pi_agent_rust

pi_agent_rust - Native AI coding agent CLI written in Rust

Why Should You Care? • TL;DR • Methodology • Quick Start • Features • Installation • Commands • Configuration

# Install latest release
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/main/install.sh?$(date +%s)" | bash

---

The Problem

You want an AI coding assistant in your terminal, but existing tools are:

• Slow to start: Managed runtimes can add noticeable startup overhead
• Resource intensive: Electron apps or heavy runtimes can add substantial overhead
• Unreliable: Streaming breaks, sessions corrupt, tools fail silently
• Hard to extend: Closed ecosystems or complex plugin systems

The Solution

pi_agent_rust is a from-scratch Rust port of Pi Agent by Mario Zechner (made with his blessing!). Official release archives install the single end-user binary pi, with streaming responses and 36 built-in tools (19 in the default --tools list; 14 always in the model’s schema, the rest reachable through the xdev dispatcher or enabled in settings).

Current product direction

This project is no longer trying to be a strict drop-in replacement for the legacy TypeScript Pi. That target became both impractical and undesirable as the legacy implementation evolved. Legacy Pi remains useful historical context, but it is not our compatibility authority or definition of completeness.

OMP is the closer reference for where the product is going: feature surface, agent workflows, look and feel, and overall UI/UX. Pi Rust still chooses Rust-native architecture and may intentionally differ from both projects when that produces a simpler, safer, or better coding agent. Historical drop-in and parity artifacts remain in the repository as records; they do not gate product work, releases, or user-facing claims.

All repository quality checks, builds, and releases run through Doodlestein Self-Releaser (DSR). Contributors must not invoke Cargo or RCH directly, and GitHub Actions is never an execution or evidence authority for this project. Use dsr quality --tool pi_agent_rust for the registered quality recipe and the canonical DSR build/release commands documented in docs/releasing.md.

For signed published payload verification and offline CLI smoke checks, see Published release consumer checks.

Rather than a direct line-by-line translation, this port builds on two purpose-built Rust libraries:

• asupersync: A structured concurrency async runtime with built-in HTTP, TLS, and SQLite
• rich_rust: A Rust port of Rich by Will McGugan, providing beautiful terminal output with markup syntax

# Start a session
pi "Help me refactor this function to use async/await"

# Continue a previous session
pi --continue

# Single-shot mode (no session)
pi -p "What does this error mean?" < error.log

Why Should You Care?

If you already use Pi Agent, especially through OpenClaw, this project keeps the core workflow while upgrading the engine under the hood:

• A native single-binary design intended to minimize startup and runtime overhead
• A bounded-resource architecture for long-running sessions
• A capability-gated security model for extension/tool execution, including command-level blocking of dangerous extension shell patterns

Security is a first-class design goal here, not a bolt-on:

• Capability-gated hostcalls (tool/exec/http/session/ui/events)
• Two-stage extension exec enforcement: capability gate first, then command mediation that blocks critical shell classes by default (for example recursive delete, disk/device writes, reverse shell) and can tighten to block high-tier classes in strict/safe policy
• Policy + runtime risk + quota enforcement on the execution path
• Per-extension trust lifecycle (pending -> acknowledged -> trusted -> killed) with kill-switch audit logs and explicit operator provenance
• Hostcall-lane emergency controls that can force compatibility-lane execution globally or for one extension when fast-lane behavior needs immediate containment
• Structured concurrency via asupersync for more predictable cancellation/lifecycle behavior
• Auditable runtime signals/ledgers and redacted security alerts for extension behavior

TL;DR (Pi/OpenClaw Users)

The Rust port is designed around large-session, multi-agent, and extension-heavy workloads. Release-facing performance numbers are published only when the checked-in evidence artifacts are current, have matching run provenance, and show data and a passing result for every declared budget, with no data-contract failures. Historical benchmark snapshots are retained in planning/evidence artifacts, but they are not treated as current README claims until the performance evidence gate is regenerated cleanly.

Extension runtime guarantees are also concrete:

Bottom line: Pi’s architecture targets lower latency, lower memory use, and stronger extension runtime safety under real workload pressure; current numeric claims must come from fresh, provenance-matched evidence artifacts.

Data source: docs/planning/BENCHMARK_COMPARISON_BETWEEN_RUST_VERSION_AND_ORIGINAL__GPT.md (latest secure-path + full orchestrator checkpoints, 2026-04-23).

README Citation Convention

Release-facing numeric performance claims in this README include inline citations with format: *(from [artifact-path], run [correlation-id])*

Example: *(from [artifact-path], run [correlation-id])*

Two additional machine-recognized citation forms exist:

• *(from [artifact-path])* — path-only citation. The cited artifact must exist and parse; freshness is enforced by the same 14-day rule as release-facing claims.
• *(from [artifact-path]; historical snapshot)* — explicit historical contract. The citation itself declares the obligation a retained snapshot: existence and validity are still checked, but staleness is not enforced because such claims never satisfy current release-facing requirements.

scripts/check_readme_evidence_freshness.py (a pre-release check listed in docs/releasing.md, not part of the code quality recipe) checks file freshness and artifact content so stale, no-data, or correlation-mismatched evidence cannot back user-facing performance claims; for release-facing citations of budget_summary.json it validates the full pi.perf.budget_summary.v2 contract, including that the header counts equal the per-budget rows. It reports line-numbered proof obligations for cited claims and extracts claim-gated performance phrases for reviewer audit. Historical snapshot citations are mapped separately and do not satisfy current release-facing claims.

Performance-Oriented Architecture

In this README, we means the project owner and collaborating coding agents.
The design concentrates performance work in several runtime layers rather than assuming one optimization proves an end-to-end result.

If you want the full implementation inventory, see Performance Engineering.

Benchmark Methodology and Claim Integrity

The benchmark evidence policy is designed to keep results realistic, reproducible, and hard to game.

What we measured:

• Matched-state workloads: resume a large session and append the same 10 messages.
• Realistic E2E workloads: resume + append + extension activity + slash-style state changes + forks + exports + compactions.
• Scale levels: from 100k up to 5M token-class session states.
• Startup/readiness: command-level readiness (--help, --version) separately from long-session workflows.

How we kept comparisons fair:

• Two scopes in the benchmark report:
  • apples-to-apples (pi_agent_rust vs legacy coding-agent)
  • apples-to-oranges (legacy stack components included where legacy behavior is outsourced)
• Release-mode binaries and repeated runs per matrix cell.
• No paid-provider noise in core latency/footprint tables (provider-call costs are excluded from these core comparisons).

How we kept claims honest:

• Security controls stayed on during secure-path measurements (no policy/risk/quota bypasses for speed claims).
• Raw artifacts are preserved (JSON/trace/time outputs) and called out in the benchmark report.
• Blockers are explicitly disclosed: when direct legacy reruns were blocked by missing workspace deps, we state that and compare against prior validated legacy artifacts instead of pretending reruns succeeded.
• Interpretation notes are explicit: the report distinguishes baseline sections vs fresh reruns so readers can see exactly which values came from which run set.
• Reproducibility over marketing: methodology, caveats, and known limits are included alongside wins.

If you want full details, see:

• docs/planning/BENCHMARK_COMPARISON_BETWEEN_RUST_VERSION_AND_ORIGINAL__GPT.md (methodology + results + caveats + raw artifact paths)

Why Pi?

Quick Example

# 1) Start an interactive session
pi

# 2) Ask a codebase question
pi "Summarize the architecture in src/"

# 3) Attach a file inline
pi @src/main.rs "Explain startup flow"

# 4) Run single-shot mode for scripting
pi -p "List likely regression risks for this diff"

# 5) Continue your last project session
pi --continue

# 6) Inspect available models/providers
pi --list-models
pi --list-providers

---

Foundation Libraries

asupersync

asupersync is a structured concurrency async runtime designed for applications that need predictable resource cleanup. Key features used by pi_agent_rust:

• Capability-based context (Cx): Async functions receive an explicit context that controls what they can do (HTTP, filesystem, time). This makes testing deterministic.
• HTTP client with TLS: Built-in HTTP API with rustls, avoiding OpenSSL dependency hell
• Structured cancellation: When a parent task cancels, all child tasks cancel cleanly. No orphaned futures.

pi_agent_rust runs on asupersync end-to-end today (runtime + HTTP/TLS + cancellation). Provider streaming uses a minimal HTTP client (src/http/client.rs) feeding a custom SSE parser (src/sse.rs).

rich_rust

rich_rust is a Rust port of Will McGugan’s Rich Python library. It provides:

• Markup syntax: [bold red]error[/] renders as bold red text
• Tables: ASCII/Unicode table rendering with alignment and borders
• Panels: Boxed content with titles
• Progress bars: Animated progress indicators
• Markdown: Terminal-rendered markdown with syntax highlighting
• Themes: Consistent color schemes across components

The terminal UI uses rich_rust for all output formatting, providing the same visual quality as Rich-based Python tools.

---

Quick Start

1. Install

# Install latest release binary
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/main/install.sh?$(date +%s)" | bash

If you already have the original TypeScript pi installed, the installer asks whether to make Rust Pi canonical as pi and automatically create legacy-pi for the old command.

2. Configure API Key

export ANTHROPIC_API_KEY="sk-ant-..."

3. Run

# Interactive mode
pi

# With an initial message
pi "Explain this codebase structure"

# Read files as context
pi @src/main.rs "What does this do?"

Using a local model

ollama, llamacpp (llama.cpp’s llama-server), mistralrs (mistral.rs), and lmstudio are built-in local providers. The first three need no API key and work out of the box against their default ports:

pi --provider ollama    --model llama3        -p "hi"
pi --provider llamacpp  --model  -p "hi"
pi --provider mistralrs --model default        -p "hi"

To point at any other OpenAI-compatible server (a custom host/port, vLLM, etc.), add it to ~/.pi/agent/models.json:

{
  "providers": {
    "ollama": {
      "api": "openai-completions",
      "baseUrl": "http://127.0.0.1:11000/v1",
      "apiKey": "ollama",
      "models": [{ "id": "Qwen3.6-35B-A3B-4bit", "contextWindow": 32768 }]
    }
  }
}

Then pi --provider ollama --model Qwen3.6-35B-A3B-4bit. See docs/models.md for the full models.json schema, provider aliases, and secret resolution (env vars and !command shell lookups).

Refreshing provider model catalogs

For OpenAI-compatible providers, model discovery is a standalone command: it prints one model ID per stdout line and exits without starting the TUI.

# Use a successful live response or (with a warning on stderr) the static
# registry when live discovery is unavailable
pi --fetch-models openrouter

# Bypass the cache and require a genuinely live response; never fall back
pi --fetch-models openrouter --refresh-models

# Make a successful live/cache catalog available to future --list-models and
# interactive /model pickers
pi --fetch-models openrouter --refresh-models --persist-models

Each standalone CLI invocation starts a new process, so its in-memory cache is fresh. The five-minute cache only avoids repeat discovery calls made within one long-lived process by SDK/library users; --refresh-models bypasses that cache.

Discovery also reads usable supported_reasoning_levels advertisements, including metadata.supported_reasoning_levels and entries shaped as {"effort":"high"}. Nested metadata takes precedence. Pi preserves the advertised tiers through the cache and persisted catalog, constrains the model’s thinking choices to those tiers, and sends the selected effort through the provider’s request format. The recognized tiers are off (also advertised as none), minimal, low, medium, high, xhigh, and max. Missing, empty, or unusable advertisements leave existing model capabilities intact. Your configured thinking level remains authoritative; discovery does not select a server-advertised default effort.

Persistence is opt-in. The v2 ~/.pi/agent/models.fetched.json schema stores provider/model IDs, optional bounded reasoning tiers, the fetch timestamp, and a non-secret SHA-256 identity binding membership to the provider, API, query-free endpoint, auth-header mode, recognized credential-query ordered name/presence shape, and request-header name/presence shape that produced it. Pi never persists a static fallback, API credentials, URL query values, or request-header values—not even as offline-verifiable hashes. Credential-channel values are rotation-tolerant. Any non-empty query/header value outside a narrowly recognized credential channel may identify a tenant or deployment, so Pi refuses to persist that catalog (and ignores legacy persisted rows for such a current route) rather than reuse unverifiable membership. If a safely bound route shape no longer matches, Pi ignores those generated rows and reports how to refresh them. Credential rotation alone does not invalidate the catalog. The generated membership is likewise not account-bound: switching accounts on the same endpoint/transport shape can retain the prior account’s saved model list until you rerun --fetch-models --refresh-models --persist-models. Inference still resolves and sends the current account’s credential; only the opt-in model-membership list can be stale across that switch. The generated catalog is loaded first; your hand-written ~/.pi/agent/models.json is loaded afterward and remains authoritative. Pi does not rewrite or merge that user-authored file. Legacy pi.models.fetched.v1 files cannot be rebound safely because they lack this provenance, so Pi preserves them instead of overwriting them. Move the legacy file aside to models.fetched.v1.backup.json, then rerun the verified live refresh and persist command above to create a v2 catalog.

---

Features

Streaming Responses

Real-time response streaming with extended thinking support:

pi "Write a quicksort implementation"

Watch the response appear incrementally, with thinking blocks shown inline.

36 Built-in Tools

Tools are tiered so the model’s live schema stays small while everything remains reachable. The tier table and default enabled list live in src/xdev.rs; src/cli.rs resolves CLI selection and optional-tool settings:

• Essential (always in the provider schema): read, write, edit, bash, grep, find, ls, hashline_edit, ask, todo, web_search, submit_plan, current_time, xdev
• Discoverable (registered, hidden from the schema until promoted via xdev list/describe/run/promote): ast_grep, ast_edit, lsp, debug, manage_skill — plus the memory-bank tools (retain, recall, reflect, memory_edit, learn) when memory.backend is local or cass (local memory plus read-only project history)
• Default-enabled: jobs (background bash job control) and hub (PTY service supervision), alongside the essential tier. The default --tools list names 19 tools; the registry adds manage_skill by default and, when any discoverable tool is enabled, the xdev dispatcher
• Settings or --tools opt-in extras: eval (tools.enableEval), github (tools.enableGithub), security_scan (tools.enableSecurityScan)
• Settings-gated extras (off until enabled in settings.json): browser (browser.enableBrowser), computer (computer.enableComputer), the media trio inspect_image, generate_image, tts (media.enableInspectImage / media.enableGenerateImage / media.enableTts), and read_media (media.enableReadMedia; inline video/audio for Gemini-family models)
• Opt-in only: subagent (tools.enableSubagent or --tools; it can start additional coding-agent processes)

Project memory learning

The local memory bank can extract reusable facts, lessons, preferences and decisions from completed saved sessions, and consolidate facts learned earlier. Enable the bank in settings, then run /memory sync in the default interactive UI to learn from the current completed branch:

{
  "memory": {
    "backend": "local",
    "autoExtract": true,
    "extractEveryTurns": 3
  }
}

autoExtract is off unless explicitly enabled. When enabled, one bounded, tool-free request runs after every three newly authored turns by default; the cadence accepts 1–100. The request uses the session’s current model and provider credentials, has a 30-second deadline, and shares any remaining foreground time budget. Extraction failure leaves the completed foreground answer intact and can be retried with /memory sync. Explicit synchronization works with autoExtract off. Ephemeral sessions and delegated child agents do not extract project memories. The cass backend uses the same local learning pipeline.

Extraction resolves the session’s current attached workdir, screens complete bounded text evidence before selecting a recent request window, and redacts both request and stored results. Oversized evidence is refused rather than partially screened. Each successful extraction commits facts, search indexes, audit records and a durable branch checkpoint together, so reopening or revisiting an already synchronized branch does not repeat the provider request. Automatic consolidation can supersede only unchanged generated memories supplied to the request; manually retained or edited facts keep their ownership. Historical versions remain auditable. Learned facts become available through recall, reflect and the existing startup memory block. /memory view, list, search and forget remain available for inspection and management.

All tools include automatic truncation for large outputs (2000 lines / 1MB), detailed metadata in responses, and process-tree cleanup for bash. Background-job logs live in a dedicated directory with a 256 MiB / 4096-entry budget. When it cannot admit another 16 MiB artifact, Pi deletes the oldest unlocked logs, only as many as needed, and always keeps active logs and at least eight recent ones. Set PI_JOBS_ARTIFACT_RETENTION=preserve to keep every log instead; Pi then refuses new background jobs once the budget is full. Job snapshots report the applied policy, removed-file count, and reclaimed bytes in artifactCleanup.

Print mode can keep working after a background job finishes. Add --wait-async or set "async": {"printWake": true} to wait for the current session’s jobs, deliver their completion notices to the agent, and run follow-up turns in the same session. A follow-up may start further jobs. The process exits when all owned work has settled and its notices have been delivered. Text continues streaming each answer; JSON mode emits the corresponding agent lifecycles.

pi -p --wait-async --max-time 3600 \
  "Run the integration suite in the background and fix any failures."

The flag is opt-in; ordinary print mode exits after its prompts and cancels unfinished jobs. With async waiting, --max-time covers the original run, idle waits, follow-up turns and retry delays. Reaching it fails the run and cancels owned jobs. Ctrl+C also interrupts idle waits and retry delays.

Hub service logs retain the first 16 MiB of raw output per run. Pi continues draining the service after that limit or an artifact write failure, so live tail logs, readiness checks, and PTY input continue working. logs, ps, and describe report logCapture byte counts, truncation, I/O errors, and output drain state separately from process status; incremental reads also report evicted lines. The cap applies to each artifact, not the total across runs. detached: true skips session-exit cleanup, but the PTY and its supervisor still belong to the current Pi process: cross-process reattachment and survival after that host process exits are not implemented.

To enable optional tools across launches, add their flags to the existing tools object in ~/.pi/agent/settings.json or a trusted project’s .pi/settings.json:

{
  "tools": {
    "enableEval": true,
    "enableGithub": true,
    "enableSecurityScan": true,
    "enableSubagent": true
  }
}

Each flag defaults to false and also accepts snake_case (enable_eval, enable_github, enable_security_scan, enable_subagent). With no --tools argument, enabled extras join the existing defaults. An explicit --tools list overrides these flags for that run, even if it equals the default list. --no-tools and an empty, whitespace-only, or comma-only --tools value disable all built-ins, including settings-gated extras. Extensions and MCP servers retain their own loading controls. Trusted project flags override global flags individually, including false; untrusted project settings are excluded. Existing approval policies and native child-agent allowlists still apply. See Optional-tool settings.

Per-tool exposure is configurable via tools.loadMode. set to essential, discoverable, or off. A configured off entry excludes a tool from implicit selection, including tools automatically added by settings and manage_skill. An explicit --tools name overrides that exclusion without enabling a separate backend such as the memory bank. The generated xdev dispatcher remains disabled by tools.loadMode.xdev: "off". Load modes do not activate an optional tool by themselves. An explicit tool selection remains available:

pi --tools read,bash,edit,write,grep,find,ls,hashline_edit,subagent \
  "Use the scout agent to inspect the provider implementation."

Model Spend Limits

Use --warn-cost to show a one-time notice and --max-cost to stop admitting model requests once known spend reaches a USD threshold:

pi --warn-cost 2 --max-cost 5 "Implement and test the change."
pi -p --wait-async --max-cost 10 "Delegate the investigation and report the findings."

The corresponding settings are session.warnCost and session.maxCost; CLI values override them. Values must be positive finite numbers. The limits cover the active session tree: foreground turns, retries, native subagents, advisor reviews, compaction, titles, memory extraction and side questions share the same account. Child requests consult the parent’s live balance; missing or ambiguous accounting acknowledgements block further child requests.

Known partial usage from an in-flight request counts toward the next admission. The provider’s final bill replaces that request’s provisional estimate exactly once, including when the final bill is lower. Cancelling or reaping a native child retains partial usage already acknowledged by its parent, so siblings cannot spend that allowance again. An already-admitted request may settle its bill during cancellation; this does not permit new model requests.

{
  "session": { "warnCost": 2, "maxCost": 5 }
}

Resuming includes recorded assistant usage from all retained branches and saved auxiliary-spend checkpoints. Forking, rewinding, retrying, changing models and reloading resources preserve the live balance. A new session gets a fresh balance; switching to another saved session restores that session’s recorded spend. A resumed session already above its warning threshold does not repeat the warning. Change launch flags or settings before reopening to use different thresholds.

The Rust SDK exposes SessionOptions.warn_cost, max_cost and an optional shared cost_budget; CostBudget clones share one account. Warnings arrive as AgentEvent::CostWarning (type: "cost_warning" in JSON). A refused request returns Error::CostCap; print mode exits with status 3. Native RPC marks the terminal failure with errorCode: "cost_cap", retryable: false, and costCap: { "spent": ..., "cap": ... }.

These are request-boundary limits based on known provider costs or catalog token pricing. Requests already in flight may finish above the cap, and unreported/unpriced usage cannot be estimated. Spend is checkpointed with session saves, including normal turn completion, replacement and SDK shutdown. Only settled charges enter saved cost checkpoints; an in-flight estimate cannot permanently inflate the restored balance if its final bill is lower. When a saved session is replaced or shuts down, captured background requests retain a durable billing destination for that original session. Their final charges go to a separate cost journal beside the session file, without rewriting its transcript or racing a newly reopened session’s saves. Resume combines those charges with saved checkpoints exactly once; a reopened live account also imports newly arrived charges before admitting another request. This applies to JSONL and SQLite sessions, including the first charge arriving after a zero-spend checkpoint. Failed replacements and same-file reloads keep the foreground account live. Cost journals are not created for --no-session.

Saved forks retain references to observed unfinished billing writers. A fork can recover their final charges even after every foreground handle has closed, without opening an ancestor’s transcript or SQLite database. Repeated forks of one live account use the same journal. Once a writer’s final closure is observed, later checkpoints retain its final total and remove the completed reference. Missing required journals or writers block further requests instead of resetting spend.

Keep the adjacent .pi-cost-*.json files with their session files when backing up a session directory. Unfinished references use the original absolute journal location: moving or copying the directory does not redirect outstanding billing to the copy. Observe final closure and save a new checkpoint before expecting a snapshot to be independent of its original journals. Failure to acknowledge a journal write blocks further requests on the affected account. If an incomplete save leaves primary usage and journal charges with an uncertain overlap, the known lower bound is retained and further requests are blocked. A crash before a normal checkpoint or before a final charge receives its durable acknowledgement can still lose uncheckpointed spend; these limits do not replace the provider’s billing records.

Native Subagents and Orchestration

Rust Pi includes a native subagent tool; it does not depend on a QuickJS extension and never resolves a child executable by assuming a pi binary on PATH. By default it starts the current Rust Pi executable. Set PI_SUBAGENT_PI_BINARY=/absolute/path/to/rpi only when an explicit binary override is needed.

Enable it persistently with "tools": { "enableSubagent": true } in settings.json, or include subagent in an explicit --tools list. It remains off by default; see built-in native subagents for the bundled agent profiles.

Agent definitions are Markdown files in $PI_CODING_AGENT_DIR/agents/*.md (normally ~/.pi/agent/agents/*.md) or the nearest .pi/agents/*.md. Project definitions take precedence over same-named user definitions. The process inherits the parent’s provider, router, authentication, and model-registry environment, including PI_CODING_AGENT_DIR.

Children choose their model in this order: the agent definition’s model: pin, the configured task role (or smol fallback), then the parent’s active provider and model. Each request captures the parent selection once, including any manual switch or failover already in effect; queued children, chain steps, schema retries, and explicit revivals retain that selection. Children resolve credentials from their inherited environment and existing authentication configuration. Credentials supplied only to the parent’s CLI or in memory are not copied into child arguments or new environment variables.

---
name: scout
description: Inspect code and return concise, evidence-backed findings.
model: ai-router/gpt-5.6-sol
reasoning: low
tools: read,grep,find,ls
skills: ../skills/codebase-archaeology/SKILL.md
---

Read relevant code before answering. Do not edit files.

The tool accepts exactly one workflow shape: a single { agent, task }, a bounded parallel tasks array (up to 8, default concurrency 4), or a sequential chain array. In chains, {previous} in a task is replaced by the preceding child’s final output. Children run headlessly in isolated ephemeral sessions, stream progress into the parent TUI/JSON output, return structured status and stderr details, and are killed/reaped if the parent is cancelled. Child agents receive the declared tool allowlist; the default child allowlist deliberately excludes subagent to prevent accidental recursive delegation.

With per-task isolation: "worktree", a parallel batch captures each source workspace before launching children, including staged, unstaged, and untracked files. Queued children and corrective retries use that same baseline. Accepted patches apply after execution in task order; conflicts and cancelled work remain in their worktrees for inspection. isoApply: "keep" retains a result without applying it, and isoApply: "drop" discards its worktree after reporting the patch.

In interactive mode, /tan starts the same task-role child machinery without interrupting the main conversation. The child runs in the current directory, appears in hub agent roster with kind=tan, and sends its bounded completion summary through the follow-up queue at the next idle turn boundary. Because /tan inherits the delegation safety gate, it is available only when the opt-in subagent tool is enabled. It captures the same model precedence when the command is submitted, so a later parent model switch does not retarget it.

Native children include hub in their default tool set and can exchange messages with other children of the same parent Pi process. Agent definitions with an explicit tool allowlist must include hub to enable this capability. A child’s hub call with {"op":"agent","action":"roster"} returns its authenticated selfId and the parent’s child roster. Use those run IDs as recipients:

{"op":"agent","action":"send","name":"receiver-1","text":"The parser change is ready."}
{"op":"agent","action":"inbox"}
{"op":"agent","action":"send","name":"parent","text":"I found a blocking dependency."}

Sibling messages enter the recipient’s steering queue between turns and remain available in its inbox. The parent reads child reports with {"op":"agent","action":"inbox","name":"parent"}. Inbox pages return nextCursor, hasMore, and oldestCursor; pass nextCursor as cursor to request later messages. Each recipient retains its latest 64 messages, so clients can use oldestCursor to detect that earlier messages have fallen outside retention.

The parent authenticates the child’s sender identity. Children can read their own inbox, list peers, and send to a live sibling or the parent; transcript, steering, kill, and revival controls remain on the parent surface. The local connection closes when its owning child run ends or is cancelled.

Session Management

Sessions persist as JSONL files with full conversation history:

# Continue most recent session
pi --continue

# Open specific session
pi --session ~/.pi/agent/sessions/--home-user-project--/2024-01-15T10-30-00.jsonl

# Ephemeral (no persistence)
pi --no-session

Sessions support:

• Tree structure for conversation branching
• Model/thinking level change tracking
• Automatic compaction for long conversations

Extended Thinking

Enable deep reasoning for complex problems:

pi --thinking high "Design a distributed rate limiter"

Thinking levels: off, minimal, low, medium, high, xhigh, max

Customization (Skills & Prompt Templates)

• Skills: Place a skill directory containing SKILL.md under .pi/skills/, .agents/skills/, ~/.pi/agent/skills/, or ~/.agents/skills/ and invoke it with /skill:name. See skill discovery and precedence.
• Prompt templates: Markdown files under ~/.pi/agent/prompts/ or .pi/prompts/; invoke via / [args].
• Packages: Share bundles with pi install npm:@org/pi-packages (skills, prompts, themes, extensions).

Autocomplete

Pi provides context-aware autocomplete in the interactive editor:

• @ file references: Type @ followed by a path fragment to attach file contents. The completion engine indexes project files (respecting .gitignore) via the ignore crate’s WalkBuilder, capping at 5,000 entries.
• / slash commands: 40+ built-in commands (see the slash-command table below) and user-defined prompt templates and skills all appear as completions.
• Fuzzy scoring: Prefix matches rank above substring matches. Results are sorted by match quality, then by kind (commands > templates > skills > files > paths).
• Background refresh: A background thread re-indexes the project file tree every 30 seconds, so completions stay current without blocking the input loop.

Four Execution Modes

Pi runs in four modes, each suited to different workflows:

ACP workspace settings. Sessions created by pi --acp resolve global settings and trusted settings from the requested workspace when creating, loading from disk, or forking a session. Reattaching to a running session keeps its admitted settings, including prompt, steering, and media policy. Launching with --trust does not approve unrelated workspaces requested later by the editor. Explicit CLI choices keep precedence. See ACP workspace settings for trust and configuration-path details.

ACP model selection. After initialize, session/new, session/load, and session/resume return model choices and the active selection. Modern clients use the model selector in configOptions and send session/set_config_option with configId: "model". Older clients can use models.currentModelId, models.availableModels, and session/set_model with modelId. Both surfaces use the same opaque provider/model values, including custom BYO IDs containing slashes or query strings. Successful selections publish a complete config_option_update before the response, including any change to the model’s supported thinking level.

Choices come from Pi’s existing model registry and credential availability, with the requested workspace’s disabled-provider policy applied. --models and enabled_models guide startup and cycling; explicit editor selections can choose other registered models. The current model stays visible when restoring a saved session or making an explicit selection outside those preferences. An explicit --api-key admits additional models only for its bound provider; other providers still need their own credentials unless they require no key. If an editor shows an empty picker, compare its session/new response with pi --list-models using the same configuration and launch environment.

ACP session forks. Pi advertises the optional ACP v1 draft sessionCapabilities.fork capability. A client can create an active branch from the source’s complete selected conversation:

{"jsonrpc":"2.0","id":7,"method":"session/fork","params":{"sessionId":"SOURCE_ID","cwd":"/absolute/workspace","mcpServers":[]}}

The response contains a new sessionId, model choices, and configOptions. The last user message and assistant reply are included in the fork; the source keeps its conversation and remains usable. Forking does not run the model or replay the transcript. Use session/load on the returned ID when an editor needs the inherited history. A live source or an exact saved ID in the configured session directory can be forked. With --session-dir, the fork is saved before success is acknowledged, including retained model spend and pending billing ancestry. Live source and fork share the configured session-tree cost budget.

Forks stay in the source’s canonical workspace; nonempty additionalDirectories are rejected. The request supplies a fresh MCP server list (omission means none), and editor file requests, terminal requests, and permissions use the new session identity. An active source prompt, unresolved editor write, unsettled terminal cleanup, or failed-persistence admission fence blocks the fork until the source is ready.

ACP attachments. session/prompt accepts ordered text, images, native audio blocks, and embedded audio/video resource blobs. Audio-only prompts work. Gemini-family providers receive native media; other transports show an explicit omission label while the complete attachment stays in session history and editor replay. Audio/video admission honors media.maxBytes (5 MiB by default), with a 10 MiB transport ceiling per attachment and at most 32 media attachments per prompt. See ACP audio and video prompts for the editor wire format and aggregate limits.

ACP live steering. The initialize response advertises _meta.steering.supported: true. Editors can send _session/steering with sessionId and the same prompt content blocks as session/prompt while a turn is running. Text, images, audio/video, and embedded references retain their native form and the same media limits. Set _meta.steering.queue to "followUp" to request delivery after the current model turn; the default "steering" lane delivers at the next native steering boundary. Steering cancels that session’s open permission dialog via $/cancel_request; a follow-up leaves the dialog open.

The response is {"outcome":"injected","queue":"steering"} (or "followUp") when the native agent claims the input. This acknowledges handoff, not completed execution or durable storage. Idle input, and input still unclaimed when its owning turn ends or is cancelled, returns "promptRequired" so the editor can submit it as a new session/prompt. Unclaimed responses follow the original prompt’s terminal response and cleanup; unclaimed input is never carried into a different prompt. The per-turn queue holds at most 100 requests and 8 MiB; each request is limited to 128 blocks, 4 MiB for serialized content and request ID together, and 256 KiB of authored prose. An overflowing queue returns "failed" without accepting the input.

ACP editor files. Clients can advertise clientCapabilities.fs.readTextFile and writeTextFile in initialize. Pi then sends standard outbound fs/read_text_file and fs/write_text_file requests, scoped to the owning sessionId and an absolute workspace path. Native read, write, edit, and hashline_edit work with unsaved editor buffers, including new files that do not yet exist on disk. Tool approval and workspace/symlink checks run before any editor request. Text buffers are limited to 10 MiB; read pagination and hashline tags are derived from the editor text. Rendered documents, notebooks, binary media, internal artifacts, and non-file URLs keep their native routes.

Capabilities are independent: without an advertised capability the corresponding operation remains local. A failed advertised editor read never falls back to stale disk content. With a read-only client, an edit can write locally only if disk still matches the exact editor preimage. Edits re-read their source before writing to catch intervening changes; ACP does not provide atomic versioned compare-and-swap. After an acknowledged editor write, Pi reads back when possible and reports formatter changes from the resulting buffer. If read-back is unavailable, Pi reports an acknowledged, unverified write and does not claim a verified diff or a disk undo record.

Cancellation discards pending reads and ignores their late responses. A write already published to the editor may still complete: cancellation or timeout stops the current prompt and holds further prompts or reloads for that session until its matching valid acknowledgment or error arrives. Pi explicitly reports that the write was not rolled back. Malformed or unrelated responses cannot clear this hold, and other sessions can continue normally.

Interactive mode provides the full experience: a multi-line text editor with history, scrollable conversation viewport, model selector (Ctrl+L), scoped model cycling (Ctrl+P/Ctrl+Shift+P), session branch navigator (/tree), and real-time token/cost tracking. Since v0.4.0 the default interactive stack is the FrankenTUI (ftui) runtime; pi --inline keeps your shell scrollback by drawing the UI at the bottom of the screen instead of on the alternate screen, and pi --classic (aliases --classic-tui, --charmed, --bubbletea) selects the previous charmed_rust stack until it is removed.

Print mode sends one message, streams the response to stdout, and exits. Useful for shell scripts and one-off queries.

Print mode and tool approval. The approval mode defaults to always-ask on every surface, including -p. The absence of a terminal is not treated as consent, so Pi does not quietly auto-approve a scripted run. Print mode also has no way to prompt, which means a gated tool call in a default -p run is denied. Pass --approval-mode yolo (or --yolo) to auto-approve tool calls, or set approval.mode in settings.json:

echo "list the files here" | pi -p --mode json --yolo

When a run does end with tool calls denied for want of an approval surface, Pi exits 3 and explains why on stderr, rather than exiting 0 on a turn that silently did nothing. In --mode json and --mode rpc the same failure also arrives as the single machine-readable record described under --mode, with "code": "approval.surface_unavailable". Exit 3 means specifically “the model could not use tools”; exit 1 remains an ordinary failure and exit 2 a usage error.

This is a deliberate divergence from the Node Pi CLI, which auto-approves in print mode. Legacy Pi is historical context here, not a compatibility authority, and inheriting its default would silently grant bash and write access to any script that never opted in.

RPC mode exposes a line-delimited JSON protocol for programmatic control. Clients send commands (prompt, steer, follow-up, abort, get-state, compact) and receive streaming events. This is how IDE extensions and custom frontends integrate with Pi. See RPC Protocol for the wire format.

Extensions

Pi supports two extension runtime families with capability-gated host connectors:

• JS/TS entrypoints run without Node or Bun in an embedded QuickJS runtime.

• *.native.json descriptors run in the native-rust descriptor runtime.

• Extension entrypoints are auto-detected:

  • .js/.ts/.mjs/.cjs/.tsx/.mts/.cts run directly in embedded QuickJS (no descriptor conversion).
  • *.native.json loads the native-rust descriptor runtime.
  • One session currently uses one runtime family at a time (JS/TS or native descriptor).

• Cold/warm extension load paths are instrumented; fresh v0.3.0 measurements are pending

• Node API shims for fs, path, os, crypto, child_process, url, and more

• Capability-based security: extensions call explicit connectors (tool/exec/http/session/ui) with audit logging

• Command-level exec mediation: dangerous shell signatures are classified and blocked before spawn, with redacted denial alerts and mediation ledger entries

• Trust-state lifecycle and kill-switch controls with audited state transitions (pending/acknowledged/trusted/killed)

• Workspace trust-on-first-use gate: project-local .pi/settings.json packages and .pi/extensions/ require a one-time interactive approval (keyed to workspace path + content digest; --trust, PI_WORKSPACE_TRUST, or global trustAllWorkspaces for automation; non-interactive launches fail closed)

• Hostcall reactor mesh with deterministic shard routing, bounded queue backpressure, and optional NUMA-aware telemetry

• Fresh owner-isolated realms on every reload, with versioned persistent transpile caching instead of mutable realm reuse

Credential-Aware Model Selection

• /model (or Ctrl+L) opens a selector focused on models that are ready to run with current credentials.
• Ctrl+P and Ctrl+Shift+P cycle through the scoped model set without opening the overlay.
• Provider IDs and aliases are matched case-insensitively in model selection and /login.
• Models that do not require configured credentials can run keyless.

Extensions can register tools, slash commands, event hooks, flags, providers, and shortcuts. See EXTENSIONS.md for the full architecture and docs/extension-catalog.json for the 223-entry catalog with per-extension conformance status and perf budgets.

Extension Validation Pipeline

This project validates extension compatibility with a three-track pipeline:

• Vendored corpus (223, plus one intentionally excluded negative test fixture): deterministic conformance, compatibility matrix, and scenario suites.
• Unvendored corpus (777): source acquisition and onboarding prioritization.
• Release-binary live-provider E2E: real target/release/pi execution against a non-mocked provider/model path.

Why this exists

• Catch runtime/API regressions in QuickJS host shims and capability policy.
• Catch dangerous extension shell call patterns with real command mediation on the release binary path.
• Verify extension behavior against real provider responses, not just fixture/mocked flows.
• Keep extension support measurable instead of anecdotal.
• Produce a prioritized queue for onboarding unvendored candidates into vendored conformance.

Pipeline components

1. Fetch unvendored source corpus

   • Binary: ext_unvendored_fetch_run
   • Execution authority: the registered DSR extension-validation lane; do not invoke the example directly with Cargo
   • Purpose:
     • Clones GitHub repos and unpacks npm tarballs into .tmp-codex-unvendored-cache/
     • Produces machine-readable acquisition status for all unvendored candidates
   • Artifacts:
     • tests/ext_conformance/reports/pipeline/unvendored_fetch_probe_report.json
     • tests/ext_conformance/reports/pipeline/unvendored_fetch_probe_events.jsonl

2. Run end-to-end validation orchestration

   • Binary: ext_full_validation
   • Execution authority: the registered DSR extension-validation lane
   • Stages (in order):
     1. refresh_onboarding_queue (runs ext_onboarding_queue)
     2. conformance_shard_0..N (runs ext_conformance_generated sharded matrix)
     3. conformance_failure_dossiers
     4. provider_compat_matrix
     5. scenario_conformance_suite
     6. auto_repair_full_corpus
     7. differential_suite (optional, enabled via --run-diff; npm diff via --run-npm-diff)
   • Artifacts:
     • tests/ext_conformance/reports/pipeline/full_validation_report.json
     • tests/ext_conformance/reports/pipeline/full_validation_report.md
     • Plus stage-specific reports under tests/ext_conformance/reports/**

3. Run dev-firstset live-provider gate (must pass before release build)

   • Binary: ext_release_binary_e2e
   • Execution authority: DSR quality with its configured live-provider inputs
   • Purpose:
     • Proves the current codepath works end-to-end on a representative first-set before paying release-build cost.
     • Serves as the promotion gate to full release-binary validation.
   • Gate:
     • Require pass=20 / total=20 with fail=0.
   • Artifacts:
     • tests/ext_conformance/reports/release_binary_e2e/ollama_firstset_dev_20260219_jobs10_timeout600.json
     • tests/ext_conformance/reports/release_binary_e2e/ollama_firstset_dev_20260219_jobs10_timeout600.md

4. Run full release-binary live-provider E2E (after step 3 passes)

   • Binary: ext_release_binary_e2e
   • Execution authority: the DSR release-validation lane after DSR quality passes
   • Purpose:
     • Executes target/release/pi directly for each selected extension case.
     • Uses a live provider/model path (default ollama + qwen2.5:0.5b) to exercise non-mocked end-to-end behavior.
     • Emits per-case stdout/stderr captures plus summary artifacts (pi.ext.release_binary_e2e.v1).
   • Artifacts:
     • tests/ext_conformance/reports/release_binary_e2e/ollama_full_release_20260219_jobs10_timeout600.json
     • tests/ext_conformance/reports/release_binary_e2e/ollama_full_release_20260219_jobs10_timeout600.md
     • tests/ext_conformance/reports/release_binary_e2e/cases/*

5. Aggregate and triage

   • full_validation_report.json combines:
     • Stage-level pass/fail (stageSummary, stageResults)
     • Corpus counts (corpus)
     • Vendored conformance totals (conformance)
     • Provider matrix totals (providerCompat)
     • Scenario totals (scenario)
     • Review queue + verdict classification (reviewQueue, verdictCounts)
   • Important interpretation rule:
     • not_tested_unvendored indicates unvendored candidates not yet in vendored conformance; this is inventory status, not a vendored regression.

Recommended run environment

These runs compile many crates and can be disk-heavy. DSR owns target/temp placement, load admission, and any remote offload. Run the registered recipe:

dsr quality --tool pi_agent_rust

Historical run snapshot (extension gate refresh 2026-05-15)

These artifacts predate v0.3.0 and do not certify the current source revision. The strict drop-in program has been retired; its contract and verdict are no longer release gates and will not be rerun to define product completeness. The figures below are retained only as a dated engineering snapshot.

The current extension must-pass corpus is the exact set selected by docs/extension-inclusion-list.json; the historical counts below do not prove coverage of that authoritative list.

From:

• tests/ext_conformance/reports/gate/must_pass_gate_verdict.json (generated 2026-05-15T17:03:02.000Z, run local-20260515T170218075Z)

• tests/ext_conformance/reports/health_delta/health_delta_report.json (generated 2026-05-13T03:37:59.568Z)

• tests/ext_conformance/reports/journeys/journey_report.json (generated 2026-05-13T02:59:58.302Z)

• tests/evidence_bundle/index.json (generated 2026-05-12T19:26:21.441Z, run local-20260512T192621Z)

• tests/full_suite_gate/certification_verdict.json (generated 2026-05-14T19:59:37.227Z)

• Historical verdict blob at 2fc4b8c0b77ded267cf5e0f517f4b6fa87f45e91:docs/evidence/dropin-certification-verdict.json (generated 2026-05-18T19:37:26Z for source 52e9fbfb24352045985b59df9d7ea63f1f8f2ef8; the live file may contain a later verdict)

• Historical strict drop-in result: 22/22 certification gates PASS, 16/16 blocking gates PASS - CERTIFIED for source 52e9fbfb24352045985b59df9d7ea63f1f8f2ef8 only (from the Git-pinned historical verdict blob above; it neither describes the live verdict file nor certifies v0.3.0)

• Unified evidence bundle, as regenerated 2026-08-28: 30 total sections, 18 present, 12 missing, 0 invalid (bundle verdict: partial) (from tests/evidence_bundle/index.json; historical snapshot)

• Extension must-pass gate, as regenerated 2026-08-17: 206/208 must-pass extensions passed (2 failures); informational stretch set 10/19 passed — the May 2026 123/123 snapshot predates the expanded corpus (from tests/ext_conformance/reports/gate/must_pass_gate_verdict.json; historical snapshot)

• Extension health delta, as regenerated 2026-08-17: 226 extensions tested at 95.6% pass rate with 0 regressions vs the 2026-02-07 baseline (from tests/ext_conformance/reports/health_delta/health_delta_report.json; historical snapshot)

• Extension journey coverage, as regenerated 2026-08-17: 125/125 journey scenarios passed (100.0%); command, event-subscriber, multi-capability, passive, and tool-provider categories are green (from tests/ext_conformance/reports/journeys/journey_report.json; historical snapshot)

• Historical stress-triage evidence is retained under tests/perf/reports/; it is not current enough to support a v0.3.0 performance claim.

---

Installation

Curl Installer (Recommended)

# Latest release
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/main/install.sh?$(date +%s)" | bash

# Non-interactive + auto PATH update
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/main/install.sh?$(date +%s)" | bash -s -- --yes --easy-mode

# Pin a release tag
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/v0.3.0/install.sh" | bash -s -- --version v0.3.0

# Install from explicit artifact URL + checksum URL
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/v0.3.0/install.sh" | \
  bash -s -- \
    --artifact-url "https://github.com/Dicklesworthstone/pi_agent_rust/releases/download/v0.3.0/pi-linux-amd64.tar.xz" \
    --checksum-url "https://github.com/Dicklesworthstone/pi_agent_rust/releases/download/v0.3.0/pi-linux-amd64.tar.xz.sha256"

# Skip completion setup (CI/non-interactive minimal install)
curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/main/install.sh?$(date +%s)" | \
  bash -s -- --yes --no-completions

The installer is idempotent and supports a migration path from TypeScript Pi:

• Detect existing TS pi command
• Prompt to install Rust Pi as canonical pi
• Preserve old CLI behind legacy-pi
• Record state for clean uninstall/restore

Notable installer flags:

• --offline [TARBALL]: enforce offline mode; optional local artifact path (.tar.gz, .tar.xz, .zip, or raw binary)
• --artifact-url: force a specific release artifact URL
• --checksum / --checksum-url: override checksum source for explicit artifacts
• --sigstore-bundle-url: override the optional Sigstore bundle URL used by cosign verify-blob; set both COSIGN_IDENTITY_RE and COSIGN_OIDC_ISSUER to the trusted signer policy (there is deliberately no GitHub Actions identity default)
• --completions auto|off|bash|zsh|fish: force shell completion install target (off is equivalent to --no-completions)
• --no-completions: disable completion installation
• --no-agent-skills: skip automatic installation of the pi-agent-rust skill into ~/.claude/skills/ and ~/.codex/skills/
• --no-verify: skip checksum + signature verification (testing only)
• --artifact-url without --version uses a synthetic tag for release mode only; if artifact download fails, install exits instead of attempting source fallback
• Installer honors HTTPS_PROXY / HTTP_PROXY for all network fetches

By default, the installer also installs a pi-agent-rust skill for both Claude Code and Codex CLI:

• Claude Code: ~/.claude/skills/pi-agent-rust/SKILL.md
• Codex CLI: ~/.codex/skills/pi-agent-rust/SKILL.md (or $CODEX_HOME/skills/pi-agent-rust/SKILL.md if CODEX_HOME is set)
• During upgrades, installer-managed legacy pre-tool entries from older versions are removed automatically (idempotent, path-scoped, and non-destructive) when prior installer state is present.

Installer regression harness (options + checksum + signature + completions):

bash tests/installer_regression.sh

Distribution Compatibility Contract (Packaging/Invocation Scope)

For migration adoption, packaging and invocation compatibility follows this contract:

• This section covers packaging and invocation behavior only. Historical drop-in/parity artifacts do not define the current product or release gate. Current compatibility statements must name the concrete behavior they cover.

• Canonical executable name is pi across release assets and installer-managed installs.

• Installer-managed installs also create an rpi compatibility launcher when no conflicting rpi command already exists on your PATH.

• Existing TypeScript pi installs can be migrated in place; the prior command is preserved as legacy-pi.

• If you keep TypeScript pi as canonical (--keep-existing-pi), Rust Pi is installed as pi-rust.

• On Apple Silicon, the installer prefers the native arm64 artifact even when launched from a Rosetta-translated shell.

• Version-pinned installs are supported via install.sh --version vX.Y.Z for deterministic rollouts.

• DSR releases ship each platform archive with a same-name .sha256 sidecar for integrity validation. The v0.3.0 release predates that contract and ships an aggregate SHA256SUMS instead; the installer verifies the per-asset sidecar first and falls back to SHA256SUMS only when the sidecar is absent.

Representative smoke checks:

# Canonical command should exist and execute
command -v pi
pi --version
pi --help >/dev/null

# If a TS migration was performed, legacy command remains available
command -v legacy-pi && legacy-pi --version

Source builds

Repository builds are tested with the exact toolchain pinned in rust-toolchain.toml (nightly-2026-08-31). The locked dependency graph requires Rust 1.95 or newer. Project builds are DSR-only:

dsr quality --tool pi_agent_rust
dsr build pi_agent_rust

End users should install a DSR-published archive through install.sh. There is no supported direct-Cargo contributor or installation lane.

Dependencies

Pi has no required external runtime dependencies: the grep and find tools search in-process using the same engines ripgrep is built from (grep-searcher/grep-regex and the ignore walker), so patterns keep ripgrep’s default (Rust regex crate) syntax. PCRE2-only features (lookaround, backreferences) were never exposed and remain unsupported.

To shell out to rg/fd instead (debugging escape hatch), set "search_backend": "external" in settings.json and install them via apt install ripgrep fd-find or brew install ripgrep fd.

Uninstall

curl -fsSL "https://raw.githubusercontent.com/Dicklesworthstone/pi_agent_rust/main/uninstall.sh" | bash

By default, uninstall removes installer-managed Rust binaries/aliases and skill directories, then restores a migrated TypeScript pi if one was preserved.

---

Commands

Basic Usage

pi [OPTIONS] [MESSAGE]...

# Examples
pi                              # Start interactive session
pi "Hello"                      # Start with message
pi @file.rs "Explain this"      # Include file as context
pi @voice.wav @clip.mp4 "Compare these recordings"
pi -p "Quick question"          # Print mode (no session)

Interactive file references:

• Type @relative/path in the editor to attach a file’s contents (autocomplete inserts the @ form).
• Images and supported audio/video files become native attachments in file-argument order. Audio/video extensions are .mp4, .mov, .webm, .mp3, .wav, .flac, .ogg, and .m4a (case-insensitive), with a 5 MiB per-file limit. Oversized files produce a visible notice and empty files are skipped.
• File arguments use the current workspace roots, including roots added with --add-dir or /add-dir; /remove-dir revokes subsequent attachment reads. CLI startup files and messages are submitted by both terminal frontends, and FTUI /retry retains the original attachment bytes.
• Gemini-family providers receive native audio/video. Providers without that transport receive a visible media-omission label; session history retains the attachment. See native RPC content and SDK native prompts.

Options

Additional high-leverage flags:

• --no-migrations to skip startup migration checks
• --explain-extension-policy to print effective capability decisions and exit
• --explain-repair-policy to print effective repair-policy resolution and exit

Subcommands

# Package management
pi install  [-l|--local]    # Install a package source and add to settings
pi remove  [-l|--local]     # Remove a package source from settings
pi update [source]                 # Update all (or one) non-pinned packages
pi list                            # List user + project packages from settings

# Configuration
pi config                          # Show settings paths + precedence

More utility subcommands:

# Extension catalog index + discovery
pi update-index
pi search "git"
pi info pi-search-agent

# Environment and extension diagnostics
pi doctor
pi doctor --only sessions --format json
pi doctor --only swarm --format json
pi doctor ./path/to/extension --policy safe --fix

# Read-only swarm progress SLO evaluation from normalized evidence
pi swarm-progress --input progress-slo-input.json --format json
pi swarm-progress --input progress-slo-input.json --since HEAD~1 --out-json progress-slo.json

# Session storage migration (JSONL -> v2 sidecar store)
pi migrate ~/.pi/agent/sessions --dry-run
pi migrate ~/.pi/agent/sessions

• update-index refreshes extension index metadata used by search and info.
• search and info let you discover and inspect extension metadata without leaving the CLI.
• pi doctor checks ordinary installation health: config, directories, auth, shell setup, and sessions. Swarm coordination diagnostics are opt-in with --only swarm (or a comma-separated selection including swarm); they require development infrastructure such as Beads, Agent Mail, RCH, and build scratch space. pi doctor --only swarm --format json reports cgroup CPU quota, cpuset size, NUMA topology, cgroup memory limits, target/tmp headroom, and recommended concurrency budgets before large multi-agent runs. These explicit checks remain strict and fail closed on operational hazards. Text, JSON, and markdown reports use the same selected findings and exit with status 1 when any selected check fails; warnings and informational findings exit 0. Supplying an extension path checks extension compatibility by default.
• swarm-progress evaluates a normalized progress SLO snapshot and emits advisory JSON/text only; it does not mutate Beads, git, Agent Mail, RCH, validation broker slots, runpacks, or source files. Operator workflow, privacy boundaries, degraded Agent Mail/RCH interpretation, stale-Beads handling, and no-open-work convergence guidance lives in docs/swarm-operations-runbook.md#progress-slo-operator-workflow.
• migrate validates or creates the v2 session sidecar format for faster resume on larger histories.

---

Configuration

Pi reads configuration from ~/.pi/agent/settings.json. Keys are shown in their canonical snake_case form below, but every field also accepts the original TypeScript Pi’s camelCase spelling (e.g. defaultProvider, shellCommandPrefix) as a serde alias, so an existing pi-mono settings.json is parsed as-is and is never rewritten to snake_case:

{
  "default_provider": "anthropic",
  "default_model": "claude-opus-4-5",
  "default_thinking_level": "medium",

  "compaction": {
    "enabled": true,
    "reserve_tokens": 8192,
    "keep_recent_tokens": 20000
  },

  "retry": {
    "enabled": true,
    "max_retries": 3,
    "base_delay_ms": 1000,
    "max_delay_ms": 30000
  },

  "images": {
    "auto_resize": true,
    "block_images": false
  },

  "terminal": {
    "show_images": true,
    "clear_on_shrink": false
  },

  "http": {
    "proxy": "http://127.0.0.1:2080",
    "no_proxy": ["localhost", "127.0.0.1", ".internal.example"]
  },

  "shell_path": "/bin/bash",
  "shell_command_prefix": "set -e"
}

Network Proxy

Every request Pi makes — provider APIs, OAuth logins, update checks, URL reads, package fetches — goes through the proxy resolved here. https:// targets use a CONNECT tunnel, so TLS stays end-to-end to the origin and the proxy never sees plaintext; http:// targets use an absolute-form request line. Credentials in the proxy URL (http://user:pass@host:port) become Proxy-Authorization, and are never logged or passed to child processes.

Resolution order for a request, first match wins:

1. A bypass match — http.no_proxy in settings.json, else NO_PROXY / no_proxy. Entries follow the usual convention: * bypasses everything, example.com or .example.com matches the domain and its subdomains, and example.com:8443 restricts the match to that port. Loopback destinations also bypass the proxy by default, unless http.proxyLoopback is enabled.
2. http.https_proxy / http.http_proxy (scheme-specific settings).
3. http.proxy (both schemes).
4. PI_HTTPS_PROXY / PI_HTTP_PROXY.
5. The standard HTTPS_PROXY, HTTP_PROXY, ALL_PROXY (lowercase spellings accepted).

Ambient proxy variables are honored by default, the same as git and curl. Where they are set for some other tool — a capture proxy, a stale VPN helper — turn the inheritance off with "http": { "ignore_env_proxy": true } or PI_HTTP_PROXY=off; explicit settings and PI_*_PROXY still apply. Proxy endpoints may be http://, socks5:// (local DNS) or socks5h:// (DNS on the proxy), with optional username/password authentication; SOCKS credentials are not encrypted on the hop to the proxy. Requests to localhost (case-insensitive, with an optional trailing dot), literal IPv4 addresses in 127.0.0.0/8, and IPv6 ::1 or IPv4-mapped loopback addresses go directly to the local server. This keeps Ollama and LM Studio reachable when a global proxy is configured. Pi makes this decision without DNS lookups; private-network addresses and other hostnames retain the configured proxy routing.

To intentionally send loopback destinations through the proxy, such as to reach a service on the proxy host or capture local traffic, set "http": { "proxyLoopback": true }. Explicit NO_PROXY / http.noProxy matches still take precedence. An unusable ambient value (e.g. an https:// or socks4:// endpoint) is skipped with a warning rather than failing requests.

The bash tool’s child environment overrides advertise only HTTP proxy endpoints (as HTTP_PROXY/HTTPS_PROXY/NO_PROXY and their lowercase spellings). Credentials are deliberately omitted from that per-child copy, so authenticated HTTP proxies need their own configuration in those tools (e.g. git config --global http.proxy). Pi’s own environment and the system environment are never modified. SOCKS endpoints leave child proxy variables as inherited, including authenticated ALL_PROXY values and NO_PROXY. This preserves existing curl/git SOCKS configuration without forcing tools that lack SOCKS support to use it through HTTP-specific variables. A SOCKS proxy set only in Pi settings or PI_*_PROXY does not configure child tools; use their standard proxy environment variables or their own configuration. The automatic loopback bypass applies to Pi’s own HTTP requests; child tools continue to follow their own proxy and bypass configuration.

Configuration Precedence

Settings are resolved in priority order (first match wins):

1. CLI flags (--model, --thinking, --provider, etc.)
2. Environment variables (ANTHROPIC_API_KEY, PI_CONFIG_PATH, etc.)
3. Project settings (.pi/settings.json in the working directory)
4. Global settings (~/.pi/agent/settings.json)
5. Built-in defaults

This means a CLI flag always overrides a settings.json value, and a trusted project-level setting overrides the global one. Untrusted project settings are excluded. Optional-tool activation flags merge individually; a project can set one flag to false while inheriting the other global flags.

Resource Resolution

Resource resolution gives explicit CLI paths priority, followed by CLI -e sources, configured project resources, configured user resources, and installed packages. Project packages take priority over user packages.

Skills additionally discover universal .agents/skills/ directories from the working directory through its ancestors up to the Git root (or filesystem root when no Git root exists), plus ~/.agents/skills/ in the user’s home directory. Ancestor probing is capped at 64 locations and each selected skill root has a 16,384-entry discovery budget; an over-budget root is rejected rather than partially loaded. Within each scope, Pi-specific skills and settings entries take priority over universal skills; nearer project universal roots take priority over farther ancestors, and project universal skills still take priority over user skills. Managed skills from /skills.managed/ load last. See the complete skill precedence table.

When multiple resources share the same name, the first occurrence wins. Collisions are logged as diagnostics.

--no-skills disables automatic discovery (including universal and managed skills), installed package skills, and skills entries in settings.json. Explicit CLI --skill paths and skills supplied by CLI -e sources still load, so pi --no-skills --skill /path/to/skill-a --skill /path/to/skill-b selects an isolated skill set. The sibling --no-prompt-templates, --no-themes, and --no-extensions flags likewise preserve explicit CLI resources while disabling configured discovery for their categories.

Project context files are a separate switch. By default pi appends AGENTS.md / CLAUDE.md from ~/.pi/agent/, the working directory, and every ancestor directory to the system prompt (plus imported foreign-format workspace rules). --no-context-files (or PI_NO_CONTEXT_FILES=1) disables that discovery entirely; --no-skills does not cover it. Hosts that compose the whole prompt with --system-prompt typically pass --no-context-files --no-skills together.

Prompt template expansion supports positional arguments: $1, $2, $@ (all args), and slice syntax ${@:start}, ${@:start:length} (or OMP’s $@[start:length]). For example, a template invoked as /review src/main.rs --strict receives src/main.rs as $1 and --strict as $2. A template with no placeholder gets its arguments appended after a blank line.

Environment Variables

---

Architecture

┌─────────────────────────────────────────────────────────────────┐
│                           CLI (clap)                            │
│  • Argument parsing    • @file expansion    • Subcommands       │
└─────────────────────────────────┬───────────────────────────────┘
                                  │
┌─────────────────────────────────▼───────────────────────────────┐
│                          Agent Loop                             │
│  • Message history     • Tool iteration    • Event callbacks    │
└────────┬──────────────────────┬──────────────────────┬──────────┘
         │                      │                      │
┌────────▼────────┐  ┌─────────▼──────────┐  ┌───────▼──────────┐
│ Provider Layer  │  │  Tool Registry     │  │  Extension Mgr     │
│ • Anthropic     │  │  • read  • bash    │  │  • QuickJS JS/TS   │
│ • OpenAI (Chat/ │  │  • write • grep    │  │  • Native descriptor│
│   Responses)    │  │  • edit  • find    │  │    runtime          │
│ • Gemini/Cohere │  │  • ls              │  │  • Capability policy│
│ • Azure/Bedrock │  │  • ext-registered  │  │  • Node shims       │
│ • Vertex/Copilot│  │                    │  │  • Event hooks      │
│ • GitLab/Cursor │  │                    │  │  • Runtime risk ctl │
│   /Ext          │  │                    │  │                    │
└────────┬────────┘  └─────────┬──────────┘  └───────┬──────────┘
         │                     │                      │
┌────────▼─────────────────────▼──────────────────────▼──────────┐
│                     Session Persistence                         │
│  • JSONL format (v3)   • Tree structure   • Session index/cache  │
│  • Per-project dirs    • Default-enabled SQLite backend support │
└─────────────────────────────────────────────────────────────────┘

Provider-count rule: Pi has 11 native provider implementation modules. Those modules are anthropic, openai, openai_responses, gemini, cohere, azure, bedrock, vertex, copilot, gitlab, and cursor. User-visible provider IDs, aliases, OpenAI-compatible presets, and extension-provided streamSimple providers are counted separately because several native modules expose multiple routes. The factory module mod.rs and shared live-catalog helper model_fetch.rs are support modules, not provider implementations, and are excluded.

Key Design Decisions

1. No unsafe code: #![forbid(unsafe_code)] enforced project-wide
2. Streaming-first: Custom SSE parser, no blocking on responses
3. Process tree management: sysinfo crate ensures no orphaned processes
4. Structured errors: thiserror with specific error types per component
5. Size-budgeted release profile: LTO + strip + opt-level = "z" for the size budget; published binaries use the speed-optimized dist profile (opt-level = 3)

asupersync Context vs TypeScript Pi (pi-mono)

This Rust port preserves Pi’s user experience, but intentionally changes the runtime substrate. The original TypeScript Pi (pi-mono, packages/coding-agent) is built on Node.js + package-level abstractions. pi_agent_rust moves those same behaviors onto asupersync primitives so lifecycle guarantees are explicit in the runtime model.

Why this is useful in practice:

• More predictable failure behavior during aborts/timeouts because cancellation is checked in explicit loop boundaries and tool runners.
• Cleaner resource lifetimes because the runtime, timers, and I/O paths all share one concurrency substrate.
• Less hidden coupling because the main invariants live in Rust types/algorithms rather than spread across framework conventions.

Runtime Invariants (and Why They Matter)

These are the concrete invariants we rely on in this implementation:

1. Turn-scoped agent lifecycle

   • The main loop emits AgentStart, TurnStart, TurnEnd, and AgentEnd in a stable order.
   • Tool recursion is bounded by max_tool_iterations (default 50) to avoid unbounded self-tool loops.
   • Benefit: stable event ordering for TUI/RPC consumers and predictable termination behavior.

2. Abort and timeout behavior is explicit

   • Agent abort checks happen at turn boundaries and around tool execution.
   • bash timeout follows a clear escalation path: terminate process tree, grace period, then hard kill.
   • Benefit: fewer “hung” sessions and reduced orphan-process risk during aggressive tool use.

3. Session writes are crash-resilient

   • JSONL saves write to a temp file and persist atomically.
   • Session indexing uses SQLite WAL + lock file coordination for concurrent instances.
   • Benefit: better durability and resume reliability under multi-process usage.

4. Compaction is threshold-driven and boundary-aware

   • Trigger: estimated context tokens exceed context_window - reserve_tokens.
   • Cut-point logic prefers user-turn boundaries and preserves recent context budget.
   • Benefit: compaction recovers context without collapsing near-term task continuity.

5. Capability policy is fail-closed and precedence-defined

   • Resolution order: per-extension deny -> global deny -> per-extension allow -> default caps -> mode fallback.
   • Benefit: policy outcomes are explainable, deterministic, and auditable.

6. Streaming parser tolerates real network chunking

   • SSE parser handles CR/LF variants, multi-line data: fields, partial UTF-8 tails, and end-of-stream flush.
   • Benefit: incremental rendering remains robust across providers and network fragmentation.

Design Principles Carried From asupersync Into Pi

The following asupersync principles are reflected directly in pi_agent_rust architecture:

• Single async substrate: runtime, timers, fs, and HTTP/TLS all run on one coherent foundation.
• Explicit context threading: AgentCx wraps asupersync::Cx at subsystem boundaries (agent/tools/session/rpc).
• Bounded operations over best-effort cleanup: timeout paths and compaction thresholds are parameterized and enforceable.
• Determinism hooks for tests: timer-driver aware sleeps and asupersync test helpers reduce nondeterministic flakiness.

Compared to the original TypeScript implementation, this shifts more correctness responsibility into the runtime and core algorithms themselves, instead of relying primarily on ecosystem conventions.

Additional Major Deltas (Original pi-mono vs Rust Port)

This is a second comparison pass focused on high-impact architectural deltas and rationale.

Practical consequence of these deltas:

• Extension and package workflows are selected according to current user value, with OMP as the closer product-surface and UX reference. Legacy pi-mono behavior may be adopted selectively when it remains useful.
• The goal is a coherent Rust-native coding agent, not parity with pi-mono.
• The Rust SDK provides a companion API for core embedding workflows without requiring TypeScript-specific adaptation patterns.
• docs/parity-certification.json tracks informational functional-parity progress; it does not authorize strict replacement claims.

Algorithmic Mechanics: pi-mono Baseline vs Rust Implementation

This section compares concrete implementation mechanics for equivalent high-level behavior.

Feature Superset Highlights (Beyond pi-mono Baseline)

The sections above compare mechanics. This section calls out concrete features present in this Rust port that are not part of the pi-mono baseline implementation model.

---

Deep Dive: Core Algorithms

Math-Driven Decision Systems

Pi deliberately uses advanced math where it improves runtime behavior or benchmark confidence. The goal is not “fancy formulas in docs”; it is safer policy decisions, faster recovery from workload shifts, and more trustworthy performance attribution.

Regime-Shift Detection (CUSUM + BOCPD)

In the extension dispatcher, Pi combines CUSUM and Bayesian online change-point detection to detect load-regime changes early (for example when hostcall traffic suddenly spikes or stalls).

$$ S_t^+ = \max\left(0,;S_{t-1}^+ + (-z_t - k)\right), \quad S_t^- = \max\left(0,;S_{t-1}^- + (z_t - k)\right) $$

$$ H®=\frac{1}{\lambda}, \quad P(r_t=0 \mid x_{1:t}) \propto \sum_r P(r_{t-1}=r),H®,P(x_t \mid r) $$

Intuition: CUSUM catches persistent drift; BOCPD catches sudden regime changes without brittle fixed thresholds.

Conformal Prediction Envelope

Pi tracks nonconformity scores (absolute residuals from the running mean) and treats out-of-interval events as anomalies.

$$ q = \text{score}_{\lceil (n+1)\cdot \text{confidence} \rceil - 1}, \quad \text{anomaly if } |x_t - \mu_t| > q $$

Intuition: thresholds adapt from recent behavior instead of hard-coding one static latency cutoff.

PAC-Bayes Safety Bound

Pi’s safety envelope includes a PAC-Bayes-kl bound over extension outcomes, and can veto aggressive optimization when the bound is too high.

$$ \mathrm{kl}(\hat q ,|, q_{\text{bound}});\le;\frac{\mathrm{KL}(Q|P)+\ln!\left(2\sqrt{n}/\delta\right)}{n} $$

Intuition: this gives an explicit uncertainty-aware ceiling on true error risk before allowing more aggressive runtime behavior.

Off-Policy Evaluation (IPS/WIS/DR + ESS + Regret Gate)

Before approving policy moves, Pi evaluates candidate behavior from trace data:

$$ w_i=\frac{\pi(a_i\mid x_i)}{\mu(a_i\mid x_i)}, \quad \hat V_{\text{IPS}}=\frac{1}{n}\sum_i w_i r_i $$

$$ \hat V_{\text{WIS}}=\frac{\sum_i w_i r_i}{\sum_i w_i}, \quad \hat V_{\text{DR}}=\frac{1}{n}\sum_i\left(\hat r_i + w_i(r_i-\hat r_i)\right) $$

$$ N_{\text{eff}}=\frac{(\sum_i w_i)^2}{\sum_i w_i^2}, \quad \Delta_{\text{regret}}=\bar r_{\text{baseline}}-\hat V_{\text{DR}} $$

Intuition: Pi fails closed if sample support is weak, uncertainty is high, or estimated regret is above threshold.

VOI-Driven Experiment Selection

The VOI planner prioritizes probes that provide the most expected learning under a strict overhead budget.

$$ \text{priority}_i \propto \frac{\text{utility}_i}{\text{overhead}_i} $$

Intuition: run only the experiments that are likely to change decisions; skip stale or low-value probes.

Weighted Bottleneck Attribution (Benchmarking)

For phase-1 matrix benchmarking, Pi computes stage attribution weighted by realistic workload size (session_messages) and reports confidence intervals.

$$ \text{weighted_contribution}_s

\frac{\sum_i w_i,m_{i,s}}{\sum_i w_i,t_i}\cdot 100, \quad w_i=\text{session_messages}_i $$

$$ n_{\text{eff}}=\frac{(\sum_i w_i)^2}{\sum_i w_i^2}, \quad \mathrm{CI}{95}=\mu \pm 1.96\sqrt{\frac{\sigma^2}{n{\text{eff}}}} $$

Intuition: prioritize what dominates real end-to-end latency, not just isolated microbench hotspots.

Online Convex Control + Regret Tracking

Pi also includes an online tuner path for batch/time-slice controls with explicit rollback behavior:

$$ \tau_{t+1}

\mathrm{clip}!\left(\tau_t - \eta\nabla_{\tau}\mathcal{L}t,;\tau{\min},\tau_{\max}\right) $$

Intuition: the system adapts continuously, but if instantaneous loss exceeds a rollback threshold it immediately returns to a safer profile.

Math At a Glance

SSE Streaming Parser

The SSE (Server-Sent Events) parser is a custom implementation that handles Anthropic’s streaming response format. Unlike library-based approaches, the parser operates as a state machine that processes bytes incrementally:

Bytes → Line Accumulator → Event Parser → Typed StreamEvent

Key characteristics:

The parser handles edge cases like:

• Multi-line data: fields (concatenated with newlines)
• Events split across TCP packet boundaries
• The event: field appearing before or after data:
• CRLF and LF line endings interchangeably

Truncation Algorithm

Large outputs from tools (file reads, command output, grep results) must be truncated to avoid exhausting the LLM’s context window. The truncation algorithm preserves usefulness while staying within limits:

┌─────────────────────────────────────────┐
│           Original Content              │
│         (potentially huge)              │
└─────────────────────────────────────────┘
                    │
                    ▼
┌─────────────────────────────────────────┐
│  HEAD: First N/2 lines                  │
│  ─────────────────────────              │
│  [... X lines truncated ...]            │
│  ─────────────────────────              │
│  TAIL: Last N/2 lines                   │
└─────────────────────────────────────────┘

Constants:

The algorithm:

1. Splits content into lines
2. If line count exceeds MAX_LINES, takes first 1000 and last 1000
3. Inserts a marker showing how many lines were omitted
4. If byte count still exceeds MAX_BYTES, applies byte-level truncation
5. Returns metadata indicating truncation occurred, enabling the LLM to request specific ranges

Process Tree Management

The bash tool must handle runaway processes, infinite loops, and fork bombs without leaving orphans. The implementation uses the sysinfo crate to walk the process tree:

// Pseudocode for process cleanup
fn kill_process_tree(root_pid: Pid) {
    let system = System::new();
    let children = find_all_descendants(root_pid, &system);

    // Kill children first (deepest first), then parent
    for child in children.iter().rev() {
        kill(child, SIGKILL);
    }
    kill(root_pid, SIGKILL);
}

Timeout behavior:

1. Command starts with configurable timeout (default 120s)
2. Output streams to a rolling buffer in real-time
3. On timeout: SIGTERM sent, 5s grace period, then SIGKILL
4. Process tree walked and all descendants killed
5. Exit code set to indicate timeout vs. normal termination

To avoid orphaned background jobs (e.g. cmd &), the bash script installs an EXIT trap that waits for any remaining child processes and then exits with the original command’s status.

This prevents the common failure mode where killing a shell leaves its children running.

Session Tree Structure

Sessions use a tree structure rather than a flat list, enabling conversation branching (useful when exploring different approaches):

                    ┌─────────┐
                    │ Message │ (root)
                    │   #1    │
                    └────┬────┘
                         │
                    ┌────▼────┐
                    │ Message │
                    │   #2    │
                    └────┬────┘
                         │
              ┌──────────┼──────────┐
              │                     │
         ┌────▼────┐          ┌────▼────┐
         │ Message │          │ Message │ (branch)
         │   #3    │          │   #3b   │
         └────┬────┘          └────┬────┘
              │                    │
         ┌────▼────┐          ┌────▼────┐
         │ Message │          │ Message │
         │   #4    │          │   #4b   │
         └─────────┘          └─────────┘

JSONL format (v3):

Each line is a self-contained JSON object with a type discriminator:

{"type":"session","version":3,"cwd":"/project","created":"2024-01-15T10:30:00Z"}
{"type":"message","id":"a1b2c3d4","parent":"root","role":"user","content":[...]}
{"type":"message","id":"e5f6g7h8","parent":"a1b2c3d4","role":"assistant","content":[...]}
{"type":"model_change","id":"i9j0k1l2","parent":"e5f6g7h8","model":"claude-sonnet-4-20250514"}

The parent field creates the tree. Replaying a session walks the tree from root to the current leaf. Branching creates a new message with a different parent than the previous continuation.

Provider Abstraction

The Provider trait abstracts over different LLM backends:

#[async_trait]
pub trait Provider: Send + Sync {
    fn name(&self) -> &str;
    fn api(&self) -> &str;
    fn model_id(&self) -> &str;

    async fn stream(
        &self,
        context: &Context,
        options: &StreamOptions,
    ) -> Result> + Send>>>;
}

Context structure:

pub struct Context {
    pub system_prompt: Option>, // System prompt
    pub messages: Cow,        // Conversation history
    pub tools: Cow,           // Available tools with JSON schemas
}

StreamOptions:

pub struct StreamOptions {
    pub temperature: Option,
    pub max_tokens: Option,
    pub api_key: Option,
    pub cache_retention: CacheRetention,
    pub session_id: Option,
    pub headers: HashMap,
    pub thinking_level: Option,
    pub thinking_budgets: Option,
}

This design allows adding new providers (OpenAI, Gemini) without modifying the agent loop. Each provider translates the common types to its wire format and emits a unified StreamEvent stream.

Compaction Algorithm

Long conversations eventually exceed the model’s context window. Pi’s compaction algorithm reclaims space by summarizing older messages while preserving recent context.

The algorithm runs automatically after each agent turn when estimated token usage exceeds context_window - reserve_tokens:

┌──────────────────────────────────────────────────────────────┐
│                     Full Conversation                         │
│  msg1 → msg2 → msg3 → ... → msgN-5 → msgN-4 → ... → msgN   │
│  ├──── older messages ─────┤ ├─── recent messages ──────────┤ │
│                                                              │
│  Step 1: Find cut point at a valid turn boundary             │
│  Step 2: LLM summarizes msgs 1..N-5 into compact paragraph  │
│  Step 3: Store Compaction entry in session JSONL             │
│  Step 4: Next agent call uses [summary] + msgs N-4..N       │
└──────────────────────────────────────────────────────────────┘

Token estimation counts tokens with real BPE tables (O200k/Cl100k, enabled by the default bpe-tokens feature); when BPE is unavailable it falls back to a conservative chars ÷ 3 heuristic for text, plus a flat 1,200 tokens per image. When an assistant message includes a usage field from the API, that measured value takes precedence over the estimate.

Cut point selection prefers boundaries between complete user-assistant turns. If the budget forces a mid-turn cut, the algorithm includes prefix messages from the split turn so the model retains context about what was being discussed at the boundary.

File operation tracking extracts read, write, and edit tool calls from the messages being summarized. The compaction prompt includes these paths so the summary preserves awareness of which files were examined or modified:

src/main.rs
src/config.rs

src/auth.rs

Configurable parameters:

Compaction can also be triggered manually with /compact in interactive mode or the compact RPC command.

Multi-Provider Routing & Model Registry

Pi routes model requests through a provider factory that resolves the correct backend implementation from a (provider, model, api) tuple.

Resolution flow:

User specifies --provider openai --model gpt-4o
               │
               ▼
  ┌──────────────────────────┐
  │  Provider Metadata Table  │  Maps "openai" → canonical ID,
  │                           │  determines API type (Completions
  │                           │  vs Responses vs custom)
  └────────────┬──────────────┘
               │
  ┌────────────▼──────────────┐
  │  URL Normalization         │  Appends /chat/completions,
  │                            │  /responses, or /chat depending
  │                            │  on detected API type
  └────────────┬──────────────┘
               │
  ┌────────────▼──────────────┐
  │  Compat Config             │  Applies per-model overrides:
  │                            │  system_role_name, max_tokens
  │                            │  field name, feature flags
  └────────────┬──────────────┘
               │
  ┌────────────▼──────────────┐
  │  Provider Instance         │  Anthropic | OpenAI | Gemini
  │                            │  Cohere | Azure | Bedrock | ...
  └───────────────────────────┘

models.json overrides: Users can define custom providers in ~/.pi/agent/models.json or .pi/models.json. Each entry specifies a model ID, base URL, API type, and optional compat flags, letting you route to self-hosted models, proxies, or providers that Pi does not natively support.

Compat config handles the differences between OpenAI-compatible APIs:

Fuzzy matching: When a provider name doesn’t match any known provider, Pi computes edit distance against all registered names and suggests the closest match in the error message.

Extension Hostcall Protocol

Extensions run in an embedded QuickJS runtime (rquickjs crate) and communicate with Pi through a structured hostcall protocol. This is the mechanism that lets JavaScript code invoke Pi’s built-in tools, make HTTP requests, and interact with the session, all without direct OS access.

Execution model:

┌─────────────────── QuickJS VM ───────────────────┐
│                                                   │
│  extension.js calls:                              │
│    pi.tool("read", {path: "src/main.rs"})         │
│      │                                            │
│      ▼                                            │
│    enqueue HostcallRequest {                      │
│      call_id: "hc-0042",                          │
│      kind: Tool { name: "read" },                 │
│      payload: { path: "src/main.rs" },            │
│    }                                              │
│      │                                            │
│      ▼                                            │
│    return Promise (resolve/reject stored in map)  │
│                                                   │
└────────────────────────┬──────────────────────────┘
                         │
    drain_hostcall_requests()
                         │
                         ▼
┌─────────────── ExtensionDispatcher ──────────────┐
│                                                   │
│  1. Check capability policy:                      │
│     read tool → requires "read" capability        │
│     → Policy says: Allow / Deny / Prompt          │
│                                                   │
│  2. If allowed → dispatch to ToolRegistry         │
│     → Execute read tool                           │
│     → Get ToolOutput                              │
│                                                   │
│  3. complete_hostcall("hc-0042", Ok(result))      │
│     → Resolves the Promise in QuickJS             │
│                                                   │
│  4. runtime.tick()                                │
│     → Drains Promise .then() chains               │
│     → Extension JS continues execution            │
│                                                   │
└───────────────────────────────────────────────────┘

Capability mapping: Each hostcall kind maps to a required capability:

Deduplication: Each hostcall’s parameters are canonicalized (object keys sorted, structure normalized) and SHA-256 hashed. Identical requests within a short window can be deduplicated to avoid redundant tool executions.

Fast lane vs compatibility lane: Pi has two execution lanes for hostcalls:

• Fast lane is used when the call shape matches known safe patterns (for example common tool and session operations). This avoids extra allocation and parsing work.
• Compatibility lane is the fallback for uncommon or partially-specified calls.
• Both lanes still enforce the same capability policy and permission checks.
• Operators can force compatibility-lane routing globally or per extension as an emergency control path.

For observability, each call is tagged with a stable lane key (for example tool|tool.read|filesystem or tool|fallback|filesystem) so latency and failure trends can be grouped consistently.

Built-in consistency guard (shadow dual execution): Pi can sample a small subset of read-only hostcalls, execute them through both lanes, and compare canonical output fingerprints. If divergence crosses a configured budget, Pi automatically backs off the fast lane for a period. This gives performance wins without silently changing behavior.

Adaptive dispatch mode under load: Pi can switch between:

• sequential_fast_path for simpler/low-contention workloads
• interleaved_batching when contention and queue pressure rise

Mode changes are gated by sample coverage and risk checks, so Pi does not switch based on thin or cherry-picked evidence.

Runtime telemetry for debugging and tuning: Pi records structured hostcall telemetry (pi.ext.hostcall_telemetry.v1) with lane choice, fallback reason, dispatch latency share, marshalling path, and optimization hit/miss fields. This is used by perf reports and reliability diagnostics.

Auto-repair pipeline: When an extension fails to load or produces runtime errors, Pi’s repair system can automatically fix common issues:

Compatibility scanner: Before loading, Pi statically analyzes extension source code for imports, require() calls, and forbidden patterns (eval, Function(), process.binding, dlopen). The scan produces a capability evidence ledger that informs policy decisions.

Environment variable filtering: Extensions calling pi.env() hit a blocklist that denies access to API keys, credentials, tokens, and private keys. The filter blocks exact matches (ANTHROPIC_API_KEY, AWS_SECRET_ACCESS_KEY), suffix patterns (*_API_KEY, *_SECRET, *_TOKEN), and prefix patterns (AWS_SECRET_*, AWS_SESSION_*). Variables whose names do not match any secret pattern — including most PI_* variables — are served normally; there is no unconditional PI_* exemption, so a name like PI_EXAMPLE_API_KEY is blocked by the suffix rule.

Trust lifecycle and kill switch: Extension trust state is tracked explicitly (pending, acknowledged, trusted, killed). A kill switch demotes an extension to killed, quarantines it in the runtime risk controller, emits a critical alert, and writes an audit record. Lifting the switch requires an explicit operator action and moves the extension back to acknowledged.

Extension Runtime Decision Logic (Plain English)

The extension runtime includes a few small decision engines so behavior stays stable as workload patterns change:

• Value-of-information planner (VOI): Ranks candidate probes by “expected learning per millisecond” and picks the best set under a strict overhead budget. Stale or low-value candidates are skipped with explicit reasons.
• Shard load controller: Adjusts routing weights, batch budgets, and backoff/help factors based on queue pressure, latency, and starvation risk. Damping and oscillation guards prevent overreaction.
• Policy safety evaluator: Replays historical samples with multiple estimators and only approves a policy when sample support is strong, uncertainty is low, and predicted regret stays within limit.

These pieces are intentionally conservative: if confidence is weak, Pi holds steady instead of making an aggressive switch.

Interactive TUI Architecture

The default interactive stack is FrankenTUI (src/interactive_ftui.rs, feature ftui, on by default since the 2026-08-25 cutover). It keeps the Elm Architecture (Model-Update-View): a driver thread owns an asupersync runtime plus an SDK agent session, agent events arrive through an AgentEventSubscription, and PiFtuiModel renders header, markdown conversation, status line, growing editor, and footer regions with tail-follow scrolling, per-entry render caching, inline ask cards, and modal overlays. All agent- and tool-originated text is sanitized before it reaches a frame. pi --inline draws the UI at the bottom of the terminal and preserves shell scrollback.

The previous stack, built on the charmed_rust library family (a Rust port of Go’s Bubble Tea), lives in src/interactive.rs and is still selectable with pi --classic until it is deleted. The diagram below describes that classic stack; the FrankenTUI stack keeps the same agent/UI split and the same PiMsg event vocabulary.

Component stack (classic --classic stack):

┌────────────────────────────────────────────────────┐
│                 Terminal (crossterm)                │
│  Raw mode │ Alt screen │ Keyboard/Mouse events      │
└──────────────────────┬─────────────────────────────┘
                       │
┌──────────────────────▼─────────────────────────────┐
│             bubbletea Program Loop                  │
│  Init() → Update(Msg) → View() → render cycle      │
└──────────────────────┬─────────────────────────────┘
                       │
┌──────────────────────▼─────────────────────────────┐
│                  PiApp (Model)                      │
│                                                     │
│  ┌─────────────┐ ┌──────────────┐ ┌─────────────┐  │
│  │  TextArea    │ │  Viewport    │ │  Spinner     │  │
│  │  (editor)    │ │  (convo)     │ │  (status)    │  │
│  └─────────────┘ └──────────────┘ └─────────────┘  │
│                                                     │
│  ┌─────────────────────────────────────────────┐    │
│  │           Overlay Stack                      │    │
│  │  Model Selector │ Session Picker │ /tree     │    │
│  │  Settings UI    │ Theme Picker   │ Branches  │    │
│  │  Capability Prompt (extension UI)            │    │
│  └─────────────────────────────────────────────┘    │
└──────────────────────┬─────────────────────────────┘
                       │
              async channels (mpsc)
                       │
┌──────────────────────▼─────────────────────────────┐
│             Agent Async Task                        │
│  Runs on asupersync runtime                         │
│  Streams provider responses                         │
│  Executes tools                                     │
│  Sends PiMsg events back to TUI thread              │
└────────────────────────────────────────────────────┘

The async/sync bridge: The agent runs on the asupersync async runtime in a separate thread. It communicates with the bubbletea UI thread through mpsc channels. Each streaming event (text delta, tool start, tool update, agent done) becomes a PiMsg variant delivered to PiApp::update(), keeping the UI responsive during API streaming and tool execution.

Viewport scrolling: The conversation viewport tracks whether the user is at the bottom. When new content arrives and the user hasn’t scrolled up, the viewport auto-follows the stream tail. Scrolling up disables auto-follow; pressing End or typing a new message re-enables it.

Overlay system: Modal UIs (model selector, session picker, branch navigator, extension capability prompts) stack on top of the main conversation view. Each overlay captures keyboard input until dismissed. Only the topmost active overlay receives events.

Slash commands available in the interactive editor:

Live stream rules: /rules manages corrections enforced during generation on every native agent surface. Matching assistant text or thinking interrupts the response before its tool calls execute, adds the rule’s directive, and retries within a three-correction turn budget. Project rules require workspace trust; triggered corrections survive compaction on their active branch. See Stream rules for configuration and examples.

MCP server authentication and project configuration

Use /mcp list to inspect each server’s selected source, trust decision, and connection health. /mcp trust acknowledges that exact definition; HTTP servers that need OAuth can then sign in with /mcp reauth . Follow the authorization URL and complete the browser callback. A successful reauthentication replaces the live connection and reloads its tools, including in RPC and ACP sessions.

Reauthentication also repairs an expired configured Authorization header. After login succeeds, Pi removes that header from the file that supplied the winning definition and preserves its existing trust acknowledgement for that specific repair. Explicit --mcp-config files take precedence over .pi, .agents, global Pi, and imported tool configurations. Shadowed files, unrelated settings, and authored variable or secret references remain intact. If the selected file changes while the browser is open, Pi preserves the edit and refuses the repair. ACP and extension definitions remain owned by their clients: their explicit login takes over only the exact supplied definition without creating a replacement configuration file.

/mcp unauth removes the managed login and retires the connection and tools. Logout suppresses the old configured bearer as well, and a late browser callback cannot restore a logged-out session. Cancelling the browser wait preserves the previous credentials, source, and trust. If login succeeds but source persistence or reconnection fails, Pi reports that the login was saved so the remaining problem can be repaired without mistaking it for a failed authorization.

Native settings support the OMP project-configuration switch:

{
  "mcp": {
    "enableProjectConfig": false
  }
}

The default is true. When the merged setting is false, automatic discovery skips project .pi and .agents MCP files and imported project configurations. Global Pi configuration and explicit --mcp-config arguments remain eligible. Project configuration also requires workspace trust when the setting is enabled. Client-supplied ACP servers remain explicit session definitions, and each server still requires its own trust acknowledgement.

RPC Protocol

The RPC mode (pi --mode rpc) exposes a line-delimited JSON protocol over stdin/stdout for programmatic integration. Each line is a self-contained JSON object.

Client → Pi (stdin):

{"type": "prompt", "message": "Explain this function", "id": "req-001"}
{"type": "steer", "message": "Focus on error handling"}
{"type": "follow_up", "message": "Now add tests"}
{"type": "abort"}
{"type": "get_state"}
{"type": "compact", "reserveTokens": 8192, "keepRecentTokens": 20000}

Pi → Client (stdout):

{"type": "agent_start", "sessionId": "..."}
{"type": "message_update", "message": {...}, "assistantMessageEvent": {"type": "text_delta", "delta": "The function", "contentIndex": 0}}
{"type": "tool_execution_start", "toolCallId": "...", "toolName": "read", "args": {}}
{"type": "tool_execution_end", "toolCallId": "...", "toolName": "read", "result": {}, "isError": false}
{"type": "agent_end", "sessionId": "...", "messages": [...]}
{"type": "response", "id": "req-001", "command": "prompt", "success": true, "data": {"status": "ok"}}

Fatal errors (JSON and RPC modes): a failure that ends the process — a bad config file, no credentials for the selected model, an unwritable state directory, a usage error — prints exactly one record on stdout before the non-zero exit, so a host never has to parse stderr prose:

{"type": "error", "phase": "startup", "code": "auth.missing_api_key", "message": "No API key found for provider anthropic. Set env var or use --api-key.", "exit_code": 1}

phase is startup when nothing had been written to stdout yet (no session header, no RPC loop) and run otherwise. code is stable: the auth diagnostic codes (auth.missing_api_key, auth.no_models_available, auth.invalid_api_key, auth.quota_exceeded, auth.oauth.token_refresh_failed, …) when the failure classifies as one, else the family — config, session, provider, auth, tool, usage (argument/validation errors, exit code 2), extension, io, json, state_store, aborted, api, or internal. The human-readable diagnosis with hints still goes to stderr. Text mode prints nothing on stdout.

I/O architecture: Two dedicated threads handle stdin reading and stdout writing, bridged to the async agent runtime via channels. The stdin thread retries on transient errors to prevent dropped input. The stdout thread flushes after every line to prevent buffering delays.

Message queuing: While the agent is streaming a response, incoming messages are routed to one of two queues:

Queue modes (All or OneAtATime) control whether multiple queued messages are batched into a single turn or processed individually.

Extension UI over RPC: When an extension requests user input (capability prompt, selection dialog), Pi emits an extension_ui_request event. Response-bearing events include a requestGeneration correlation token, which the client must echo alongside requestId in its extension_ui_response; this prevents a late response from resolving a newer request that reused the same public ID. IDE extensions can then present native UI for capability decisions instead of falling back to terminal prompts.

Session Indexing

Session resume (pi -c or pi -r) needs to find the most recent session for the current project without scanning every JSONL file on disk. Pi maintains a SQLite index (session-index.sqlite) that provides constant-time lookups.

Schema:

CREATE TABLE sessions (
    path            TEXT PRIMARY KEY,
    id              TEXT NOT NULL,
    cwd             TEXT NOT NULL,
    timestamp       TEXT NOT NULL,
    message_count   INTEGER NOT NULL,
    last_modified   INTEGER NOT NULL,
    size_bytes      INTEGER NOT NULL,
    name            TEXT
);

Update lifecycle:

1. After saving a session JSONL file, Pi upserts its metadata into the index
2. pi -c queries WHERE cwd = ? ORDER BY last_modified DESC LIMIT 1
3. pi -r queries the same table and presents a picker sorted by recency

Concurrency: A file-based lock (session-index.lock) serializes writes from concurrent Pi instances. Reads use WAL mode for non-blocking access.

Staleness-based reindexing: If the index is older than a configurable threshold, Pi runs a full re-scan of the sessions directory to catch files created by other instances or manual edits. The re-scan keeps the index accurate without a centralized daemon.

Session Store V2 Sidecar (Large Session Fast-Path)

Pi also supports a v2 sidecar store next to JSONL sessions for faster resume and stronger corruption checks on long histories.

What it adds:

• Segmented append log files (instead of one ever-growing JSONL file)
• Offset index rows for direct seeks and fast tail reads
• Periodic checkpoints and a manifest snapshot
• Migration ledger entries for auditability
• Checkpoint-based rollback path with explicit rollback event logging

How resume works:

1. A verified migration installs a durable clean source-state record, allowing the normal v2 open to use the sidecar index + segments without rescanning the full JSONL file.
2. Before any JSONL mutation, Pi durably marks the sidecar dirty; dirty, invalid, or otherwise stale sidecars fall back to authoritative JSONL parsing. Filesystem mtimes remain a secondary external-change check, and legacy sidecars are fully verified once before use.
3. If index data needs repair, Pi marks the sidecar dirty first and adopts the repaired store only after its entry IDs and hash chain match the source JSONL exactly.

Integrity strategy:

• Segment frames carry payload and chain hashes.
• Index rows store byte offsets plus CRC32C checksums.
• Validation checks offset bounds, checksum matches, and frame/index alignment before trusting the sidecar.
• Truncated trailing frames are recoverable during rebuild; non-EOF frame corruption fails closed instead of silently dropping data.

CLI support:

• pi migrate --dry-run validates migration without writing.
• pi migrate performs JSONL-to-v2 migration and verifies parity.

Authentication & Credential Management

Beyond simple API keys, Pi supports OAuth, AWS credential chains, service key exchange, and bearer-token auth. Credentials are stored in ~/.pi/agent/auth.json with file-locked access to prevent corruption from concurrent instances. Stored API keys can be literal strings, $ENV:VAR_NAME references, or $CMD:shell command / $COMMAND:shell command sources that resolve trimmed stdout at request time.

OAuth token lifecycle:

1. User runs pi with an OAuth-configured provider
2. Pi checks auth.json for an existing token
3. If missing: opens browser to authorization URL, user authenticates, Pi receives authorization code, exchanges it for access + refresh tokens, stores both with expiry timestamp
4. If expired but refresh token valid: exchanges refresh token for new access token, updates auth.json
5. Bearer token attached to API requests

Google CLI-style OAuth providers carry project metadata with the token payload. Pi preserves and refreshes that payload and can resolve project IDs from GOOGLE_CLOUD_PROJECT or local gcloud config when needed.

Credential status reporting: pi config shows the status of each configured provider’s credentials: Missing, ApiKey, OAuthValid (with time until expiry), OAuthExpired (with time since expiry), AwsCredentials, or BearerToken.

Diagnostic codes: Auth failures produce specific diagnostic codes (MissingApiKey, InvalidApiKey, QuotaExceeded, OAuthTokenRefreshFailed, MissingAzureDeployment, MissingRegion, etc.) with context-specific error hints rather than generic messages.

---

Tool Details

read

Read file contents (optionally images):

Input: { "path": "src/main.rs", "offset": 10, "limit": 50 }

• Supports images (jpg, png, gif, webp) with optional auto-resize
• Streams file bytes in chunks with hard size limits to reduce peak memory usage
• Applies defensive image decode limits to block decompression-bomb/OOM inputs
• Truncates at 2000 lines or 1MB
• Returns continuation hint if truncated

bash

Execute shell commands with timeout and output capture:

Input: { "command": "cargo test", "timeout": 120 }

• Default 120s timeout, configurable per-call
• Set timeout: 0 to disable the default timeout
• Process tree cleanup on timeout (kills children)
• Rolling buffer for real-time output
• Full output saved to temp file if truncated

edit

Surgical string replacement:

Input: { "path": "src/lib.rs", "old": "fn foo()", "new": "fn bar()" }

• Exact string matching (no regex)
• Fails if old string not found or ambiguous
• Returns diff preview

grep

Search files, internal documents, or a web page:

{ "pattern": "TODO", "path": "src/", "context": 2, "limit": 100 }
{ "pattern": "FEATURE_X", "path": ".;!vendor;!gen" }
{ "pattern": "track\\(", "path": "src;!*.test.ts;!src/legacy" }
{ "pattern": "critical|warning", "path": "agent://reviewer/findings;local://review-notes", "ignoreCase": true }
{ "pattern": "migration", "path": "skill://release/references/checklist.md" }
{ "pattern": "deprecated", "path": "https://example.com/changelog" }
{ "pattern": "data-version", "path": "https://example.com/changelog:raw", "literal": true }

• Regex and literal matching, case folding, and merged context lines work across all sources. Each result identifies its source and line number.
• Filesystem searches respect .gitignore and workspace boundaries. hashline: true provides edit anchors for filesystem results.
• Exclude files or entire directory trees with path: "root;!glob;!glob" (up to 64 entries and 64 KiB). A bare name such as !vendor matches at any depth below the root; a slash-containing pattern such as !src/legacy or !./gen is anchored to the working directory. Use !**/generated to match a directory at any depth. Exclusions alone default to the working directory, and explicitly named roots remain searchable. Existing filenames containing ! or semicolons retain their literal meaning. Both search backends prune exclusions before counting matches, while included paths retain normal permission checks. Use separate calls for multiple filesystem roots or resource URLs.
• Internal documents use the same resolver as read, including accepted agent:// results and JSON fields, skills, prompts, scratch notes, merge conflicts, PRs, issues, and SSH documents. Pending or rejected child output retains its normal access restrictions. Separate up to 16 internal URLs with semicolons; filesystem paths and HTTP URLs use separate calls.
• HTTP(S) searches use reader text by default; append :raw to search the fetched source. Semicolons inside an HTTP URL remain part of that URL. The existing read.urlAllowPrivateTargets setting governs private-network access, with the same redirect checks as read. PR, issue, SSH, and conflict lookups retain their process effects for execution approval.
• The match limit applies across the entire resource list. Resource searches accept up to 10,000 matches and 10 MiB per document; output still truncates at 1 MiB and 500 characters per line. Results report skipped resources and partial source documents. glob and hashline apply to filesystem searches; virtual document line numbers are search locations.

find

Discover files by pattern:

Input: { "pattern": "*.rs", "path": "src/", "limit": 1000 }

• Glob patterns matched in-process (no fd required)
• Sorted by modification time
• Respects .gitignore

ls

List directory contents:

Input: { "path": "src/", "limit": 500 }

• Alphabetically sorted
• Directories marked with trailing /
• Truncates at limit

---

Performance Engineering

Why Rust Matters for CLI Tools

CLI tools have different performance requirements than servers or GUI applications. The critical metric is time-to-first-interaction: how quickly can the user start typing after invoking the command?

Current comparative timing numbers are intentionally omitted until a fresh, provenance-matched benchmark run is available.

Optimization Playbook

Pi applies performance-oriented design at multiple layers, not only in the Cargo release profile.

Where performance work is concentrated:

• Startup path kept minimal: no JS runtime bootstrap, no module graph loading, no JIT warmup.
• Hot hostcall specialization: common extension hostcalls use typed fast paths; uncommon shapes fall back to compatibility paths.
• Adaptive dispatch under load: hostcall scheduling can switch modes when contention rises, then switch back when pressure drops.
• Fast-path safety guardrails: sampled shadow dual-execution checks ensure optimizations do not silently change behavior.
• Low-allocation rendering: TUI render buffers and markdown render results are cached/reused instead of rebuilt every frame.
• Fast resume internals: session indexing plus the v2 sidecar layout avoid expensive full-history scans on resume.
• Bounded growth controls: compaction and truncation keep token/context growth and tool-output payload growth from degrading responsiveness over long sessions.
• Measurement-first culture: perf artifacts are schema-validated and claim-gated in CI, so optimization work is driven by evidence and regressions are caught early.

These mechanisms are intended to preserve responsiveness under streaming, tool, large-session, and extension workloads. Achieved release-level latency and memory results require fresh, provenance-matched measurements.

Optimization Catalog (Code + Commit History)

This catalog reflects a long sequence of changes across runtime, storage, streaming, and UI.

Concrete engineering work in this codebase includes:

Optimization spans algorithms, execution lanes, memory movement, queue discipline, storage layout, and validation policy as one system.

Release Profile and Binary Size Gate

The shipping release profile is tuned to keep release artifacts inside the binary-size budget while preserving cross-crate optimization:

[profile.release]
opt-level = "z"      # Optimize generated code for size
lto = true           # Link-time optimization across all crates
codegen-units = 1    # Single codegen unit (slower compile, better optimization)
panic = "abort"      # No unwinding machinery
strip = true         # Remove symbol tables

Binary size is explicitly budgeted by the DSR quality recipe via binary_size_release, with a target threshold of 48.0 MiB (the harness computes bytes / 1024 / 1024; raised from 26.0 MiB for the v0.3.0 capability wave — BPE token tables, LSP/DAP bridges, the MCP client, and eval kernels — and previously from 22.0 MiB with the FrankenSQLite cutover). A fresh v0.3.0 release measurement is required before reporting the achieved size. Default release builds keep heavyweight extras opt-in; use --features full when you need the image, clipboard, wasm, jemalloc, and syntax-highlighting extras in one build.

Published binaries (GitHub Releases, the installer, pi self-update) are built by DSR with --profile dist. It inherits every setting above and sets opt-level = 3, so the binaries people download are optimized for speed and are larger than a release build. The binary_size_release budget keeps measuring the release profile.

Benchmark Evidence vs Shipping Artifacts

Pi separates performance evidence artifacts from distributable release artifacts:

• Benchmark evidence artifacts (PERF-3X and certification inputs) are produced by scripts/perf/orchestrate.sh and scripts/bench_extension_workloads.sh, and must carry run provenance (correlation_id) plus profile labels (for example build_profile=perf).
• Shipping artifacts (end-user binaries) are built with the Cargo dist profile (release with opt-level = 3) and distributed via GitHub Releases + installer paths.

Policy implication: release/size artifacts alone are not valid evidence for global performance claims. Performance claims must cite benchmark evidence bundles with reproducible provenance. See docs/testing-policy.md and docs/releasing.md for normative policy details.

Current checked-in performance evidence state:

• Run output: tests/perf/reports/ (budget_summary.json, PERF_BUDGETS.md)

• Current budget summary (generated 2026-09-21 from source 9c887d33; no run id, no correlation id, strict_mode false): 19 declared budgets, of which 0 PASS, 0 FAIL and 19 NO_DATA. Header and per-budget budget_results rows now agree, and they agree on this: the tree carries no performance measurements at all. Claim readiness is blocked, performance_claims_authorized is false, and the blocking reason codes are budget_data_missing, ci_budget_data_missing, correlation_id_missing, data_contract_failure, run_id_missing and strict_mode_disabled. Counts are value-bound to the artifact: (from tests/perf/reports/budget_summary.json; no performance claim)

• Before spending time on a definitive refresh, run python3 scripts/perf/preflight_budget_inputs.py to list missing budget inputs, expected artifact paths, and RCH-only refresh commands.

• Orchestrated perf runs also write results/perf_budget_preflight_before_refresh.json before any budget_summary.json refresh and results/perf_artifact_staging_manifest.json after collection. The staging manifest records each required artifact’s source path, RCH remote-source prefix when supplied via PERF_REMOTE_TARGET_DIR, retrieved local path, schema, size, mtime, checksum, and explicit blocker status.

• Perf evidence cache entries live under PERF_EVIDENCE_CACHE_DIR (default: $CARGO_TARGET_DIR/perf/evidence_cache) with schema pi.perf.evidence_cache.v1. Preflight and staging may reuse cached evidence only when the entry’s schema, command, git commit, build profile, run ID/correlation ID, host/toolchain provenance, checksum, and TTL validate; reused artifacts are labeled source_kind=cache/evidence_source=cache in the JSON outputs. Override the cache lifetime with PI_PERF_EVIDENCE_CACHE_TTL_HOURS.

• env_fingerprint.json records cgroup-aware host topology with schema pi.perf.host_topology_fingerprint.v1: cgroup v2 CPU quota, cpuset size, memory limits, NUMA node count, caveats, and a constrained budget_profile for containerized hosts.

• Regenerate the perf evidence bundle before adding release-facing speed, throughput, memory, or startup numbers to this README.

• Full-suite gate at its latest regeneration (2026-08-04): 17/20 gates passed with 12/14 blocking gates. The ext_must_pass blocker is now substantive rather than staleness-only: the 2026-08-17 regeneration recorded 206/208 with two marckrenn-pi-sub family failures (tracked for remediation), alongside the practical-finish checkpoint (from tests/full_suite_gate/full_suite_verdict.json; historical snapshot)

• Unified evidence bundle, as regenerated 2026-08-28: 30 total sections, 18 present, 12 missing, 0 invalid (bundle verdict: partial) (from tests/evidence_bundle/index.json; historical snapshot)

• Historical drop-in result: 22/22 certification gates passed, overall verdict CERTIFIED for source 52e9fbfb24352045985b59df9d7ea63f1f8f2ef8 only (from the verdict blob stored at Git revision 2fc4b8c0b77ded267cf5e0f517f4b6fa87f45e91, not from the live verdict file)

• Extension must-pass gate, as regenerated 2026-08-17: 206/208 must-pass extensions passed (2 failures); informational stretch set 10/19 passed (from tests/ext_conformance/reports/gate/must_pass_gate_verdict.json; historical snapshot)

• Context-intelligence closeout gate: pass, with child Beads mapped to code, tests, docs/evidence, validation commands, pushed commits, redaction posture, perf-budget evidence, README freshness, staged UBS, and Beads ledger reconciliation (from docs/evidence/context-intelligence-closeout-gate.json; historical snapshot)

• Progress SLO closeout gate: pass, with child Beads mapped to code, tests, docs/evidence, validation commands, pushed commits, source-boundary checks, stress-budget evidence, README freshness, staged UBS, and Beads ledger reconciliation (from docs/evidence/swarm-progress-slo-closeout-gate.json; historical snapshot)

• Runtime-intelligence closeout gate: pass, with child Beads mapped to compaction admission, tool-output artifacts, provider routing, scheduler fairness, frame-budget telemetry, cancellation cleanup, extension safety provenance, docs/evidence, source-boundary checks, pushed commits, staged UBS, and Beads ledger reconciliation (from docs/evidence/runtime-intelligence-closeout-gate.json; historical snapshot)

Quality and benchmark authority

Use the registered DSR recipe for day-to-day validation and definitive benchmark evidence:

• Quality gate: dsr quality --tool pi_agent_rust.
• Definitive evidence: require the DSR run to regenerate and validate these artifacts:
  • tests/perf/reports/phase1_matrix_validation.json
  • tests/full_suite_gate/full_suite_verdict.json
  • tests/full_suite_gate/certification_verdict.json
  • tests/full_suite_gate/extension_remediation_backlog.json

Direct Cargo, RCH, and repository wrapper invocations are not evidence and are not an authorized fast path around DSR.

Extension Workload Hotspot Profiling

Pi includes a dedicated ext_workloads workload harness for extension runtime bottlenecks. Run it only through the registered DSR performance lane, which writes ext_workloads.jsonl, ext_hostcall_hotspot_matrix.json, and ext_hostcall_bridge_trace.jsonl with DSR provenance.

This harness does more than raw timing:

• Breaks hostcall cost into six stages: marshal, queue, schedule, policy, execute, io
• Produces a hotspot matrix (pi.ext.hostcall_hotspot_matrix.v1) for quick bottleneck ranking
• Produces bridge trace events (pi.ext.hostcall_trace.v1) for per-call debugging
• Measures how stage pairs interact and verifies full stage-pair coverage
• Generates a VOI scheduler plan (pi.ext.voi_scheduler.v1) to recommend the next highest-value experiments under a fixed overhead budget

In plain terms: it helps answer “what should we optimize next?” with data, not guesswork.

For multi-agent operator handoff, scripts/build_swarm_operator_runpack.py can also project a diagnostic-only bottleneck_attribution dashboard from the tail latency report, swarm flight recorder report, swarm resource preflight, hostcall admission swarm profile, session recovery swarm profile, concurrent RPC swarm E2E evidence, and RCH artifact-sync preflight. The dashboard is operator evidence only; release-facing speed, drop-in, or performance claims still require the claim-integrity gates below to pass.

The runpack also embeds predictive_telemetry_ledger (pi.swarm.predictive_telemetry_ledger.v1) and can write it separately with --out-predictive-telemetry-ledger-json or print it with --print-predictive-telemetry-ledger. The ledger is read-only advisory evidence: it joins existing RCH, Agent Mail, Beads, turn-pressure, bottleneck-attribution, and claim-readiness signals into ordered pressure observations and next-bottleneck hypotheses, but it does not mutate the scheduler, Agent Mail, RCH, Beads, git, or release/capacity claims.

The runpack also embeds validation_scheduler_plan (pi.swarm.validation_scheduler_plan.v1) and can write it separately with --out-validation-scheduler-plan-json or print it with --print-validation-scheduler-plan. This is an advisory RCH-aware simulator: it ranks fast script checks, evidence regeneration, focused tests, E2E/conformance, compiler checks, and linting from the current git change profile, predictive telemetry, RCH admission, remote proof, and target cache evidence. It preserves exact command strings and required RCH env, but it does not execute cargo, mutate RCH, reserve Agent Mail, claim Beads, delete temp artifacts, or allow heavy cargo to fail open into a local build.

If validation-broker status or plan JSON is supplied with --validation-broker-json, the runpack projects source status, slot counts, stale slot warnings, duplicate-gate opportunities, and recommended next actions as advisory handoff data. This projection does not replace DSR, Doctor, Beads, Agent Mail, UBS, or release-claim gates. Operator workflow, privacy, and degraded-data guidance lives in docs/swarm-operations-runbook.md#validation-broker-operator-workflow.

If progress SLO JSON is supplied with --progress-slo-json, the runpack projects the current swarm progress posture, confidence, saturation summary, source freshness, redaction posture, and next actions as advisory handoff data. This projection does not replace Beads, Agent Mail, RCH, Doctor, validation gates, git, or release-claim gates. Operator workflow, privacy, degraded-source handling, stale-Beads handling, RCH pressure interpretation, and no-open-work convergence guidance lives in docs/swarm-operations-runbook.md#progress-slo-operator-workflow.

The same runpack command can emit a dry-run swarm autopilot input pack and plan beside the handoff bundle. When those companion artifacts are requested, the runpack JSON/Markdown includes an autopilot_handoff summary (pi.swarm.autopilot_handoff.v1) with pi.swarm.autopilot_input_pack.v1 and pi.swarm.autopilot_plan.v1 schema references, selected advisory action, artifact paths, and source provenance. The autopilot never mutates ownership or replaces Doctor, Beads, Agent Mail, RCH, git, or the source artifacts; it only turns those inputs into reproducible operator next-action guidance. Its work-admission gate includes a read-only dry-run executor that classifies plan items as would_execute, blocked, requires_operator, or never_execute so deletion requests, Agent Mail/RCH mutation, local heavyweight Cargo, and Beads ownership bypasses are rejected without executing anything.

For degraded-coordination handoff proof, the runpack script also has a no-mock E2E mode (--run-degraded-coordination-e2e) that feeds real temporary Beads state plus fixture-captured Agent Mail and RCH failures through the runpack builder. The emitted evidence stays operator-only: it proves Beads soft-lock fallback, degraded validation, and no cleanup/deletion-command output without supporting release-facing speed or drop-in claims.

Fourth-wave swarm self-healing guidance layers a read-only stale-evidence renewal queue, dry-run action plan, work admission gate, budget lease simulation, turn pressure ledger, extension quarantine rehearsal, and redacted handoff summaries onto that runpack flow. These artifacts can recommend renew_stale_evidence, wait_for_pressure, use_beads_soft_lock, or similar operator actions, but they do not mutate Beads, Agent Mail, extension configuration, git, evidence files, RCH jobs, or release claims. The full workflow and safe handoff wording live in docs/swarm-operations-runbook.md#fourth-wave-self-healing-workflow.

The fourth-wave closeout gate emits pi.swarm.fourth_wave_self_healing.closeout_gate.v1. It maps the bd-63x3v.7 child Beads to implementation artifacts, contracts, fixtures, operator docs, validation commands, pushed refs, source-boundary checks, and remaining advisory limits. It is governed by docs/contracts/fourth-wave-self-healing-closeout-gate-contract.json; the current closeout artifact is docs/evidence/fourth-wave-self-healing-closeout-gate.json.

Offline swarm replay traces can be previewed before handoff with pi swarm-replay-preview --trace --format json. The command is read-only, emits pi.swarm.replay_preview.v1 JSON or concise text, refuses to overwrite requested output files, and can feed scripts/build_swarm_operator_runpack.py --swarm-replay-preview-json so runpacks carry the replay policy comparison without treating the runpack as source-of-truth replay evidence. Operator workflow, privacy, and degraded-data guidance lives in docs/swarm-replay-operator-workflow.md.

The autopilot closeout gate emits pi.swarm.autopilot_decision_gate.v1 to audit the shipped input pack, planner, work partitions, failure actions, budget drift watcher, E2E/logging evidence, runpack handoff, safety guards, pushed commits, and quality gates before the swarm-autopilot epic is closed.

Context intelligence has its own operator guide and closeout gate. See docs/context-intelligence.md for configuration, preview workflows, failure modes, privacy posture, examples, troubleshooting, and final closeout evidence. The context-intelligence closeout gate emits pi.context_intelligence.closeout_gate.v1 and is governed by docs/contracts/context-intelligence-closeout-gate-contract.json; the current closeout artifact is docs/evidence/context-intelligence-closeout-gate.json.

The progress-SLO closeout gate emits pi.swarm.progress_slo.closeout_gate.v1. It audits the shipped contract/source inventory, deterministic evaluator, read-only CLI, Doctor/runpack projection, no-mock E2E evidence, large-host stress budgets, operator docs, pushed child commits, source-boundary checks, staged UBS, and Beads ledger reconciliation. It is governed by docs/contracts/swarm-progress-slo-closeout-gate-contract.json; the current closeout artifact is docs/evidence/swarm-progress-slo-closeout-gate.json.

The runtime-intelligence closeout gate emits pi.runtime_intelligence.closeout_gate.v1. It maps the bd-h66tp child Beads to compaction admission, tool-output artifacts, provider routing, scheduler fairness, frame-budget telemetry, cancellation cleanup, extension safety provenance, docs/evidence, source-boundary checks, pushed refs, and quality gates. It is governed by docs/contracts/runtime-intelligence-closeout-gate-contract.json; the current artifact is docs/evidence/runtime-intelligence-closeout-gate.json. The gate is advisory closeout evidence only and does not replace Beads, git, RCH, UBS, CI, claim-integrity gates, or source artifacts.

The sixth-wave validation-hardening closeout gate emits pi.swarm.validation_hardening.closeout_gate.v1. It maps the bd-63x3v.9 child Beads to RCH workspace-shadow detection, Agent Mail degraded Beads soft-lock evidence, validation proof replay, temp artifact inventory, stale claim heuristics, degraded-coordination E2E proof, cgroup/NUMA/RCH budget modeling, provider/RPC/TUI backpressure budgets, operator explanations, pushed refs, source-boundary checks, and quality gates. The provider/RPC/TUI contract now includes a synthetic fairness stress gate for balanced load, one slow provider stream, RPC output flood, and TUI frame-budget pressure, with fail-closed negative controls when one surface reports success while another starves. It is governed by docs/contracts/sixth-wave-validation-hardening-closeout-gate-contract.json; the current artifact is docs/evidence/sixth-wave-validation-hardening-closeout-gate.json. The gate is advisory closeout evidence only and does not replace Beads, git, RCH, UBS, CI, claim-integrity gates, or source artifacts.

The seventh-wave runtime-autonomy closeout gate emits pi.swarm.runtime_autonomy.closeout_gate.v1. It maps the bd-63x3v.10 child Beads to effect-aware tool batching, fail-closed validation-proof reuse, cgroup/NUMA lane placement, large-session replay acceleration, provider/RPC/TUI fairness, work-admission dry-run execution, hostcall QoS starvation evidence, source-boundary checks, pushed refs, and quality gates. It is governed by docs/contracts/seventh-wave-runtime-autonomy-closeout-gate-contract.json; the current artifact is docs/evidence/seventh-wave-runtime-autonomy-closeout-gate.json. The gate is advisory closeout evidence only and does not replace Beads, git, RCH, UBS, CI, claim-integrity gates, child evidence, or source artifacts.

The proof-carrying swarm test-fabric closeout gate emits pi.swarm.proof_carrying_test_fabric.closeout_gate.v1. It maps the bd-zeccr child Beads to no-mock lifecycle E2E evidence, cross-surface conformance, operator evidence goldens, structure-aware fuzz/property coverage, metamorphic replay equivalence, source-boundary checks, pushed refs, negative controls, and quality gates. It is governed by docs/contracts/proof-carrying-swarm-test-fabric-closeout-gate-contract.json; the current artifact is docs/evidence/proof-carrying-swarm-test-fabric-closeout-gate.json. The gate is advisory closeout evidence only and does not replace Beads, git, RCH, Agent Mail, UBS, CI, claim-integrity gates, child evidence, or source artifacts.

The predictive-operations closeout gate emits pi.swarm.predictive_operations.closeout_gate.v1. It maps the bd-63x3v.11 child Beads to predictive telemetry fusion, RCH-aware validation scheduling, semantic compaction quality, extension hostcall cost attribution, operator-perceived latency, redundant-agent-work detection, source-boundary checks, pushed refs, staged UBS, and Beads ledger reconciliation. It is governed by docs/contracts/predictive-operations-closeout-gate-contract.json; the current artifact is docs/evidence/predictive-operations-closeout-gate.json. The gate is advisory closeout evidence only and does not replace Beads, git, RCH, Agent Mail, UBS, CI, claim-integrity gates, child evidence, or source artifacts.

The ninth-wave incident replay and proof-memory closeout gate emits pi.swarm.incident_replay_proof_memory.closeout_gate.v1. It maps the bd-9yq7i child Beads to incident corpus, incident replay, validation proof-memory, operator work recommendation, operator smoothness SLO, extension resource firewall matrix, and no-mock incident replay E2E evidence. It is governed by docs/contracts/ninth-wave-incident-replay-proof-memory-closeout-gate-contract.json; the current artifact is docs/evidence/ninth-wave-incident-replay-proof-memory-closeout-gate.json. The gate is advisory closeout evidence only and does not replace Beads, git, RCH, Agent Mail, UBS, CI, claim-integrity gates, child evidence, generated target/perf outputs, prior-wave evidence, or source artifacts.

The closeout evidence registry at docs/contracts/closeout-evidence-registry.json is the machine-readable index of current advisory closeout-gate artifacts. It binds each current artifact to its governing contract, decision schema, source bead or epic, Markdown references, and advisory boundary, including current artifacts that are not expanded above in prose: docs/evidence/adaptive-execution-closeout-gate.json, docs/evidence/extension-compatibility-closeout-gate.json, and docs/evidence/swarm-replay-closeout-gate.json. The registry is not a release manifest, performance evidence bundle, Beads source of truth, Agent Mail authority, RCH authority, CI/UBS replacement, drop-in certification gate, or permission to delete or mutate files.

The operator-perceived latency trace emits pi.operator.perceived_latency_trace.v1. It joins deterministic provider, RPC, TUI, tool-update, and operator-visible fixture timelines so operators can see when semantic output became visible while low-value updates were coalesced. It is governed by docs/contracts/operator-perceived-latency-trace-contract.json; the current fixture artifact is docs/evidence/operator-perceived-latency-trace.json. The trace is advisory only and does not authorize benchmark, capacity, release performance, strict drop-in, or backpressure-budget replacement claims.

The operator smoothness SLO emits pi.operator.smoothness_slo.v1. It uses deterministic high-volume fixtures for provider stream deltas, RPC output pressure, TUI frame rendering, tool-update coalescing, and session-write pressure. It is governed by docs/contracts/operator-smoothness-slo-contract.json; the current fixture artifact is docs/evidence/operator-smoothness-slo.json. Its p50/p95/p99 counters are engineering fixture counters only, proving semantic milestones remain visible while low-value updates can be coalesced; they are not release benchmark, capacity, performance, or strict drop-in evidence.

The extension resource firewall matrix emits pi.ext.resource_firewall_matrix.v1 from the deterministic extension stress fixture. It covers cheap-read floods, large payload emission, denied capability churn, slow hostcalls, repeated failure, and steady-peer progress while preserving payload redaction and existing extension capability boundaries. The contract is docs/contracts/extension-resource-firewall-matrix-contract.json; focused test runs write resource_firewall_matrix.json under the configured target/perf directory. This matrix is advisory stress evidence only and does not replace runtime enforcement, hostcall cost attribution, RCH validation, Agent Mail, Beads, UBS, CI, or benchmark/capacity/release claims.

The swarm incident corpus emits pi.swarm.incident_corpus.v1. It records deterministic degraded-source fixtures for Agent Mail schema corruption, RCH saturation with local-fallback denial, stale evidence, duplicate work risk, dirty worktree admission denial, malformed sources, and deletion or live mutation rejection. It is governed by docs/contracts/swarm-incident-corpus-contract.json; the current fixture artifact is docs/evidence/swarm-incident-corpus.json. The corpus is operator evidence only and does not replace release performance, drop-in certification, Agent Mail, RCH, Beads, git, source artifacts, or destructive-action authority.

The swarm incident replay harness emits pi.swarm.incident_replay.v1. It consumes the checked-in incident corpus and replays healthy, degraded Agent Mail, saturated RCH, duplicate work, stale evidence, malformed source, dirty worktree, and deletion-request scenarios into ordered phases, per-step assertions, redacted excerpts, selected safe actions, and follow-up recommendations. It is governed by docs/contracts/swarm-incident-replay-contract.json; the current fixture artifact is docs/evidence/swarm-incident-replay.json. Replay output is read-only operator evidence and cannot replace source systems or authorize live Agent Mail, RCH, Beads, git, deletion, release, benchmark, or drop-in claims.

The swarm incident replay E2E harness emits pi.swarm.incident_replay_e2e.v1 plus pi.swarm.incident_replay_e2e.event.v1 JSONL events. It joins real temporary Beads and git workspaces with fixture-captured degraded Agent Mail and RCH inputs, then verifies healthy replay, Beads soft-lock fallback, RCH proof refresh backoff, duplicate-work risk, dirty worktree denial, stale proof-memory refresh, extension resource firewall failure, and smoothness SLO failure. It is governed by docs/contracts/swarm-incident-replay-e2e-contract.json; the current fixture artifact is docs/evidence/swarm-incident-replay-e2e.json. This evidence is advisory only and does not authorize live source mutation, local heavyweight Cargo fallback, release, benchmark, capacity, or drop-in claims.

The validation proof-memory index emits pi.validation.proof_memory_index.v1. It indexes existing remote validation proof fixtures by command fingerprint, git head, touched paths, RCH provenance, artifact retrieval hash, freshness, and reuse eligibility. The index is governed by docs/contracts/validation-proof-memory-index-contract.json; the current fixture artifact is docs/evidence/validation-proof-memory-index.json. It is advisory operator evidence only: stale refs, missing artifacts, local fallback, dirty worktree mismatch, command mismatch, uncovered paths, or non-authoritative coverage all fail closed and require fresh validation.

The operator work recommender emits pi.swarm.operator_work_recommendation.v1. It consumes the incident replay and proof-memory artifacts, then ranks read-only next-work decisions for ready Beads, no ready work, Agent Mail corruption, RCH saturation, stale proof refresh, duplicate-work risk, and dirty-worktree admission denial. It is governed by docs/contracts/operator-work-recommendation-contract.json; the current fixture artifact is docs/evidence/operator-work-recommendation.json. Its recommendations are advisory only and never claim, reserve, launch RCH, run cargo, mutate git, delete files, or replace source systems.

For the full launch, throttling, recovery, and handoff workflow for large multi-agent runs, see docs/swarm-operations-runbook.md.

Claim-Integrity Gates for Performance Reporting

Pi’s perf pipeline includes strict evidence checks so global speed claims cannot be based on partial or stale data.

v0.3.0 performance-claim status: NOT authorized. Its checked-in pi.perf.budget_summary.v2 artifact is an explicit blocked/NO_DATA result, so this release may ship only without quantitative or global performance claims. That blocked state is not a passing benchmark result.

• scripts/perf/orchestrate.sh generates artifacts tied to a shared correlation_id for the same run.
• scripts/e2e/run_all.sh validates required schemas, freshness, and correlation_id alignment before considering claims valid.
• tests/release_evidence_gate.rs and scripts/release_gate.sh always reject malformed schemas, a non-canonical budget-inventory digest, reordered/duplicate/missing results, mismatched comparison semantics, inconsistent counts/statuses, or forged readiness fields. A coherent blocked summary is only a release warning when RELEASE_GATE_REQUIRE_PERFORMANCE_CLAIM_READY=0; it still cannot authorize performance copy.
• A claim_ready summary must come from strict mode with data for every declared budget, no declared budget failures or data-contract failures, complete CI coverage, one matching run_id/correlation_id, and fresh source-bound evidence. The aggregate budget_data_missing and budget_failed blockers prevent non-CI informational budgets from being silently excluded from the global authorization boolean. The gate also proves its exact canonical strict test was listed once, executed once, and was not ignored; that test freshly recomputes and deep-compares the checked-in definitions, results, failures, counts, and readiness instead of accepting an unrelated passing benchmark run.
• scripts/e2e/run_all.sh emits an evidence-adjudication matrix and only treats evidence as canonical when freshness and lineage checks both pass.
• Key release-facing artifacts include:
  • pi.perf.extension_benchmark_stratification.v1
  • pi.perf.phase1_matrix_validation.v1
  • pi.claim_integrity.evidence_adjudication_matrix.v1

If the evidence set is incomplete or contradictory, the claim-integrity gate stays closed and reports exactly why.

Allocator Strategy for Benchmarks

Default release builds use the platform allocator to stay inside the shipping binary-size budget. FreeBSD and MSVC builds also always use the platform allocator; FreeBSD libc already uses jemalloc internally, and mixing allocator domains across libc/pthread and C dependencies can turn heap corruption into thread-spawn crashes. For allocator experiments, Pi supports explicit system and jemalloc benchmark variants:

# System allocator baseline + jemalloc variant in one repeatable run
BENCH_ALLOCATORS_CSV=system,jemalloc \
  ./scripts/bench_extension_workloads.sh

The benchmark harness records both requested and effective allocator metadata in its JSONL output (allocator_requested, allocator_effective, allocator_fallback_reason) via PI_BENCH_ALLOCATOR.

• system: build with the explicit benchmark feature set except jemalloc
• jemalloc: build with --features jemalloc where supported by the target
• auto: prefer jemalloc, fall back to system if build fails

If jemalloc is requested but unavailable for the current build, the run fails closed to the compiled allocator and includes a fallback reason in artifacts.

Memory Usage

Rust’s ownership model avoids a garbage-collected application runtime. Current memory figures are intentionally withheld until the v0.3.0 evidence inputs are regenerated with matching source and run provenance.

Streaming Architecture

Responses move incrementally from the API to the terminal with bounded buffering for incomplete network and SSE event frames:

API Server → TCP → SSE Parser → Event Handler → Terminal
     │                              │
     └──── incremental bounded buffering ────┘

The SSE parser processes events as they arrive rather than waiting for complete responses.

TUI Rendering Performance

The interactive TUI has a 60fps design target, with several optimization layers:

Frame timing telemetry: Every render cycle is instrumented. Slow frames (>16ms) are tracked and classifiable by phase: viewport sync, message encoding, markdown rendering. This data feeds the internal performance monitor.

Message render cache: Markdown-to-ANSI conversion involves syntax highlighting, table layout, and link detection. Pi caches the rendered output per message and invalidates only on theme change or terminal resize. During streaming, only the actively-changing message is re-rendered; all prior messages hit the cache.

Pre-allocated render buffers: The RenderBuffers struct is reused across render cycles. Rather than allocating new String buffers each frame, Pi writes into pre-sized buffers and clears them before reuse, avoiding repeated small allocations during streaming.

Memory pressure monitoring: A MemoryMonitor samples process heap size and classifies it into three tiers:

Progressive degradation keeps Pi responsive during long sessions with accumulated tool output.

---

Troubleshooting

For a more complete guide, see docs/troubleshooting.md.

“fd not found”

The find tool searches in-process by default and needs no fd binary. This error only appears with "search_backend": "external" in settings.json; either remove that setting or install fd:

# Ubuntu/Debian
apt install fd-find

# macOS
brew install fd

# The binary might be named fdfind
ln -s $(which fdfind) ~/.local/bin/fd

“API key not set”

export ANTHROPIC_API_KEY="sk-ant-..."

# Or in settings.json
{ "apiKey": "sk-ant-..." }

# Or per-command
pi --api-key "sk-ant-..." "Hello"

“Session corrupted”

Sessions are append-only JSONL. If corruption occurs:

# Start fresh
pi --no-session

# Or delete the problematic session
rm ~/.pi/agent/sessions/--home-user-project--/corrupted-session.jsonl

“Streaming hangs”

Check your network connection. Pi uses SSE which requires stable connections:

# Test with curl
curl -N https://api.anthropic.com/v1/messages

“Tool output truncated”

This is intentional to prevent context overflow. Use offset/limit:

# In the conversation
"Read lines 2000-4000 of that file"

---

Limitations

Pi is honest about what it doesn’t do:

---

Design Philosophy

Specification-First Porting

This port follows a “specification extraction” methodology rather than line-by-line translation:

1. Extract behavior: Study the TypeScript implementation to understand what it does, not how
2. Document the spec: Write down expected behaviors, edge cases, and invariants
3. Implement from spec: Write idiomatic Rust that satisfies the spec
4. Conformance testing: Verify behavior matches via fixture-based tests

This approach yields better code than mechanical translation. TypeScript idioms (callbacks, promises, class hierarchies) don’t map cleanly to Rust (ownership, traits, enums). Fighting the language produces worse results than embracing it.

Conformance Testing

The test suite includes fixture-based conformance tests that validate tool behavior:

{
  "version": "1.0",
  "tool": "edit",
  "cases": [
    {
      "name": "edit_simple_replace",
      "setup": [
        {"type": "create_file", "path": "test.txt", "content": "Hello, World!"}
      ],
      "input": {
        "path": "test.txt",
        "oldText": "World",
        "newText": "Rust"
      },
      "expected": {
        "content_contains": ["Successfully replaced"],
        "details": {"oldLength": 5, "newLength": 4}
      }
    }
  ]
}

Each fixture specifies:

• Setup: Files/directories to create before the test
• Input: Tool parameters
• Expected: Output content patterns, exact field matches, or error conditions

The Rust implementation can be validated against the TypeScript original without coupling to implementation details.

Extension System

Pi supports legacy JS/TS extensions via an embedded QuickJS runtime. Unlike traditional plugin systems, extensions run in a sandboxed, capability-gated environment with no ambient OS access:

1. No Node/Bun required: QuickJS + Pi-provided shims for common Node APIs
2. Capability-based security: each host connector call is policy-checked and logged
3. Conformance-tested: status is tracked in docs/ext-compat.md and tests/ext_conformance/reports/pipeline/
4. Measured load paths: cold and warm extension loads are instrumented; fresh release measurements are pending

Legacy extension behavior is automatic:

• Existing .js/.ts extensions run directly (no manual conversion step).
• *.native.json descriptors are optional and mainly useful for native-rust runtime workflows.
• One session currently uses one runtime family at a time (JS/TS or native descriptor).

Policy preset quick-start:

# Inspect current effective policy
pi --explain-extension-policy

# Switch profile for one command (safe | balanced | permissive)
pi --extension-policy balanced --explain-extension-policy

# Legacy alias is still accepted:
pi --extension-policy standard --explain-extension-policy

# Narrow dangerous-capability opt-in (preferred over permissive)
PI_EXTENSION_ALLOW_DANGEROUS=1 pi --extension-policy balanced --explain-extension-policy

Operator rollout playbook (compatibility-first local defaults + explicit lock-down):

# 1) Baseline: verify defaults are compatibility-first (`permissive`)
pi --explain-extension-policy

# 2) Staging: use balanced prompting, dangerous caps still denied by default
pi --extension-policy balanced --explain-extension-policy

# 3) Explicit lock-down for strict local/CI runs
pi --extension-policy safe --explain-extension-policy

# 4) Narrow opt-in for dangerous capabilities (preferred path)
PI_EXTENSION_ALLOW_DANGEROUS=1 pi --extension-policy balanced --explain-extension-policy

# 5) Explicit permissive mode when you want to be unambiguous
pi --extension-policy permissive --explain-extension-policy

settings.json baseline for local/dev:

{
  "extensionPolicy": {
    "defaultPermissive": true
  }
}

Use this to restore the stricter fallback without CLI flags:

{
  "extensionPolicy": {
    "defaultPermissive": false
  }
}

Interactive TUI: open /settings and toggle extensionPolicy.defaultPermissive.

CI guidance:

# CI default: keep dangerous capabilities disabled
pi --extension-policy safe --explain-extension-policy

# CI opt-in job (only where required), keep explicit and auditable
PI_EXTENSION_ALLOW_DANGEROUS=1 pi --extension-policy balanced --explain-extension-policy

Rollback rule: remove PI_EXTENSION_ALLOW_DANGEROUS, set extensionPolicy.profile back to safe or set extensionPolicy.defaultPermissive to false, and re-run pi --explain-extension-policy to confirm deny decisions.

See EXTENSIONS.md for the full architecture, runtime contract, and conformance results.

Unsafe Forbidden

The #![forbid(unsafe_code)] directive is project-wide and non-negotiable. Rationale:

• Attack surface: Pi executes user-provided shell commands and reads arbitrary files
• Memory bugs = security bugs: Buffer overflows or use-after-free in this context could be exploitable
• No safety-for-speed trade: The project forbids unsafe Rust regardless of workload profile
• Dependency boundary is explicit: forbid(unsafe_code) governs this project’s Rust code; transitive dependencies remain subject to separate review and supply-chain checks

The safe Rust subset provides all necessary functionality without compromising security.

---

FAQ

Q: What’s the relationship to the original Pi Agent? A: This is an authorized Rust port of Pi Agent by Mario Zechner, created with his blessing. The architecture differs significantly from the TypeScript original: it uses asupersync for structured concurrency and rich_rust (a port of Will McGugan’s Rich library) for terminal rendering. The goal is idiomatic Rust while preserving Pi Agent’s UX.

Q: Why rewrite in Rust? A: Startup time matters when you’re in a terminal all day. Rust provides a native single-binary deployment without a managed application runtime. Fresh comparative measurements are required before this release makes a speed claim.

Q: Can I use providers beyond Anthropic (OpenAI/Gemini/Cohere/Azure/Bedrock/Vertex/Copilot/GitLab/Cursor/Codex)? A: Yes. Pi has 11 native provider implementation modules: Anthropic, OpenAI Chat, OpenAI Responses/Codex Responses, Gemini (native + Gemini CLI + Antigravity routes), Cohere, Azure OpenAI, Amazon Bedrock, Vertex AI, GitHub Copilot, GitLab Duo, and Cursor. Pi also supports many OpenAI-compatible presets (for example Groq, OpenRouter, OrcaRouter, Mistral, Together, DeepSeek, Cerebras, DeepInfra, Alibaba/Qwen, and Moonshot/Kimi). Provider IDs and aliases are case-insensitive. Set credentials and choose via --provider/--model; run pi --list-providers to see canonical IDs, aliases, and env keys.

Q: How do sessions work? A: By default, each session is a JSONL v3 file with message entries, parent references for branching, and compaction metadata. Builds include sqlite-sessions support by default, so configured deployments can use SQLite-backed session storage too; JSONL remains the default store unless configuration selects SQLite.

Q: Why is unsafe forbidden? A: Memory safety is non-negotiable for a tool that executes arbitrary commands. The project does not trade that policy for performance.

Q: How do I extend Pi? A: Pi has a full extension system with two runtime families: JS/TS entrypoints run in embedded QuickJS, and *.native.json descriptors run in the native-rust descriptor runtime. Both are capability-gated and audited through the same policy system. One session uses one runtime family at a time. Extensions can register tools, slash commands, event hooks, flags, and custom providers. See EXTENSIONS.md for details. For built-in tool changes, implement the Tool trait in src/tools.rs.

Q: Why isn’t X feature included? A: Pi focuses on core coding assistance, and the default tool set reflects that. Browser automation, desktop control, image analysis/generation, and text-to-speech exist as settings-gated tools (browser, computer, inspect_image, generate_image, tts) so they cost nothing in the model’s schema unless you turn them on. Anything beyond that is out of scope; use specialized tools.

Q: How does compaction work? A: When a conversation exceeds the model’s context window, Pi summarizes older messages using the LLM itself, storing the summary as a session entry. Recent messages are kept verbatim. The cut point is chosen at a turn boundary, and the summary includes a record of which files were read or modified so the model retains that awareness. Compaction runs automatically after each agent turn when needed, or manually via /compact.

Q: Can I add a custom provider that Pi doesn’t support natively? A: Yes. Create a models.json file in ~/.pi/agent/ or .pi/ with entries specifying the model ID, base URL, and API type (usually openai-completions for OpenAI-compatible endpoints). Pi’s compat config system handles field name differences and feature flag overrides. Extensions can also register entirely custom providers.

Q: How does Pi decide which session to resume? A: Pi maintains a SQLite session metadata index sidecar with WAL/lock handling and stale-index reindexing. When you run pi -c, it queries that index for the most recently modified session whose working directory matches your current project, including JSONL sessions and configured SQLite-backed sessions. This avoids scanning the filesystem on every resume.

Q: What happens if an extension tries to access something dangerous? A: Every hostcall from an extension is checked against the active capability policy before execution. Dangerous capabilities (exec, env) are denied by default under safe and balanced unless explicitly opted in (for example via PI_EXTENSION_ALLOW_DANGEROUS=1), and are available under permissive. For exec, Pi then applies command mediation before spawn: it classifies command+arg signatures and blocks critical classes by default (for example recursive delete, disk/device write, reverse shell), with strict/safe policy able to block high-tier classes as well (for example shutdown, process-kill, credential-file modification). Denied calls return errors to the extension Promise path, and denial events are recorded in redacted security-alert and exec-mediation audit