills, trusted project-local .pi/prompts/*.md, .pi/skills/*/SKILL.md, .agents/skills/*/SKILL.md, and installed package prompt/skill files under .pi/packages/*/prompts and .pi/packages/*/skills; it applies trusted .pi/settings.json prompts/skills override patterns as a project-local config delta over local or inherited global resources; it expands simple {{key}} prompt templates and injects basic /skill:name instructions into CLI conversations, but does not execute extension-provided tools or the full Agent Skills lifecycle.
CLI settings settings.json pi-cli Partial Java loads user-global and trusted project-local .pi/settings.json, applies outputPad to user/assistant/thinking lines, and supports configured externalEditor via /edit; it does not yet implement upstream's exact Ctrl+G/TUI keybinding workflow.
Extensions, themes, packages packages/coding-agent resource/package loader pi-cli file discovery only Gap Upstream can load executable extensions, themes, and Pi packages; Java only discovers local and installed-package prompt/skill files and does not install packages or execute package/extension code.
Project trust packages/coding-agent trust manager pi-cli Partial Java supports /trust and gates project-local prompt/skill resources. It does not yet model extension/package trust or interactive startup prompts.
TUI packages/tui pi-ui placeholder Gap Java version does not attempt upstream terminal UI parity yet.
Orchestration/process integration packages/orchestrator, RPC mode pi-cli --rpc Minimal Java exposes JSONL get_entries and get_tree session inspection, with rendered titles/plain-text/Markdown summaries and tail-derived shortId fields for entries/tree nodes. Broader RPC/orchestrator commands are still future work.
Spring embedding n/a pi-starter Java-specific value Starter auto-configuration is the main Java adoption advantage.

What Is Practical Now

  • Embed a pi-style agent loop in Spring Boot with starter wiring.
  • Use typed provider and model abstractions from Java code.
  • Normalize null AgentMessage content at the core model boundary before provider/session/CLI code reads it.
  • Resolve providers from preferred model ids and skip unavailable unauthenticated defaults.
  • List available models from currently available providers and resolve an exact model id to its available provider without fallback.
  • Use OpenAI with upstream-aligned gpt-5.5 default model and Azure Foundry/OpenAI base URL normalization.
  • Pass request-scoped provider headers and Java ProviderHeaderContributor headers through OpenAI, Anthropic, Bedrock, and GitHub Copilot non-streaming HTTP requests while keeping auth and SigV4 headers managed by provider credential resolution or transport boundaries.
  • Reuse an OAuth device-code polling helper and minimal GitHub Copilot device-flow client that can delay the first token poll, apply RFC 8628 slow_down +5s fallback, prefer a server-provided slow_down interval when present, exchange the GitHub OAuth access token for a Copilot API token, and persist the Copilot token plus GitHub refresh material to a local api_key credential store.
  • Register a GitHub Copilot provider that remains unavailable until a persisted access token exists, participates in model-id provider selection, refreshes expired Copilot API tokens from persisted GitHub refresh material, discovers account-specific models from the Copilot /models endpoint, derives the Copilot API base URL from the token proxy-ep, adds dynamic Copilot request headers, merges ProviderHeaderContributor headers with explicit request-header precedence, parses tested non-streaming OpenAI-compatible chat responses, exposes returned Copilot tool_calls as metadata.toolCalls, and serializes assistant tool_calls history for continuation requests.
  • Override OpenAI API key, OPENAI_API_KEY env value, model, temperature, and max-token options per ChatRequest.
  • Clamp OpenAI request output-token values below the provider minimum to 16.
  • Retry OpenAI-compatible HTTP 524 timeout responses.
  • Classify DS4-style OpenAI-compatible context-overflow response bodies.
  • Read real OpenAI chat response content and surface finish_reason=length as an incomplete response.
  • Preserve OpenAI token usage metadata including usage.reasoningTokens.
  • Preserve OpenAI provider HTTP response bodies in raised error messages.
  • Retry OpenAI, Anthropic, and Bedrock non-streaming provider errors whose response body explicitly asks callers to retry.
  • Convert CLI image attachments into OpenAI non-streaming image_url content parts.
  • Send OpenAI non-streaming tool schemas, expose returned tool_calls as metadata.toolCalls, and serialize assistant tool_calls history for Java-side continuation requests.
  • Send empty OpenAI TOOL_RESULT payloads as (no tool output) rather than an empty string.
  • Expose an Anthropic-compatible claude-sonnet-5 catalog entry with adaptive-thinking labeling.
  • Send non-streaming Anthropic Messages API requests with request-scoped apiKey/ANTHROPIC_API_KEY auth, Claude Sonnet 5 thinking payloads, and parse text, thinking content, plus input/output token usage from responses.
  • Convert CLI image attachments into Anthropic non-streaming image content blocks.
  • Send Anthropic non-streaming tool schemas and expose returned tool_use blocks as metadata.toolCalls for Java-side orchestration.
  • Send Anthropic TOOL_RESULT messages back as provider-native tool_result content blocks, including (no tool output) for empty results and is_error for failed Java tool executions.
  • Export Java tool schemas with sequential executionMode metadata for provider-side tool orchestration hints.
  • Execute provider-returned non-streaming toolCalls, append provider-native TOOL_RESULT messages to the session branch, fail length-truncated assistant tool calls with error TOOL_RESULT messages instead of executing potentially partial arguments, and continue through bounded multi-round provider turns.
  • Send SigV4-signed non-streaming Bedrock Anthropic invoke requests with Claude Sonnet 5 thinking/prompt-cache blocks, convert CLI image attachments into Bedrock image content blocks, send tool schemas, parse text, thinking content, tool_use, and input/output token usage from responses, and send TOOL_RESULT messages back as provider-native tool_result blocks with (no tool output) for empty results.
  • Configure Bedrock credentials through explicit pi.llm.bedrock.* properties, pi.llm.bedrock.profile/AWS_PROFILE shared credentials, AWS environment variables, per-request ChatOptions.env() AWS keys, or per-request AWS_PROFILE plus optional AWS_SHARED_CREDENTIALS_FILE without committing real keys.
  • Persist session trees as JSONL and reload them into memory.
  • Preserve nested usage metadata such as usage.reasoningTokens across JSONL save/load.
  • Serialize split-turn compaction history and turn-prefix summary requests before merging them, avoiding concurrent summary generations against single-concurrency providers.
  • Reject overwriting non-empty invalid JSONL session files so bad local session files are not silently lost.
  • Read BMP files by converting them to PNG data URLs with image metadata.
  • Expand CLI @file references into attached file blocks, including BMP-to-PNG payloads through the read tool.
  • Start deterministic in-memory CLI sessions with --session-id, including ephemeral --no-session runs where /save is disabled.
  • Resume saved sessions and inspect the active branch path.
  • Rename the current session, persist the session name metadata, and receive a Java-side session_info_changed event.
  • Fork a saved path from any node into a new JSONL session.
  • Export/import JSONL sessions for handoff and reproducibility.
  • Load AGENTS.md/CLAUDE.md context files from the project path.
  • Discover local and installed-package prompt templates and skill manifests in Pi-compatible directories.
  • Save project trust decisions with /trust and reload project-local prompt/skill resources after trust.
  • Apply trusted .pi/settings.json prompts and skills arrays with +, -, or ! entries to include or disable local prompt/skill resources, including project-local deltas over inherited global resources.
  • Load user-global and trusted project-local .pi/settings.json, apply outputPad to user/assistant/thinking lines, and show configured externalEditor with /settings.
  • Compose the next CLI message through /edit using configured externalEditor; commands can use {file} or receive the draft path as the final argument.
  • Expand trusted prompt templates with /<template> key=value and inject basic skill instructions with /skill:<name> [request].
  • Inspect current or requested sessions over minimal JSONL RPC with get_entries and get_tree, including rendered titles/plain text/Markdown summaries for entries and tail-derived shortId fields for entries/tree nodes.
  • Use built-in file/shell tools with permission categories, literal edit replacements that tolerate extra replacement metadata, sequential execution metadata, and invalid bash-timeout rejection before execution.

What Is Still Not Upstream Parity

  • Multi-provider breadth, subscription OAuth, provider transport selection, generated model catalogs, and request-scoped auth/env/header propagation beyond the currently tested OpenAI, Anthropic, and Bedrock AWS-env paths.
  • Full Anthropic streaming tool-use behavior and provider-specific streaming tool-result orchestration.
  • Full Bedrock runtime parity, including streaming transport behavior beyond the tested SigV4-signed non-streaming thinking/prompt-cache block.
  • Upstream's TUI, keyboard workflow, message queue, compaction UI, and rich session browser.
  • Full Agent Skills execution lifecycle, extension-provided tools, themes, Pi package installation/update flows, and upstream's interactive pi config TUI selector.
  • Full project trust for executable extensions, package-managed code, and interactive startup prompts.
  • Full upstream external editor keybinding/TUI workflow.
  • Full GitHub Copilot streaming transport parity and real-account auth/tool-call validation beyond the tested minimal OAuth device-flow client plus credential-backed non-streaming provider transport; full automatic compaction integration beyond serialized split-turn summary orchestration.
  • OpenAI Codex WebSocket connection rotation.
  • Provider-native multimodal attachment parts beyond the current OpenAI, Anthropic, and Bedrock non-streaming image paths; documents and streaming requests still carry CLI @file attachments as structured text.
  • Full upstream extension-runtime event delivery and RPC/orchestrator process integration beyond minimal get_entries/get_tree; Java has a provider-header contributor boundary for OpenAI, Anthropic, Bedrock, and GitHub Copilot non-streaming calls, but not full extension-runtime before_provider_headers dispatch.
  • Supply-chain hardening around package installs and shrinkwrap-style release checks.

Validation Pointers

Canonical target:

mvn clean compile
mvn test
printf "help\n/session\n/models\nexit\n" | mvn -pl pi-cli -DskipTests exec:java
printf "help\n/resources\n/prompts\n/skills\nexit\n" | mvn -pl pi-cli -DskipTests exec:java
printf "help\n/resources\n/trust\n/resources\nexit\n" | mvn -pl pi-cli -DskipTests -Dpi.trust.file=target/trust-smoke.txt exec:java
printf '{"id":1,"method":"get_tree"}\n' | mvn -pl pi-cli -DskipTests -Dexec.args=--rpc exec:java
mkdir -p target/pi-cli-resource-smoke/.pi/prompts target/pi-cli-resource-smoke/.pi/skills/deploy
printf 'Review {{target}} for {{focus}}\n' > target/pi-cli-resource-smoke/.pi/prompts/review.md
printf '# Deploy Skill\nUse deployment steps.\n' > target/pi-cli-resource-smoke/.pi/skills/deploy/SKILL.md
cd target/pi-cli-resource-smoke
printf "/resources\n/trust\n/prompts\n/skills\n/review target=api focus=security\n/skill:deploy release service\nexit\n" | mvn -f ../../pom.xml -pl pi-cli -DskipTests -Dpi.trust.file=trust-final.txt exec:java
mkdir -p ../pi-cli-settings-smoke/.pi
printf '{"outputPad":3,"externalEditor":"code --wait"}\n' > ../pi-cli-settings-smoke/.pi/settings.json
cd ../pi-cli-settings-smoke
printf "/settings\n/trust\n/settings\nhello settings\nexit\n" | mvn -f ../../pom.xml -pl pi-cli -DskipTests -Dpi.trust.file=trust-settings.txt exec:java

Focused session parity check:

mvn -pl pi-session -am test -Dtest=SessionPersistenceUnitTest
mvn -pl pi-session -am test -Dtest=SessionCompactionServiceTest
mvn -pl pi-tools -am test -Dtest=EditFileToolTest,BashToolTest,ReadFileToolTest
mvn -pl pi-llm -am test -Dtest=GitHubCopilotOAuthClientTest,OAuthDeviceCodePollerTest
mvn -pl pi-llm -am test -Dtest=GitHubCopilotLLMProviderTest
mvn -pl pi-llm -am test -Dtest=OpenAIClientTest,OpenAIConfigTest,OpenAILLMProviderTest
mvn -pl pi-llm -am test -Dtest=AnthropicClientTest,AnthropicConfigTest,AnthropicLLMProviderTest
mvn -pl pi-llm -am test -Dtest=BedrockClientTest,BedrockConfigTest,BedrockLLMProviderTest
mvn -pl pi-cli -am test -Dtest=PiResourceLoaderTest
mvn -pl pi-cli -am test -Dtest=PiTrustManagerTest
mvn -pl pi-cli -am test -Dtest=PiResourceCommandResolverTest
mvn -pl pi-cli -am test -Dtest=PiFileReferenceResolverTest
mvn -pl pi-cli -am test -Dtest=PiRpcCommandHandlerTest
mvn -pl pi-cli -am test -Dtest=PiCliStartupOptionsTest,PiCliSettingsLoaderTest,PiCliOutputFormatterTest,PiExternalEditorTest
printf "/rename Roadmap Review\n/session\nexit\n" | mvn -pl pi-cli -DskipTests exec:java

JDK 17+ verification commands (the release bytecode baseline is Java 17):

mvn -pl pi-session -am test -Dtest=SessionPersistenceUnitTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-session -am test -Dtest=SessionCompactionServiceTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-tools -am test -Dtest=EditFileToolTest,BashToolTest,ReadFileToolTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-llm -am test -Dtest=GitHubCopilotOAuthClientTest,OAuthDeviceCodePollerTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-llm -am test -Dtest=GitHubCopilotLLMProviderTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-llm -am test -Dtest=OpenAIClientTest,OpenAIConfigTest,OpenAILLMProviderTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-llm -am test -Dtest=AnthropicClientTest,AnthropicConfigTest,AnthropicLLMProviderTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-llm -am test -Dtest=BedrockClientTest,BedrockConfigTest,BedrockLLMProviderTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-cli -am test -Dtest=PiResourceLoaderTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-cli -am test -Dtest=PiTrustManagerTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-cli -am test -Dtest=PiResourceCommandResolverTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-cli -am test -Dtest=PiFileReferenceResolverTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-cli -am test -Dtest=PiRpcCommandHandlerTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
mvn -pl pi-cli -am test -Dtest=PiCliStartupOptionsTest,PiCliSettingsLoaderTest,PiCliOutputFormatterTest,PiExternalEditorTest -Dsurefire.failIfNoSpecifiedTests=false -Djava.version=17
printf "/rename Roadmap Review\n/session\nexit\n" | mvn -pl pi-cli -DskipTests -Djava.version=17 exec:java
printf "/session\n/save\nexit\n" | mvn -pl pi-cli -DskipTests -Djava.version=17 -Dexec.args="--no-session --session-id stable-cache-id" exec:java

These commands are the standard verification path on JDK 17 and later LTS JDKs, including JDK 21.