Skip to content

CHANGELOG.md

Latest commit

 

History

History
113 lines (81 loc) · 10.3 KB

CHANGELOG.md

File metadata and controls

113 lines (81 loc) · 10.3 KB

Changelog

All notable changes to this project will be documented in this file.

The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

[Unreleased]

Changed

  • Spawned child agents now inherit active registered parent tools executable in the child session, including MCP/extension tools such as ChunkHound when active and registered, while still excluding spawn and handoff and preserving child-local notebook tools.

0.3.0 - 2026-05-23

Added

  • Interactive /ledger TUI overlay — replaced the static notification popup with a full interactive overlay. Use arrow keys to navigate entries, press Enter to preview a selected entry body (truncated at 500 chars with ...), and Escape to close. Empty state shows a discoverability hint.
  • Visual handoff indicator — when /handoff is invoked, a live 🤝 Handoff in progress badge appears in the TUI status bar. Clears automatically when compaction completes, or when the agent finishes a turn without calling the handoff tool.
  • Ledger tool TUI renderers — ledger_add calls now render inline with a styled preview (✓ Saved "entry-name": first line...) in the conversation. Shows the full entry body when expanded.
  • Write-lock reentrancy detection — nested calls to saveLedgerEntry now throw an explicit error instead of silently corrupting the serialization chain.

Changed

  • Frame-based spawn scheduler — replaced the microtask-per-event render model with a scheduler that batches expensive component work at ~30 FPS. High-frequency streaming events (50–100+/sec) accumulate state cheaply per-event; layout, cache invalidation, and TUI invalidates are deferred to the next frame tick. Eliminates UI jank during bursty LLM streaming in child sessions.
  • ESM module type — added "type": "module" to package.json for compatibility with strict ESM projects.

Fixed

  • Stray ANSI reset codes in spawn shell — truncateToWidth no longer injects escape sequences that break background color styling in collapsed spawn renderer borders and padding.

0.2.0 - 2026-05-21

Added

  • Microtask event batching — rapid child session events are coalesced into a single parent invalidate per microtask boundary, preventing UI jank during bursty tool execution.
  • Epoch-based invalidation — the spawn renderer uses epoch counters rather than pointer comparison to detect stale sessions, making ownership checks reliable across session resets.

0.1.0 - 2026-05-21

Added

Context Management System Prompt

  • Automatic system-prompt injection — the LLM receives a context management primer at session start, teaching it how and when to use spawn, ledger, and handoff. No configuration needed.
  • Live ledger listing in system prompt — each session start injects the current ledger entries (name + first-line preview) so the agent always knows what stored knowledge is available without fetching blindly.

Spawn — Isolate Subtasks in Clean Child Contexts

  • spawn tool — delegate isolated work to an in-memory child agent with its own clean context. The child inherits the parent's model, thinking level, working directory, and built-in tools. Only the condensed result is returned — no context pollution in the parent.
  • Parallel execution — siblings run concurrently; the parent orchestrates and merges results. Tested for independent concurrent spawns producing correct results.
  • Child tool inheritance — children receive ledger tools (add/get/list) and the parent's built-in tools, but never the spawn or handoff tools, preventing recursive nesting.
  • Ledger-aware child prompts — child sessions are told what ledger entries exist so they can fetch knowledge on demand without the parent pre-loading bodies.
  • Child output truncation — child results are limited to 2000 lines / 50KB to prevent runaway output from overwhelming the parent. Truncated results include an advisory message.
  • Explicit model-required error — spawn fails immediately with a clear message when no model is configured, rather than with a cryptic runtime error.
  • No grandchildren — children cannot spawn further children, preventing explosive branching.
  • Session lifecycle management — children are properly cleaned up on completion, error, or parent session reset. Aborted children produce explicit aborted outcomes.
  • Signal-based cancellation — spawn respects AbortSignal, allowing the parent to cancel a running child mid-flight.

Ledger — Sparse Continuity Cache

  • ledger_add tool — save a named, compact continuity entry. Same-name writes overwrite (refinement). Writes are serialized via a process-local lock, preventing race conditions.
  • ledger_get tool — retrieve a full entry body by name. Returns the current entry listing alongside the body. Reports "not found" gracefully with the complete entry list.
  • ledger_list tool — list all entries with name and first-line preview. Returns empty-state placeholder when no entries exist.
  • Persistence across sessions — ledger entries survive context resets, handoffs, and session restarts. On session start, the extension scans the session history newest-to-oldest and rehydrates the in-memory ledger from the latest epoch.
  • Ledger entry persistence via custom entries — each ledger_add appends a versioned custom entry that survives compaction and session archival.
  • Epoch-based staleness — child ledger tools reject access with a clear "invalidated by reset" error when the parent session has been reset, preventing stale writes.

Handoff — Deliberate Context Compaction

  • handoff tool — triggers a deliberate compaction that replaces noisy context with a clean restart. The agent drafts a handoff brief preserving knowledge still missing from the ledger, then the session restarts with the brief at the top of context.
  • /handoff command — user-facing shortcut that asks the agent to draft a handoff brief from a user direction, capture reusable state to the ledger, and call the handoff tool — all in one turn.
  • Rich compaction summary — the handoff brief includes inlined ledger entries referenced in the task text (up to 3 entries, 4000 chars total), and a structured handoff primer explaining the continuation contract.
  • Clean slot handoff — compaction replaces all pre-handoff messages with just the handoff summary. Full history remains in the session file for later reference.
  • Post-handoff auto-resume — the agent receives a "Proceed." message after compaction completes, keeping momentum without manual re-prompt.
  • Enforcement tracking — the /handoff command tracks whether the agent actually called the tool, preventing the command from being silently ignored. The watchdog cleans up stale enforcement state.

Primacy-Zone Watchdog

  • Advisory context-usage reminders — a context hook injects a watchdog message before each LLM call when context usage exceeds 30%, 50%, or 70%. The message is advisory only — the agent decides whether to act.
  • Tiered nudge messages — at 30-50% the agent is reminded of the primacy-zone heuristic; at 50-70% it's advised to consider handoff soon; at 70%+ it's warned that automatic compaction may trigger soon.
  • Hidden from user — watchdog messages use a custom role with display: false, so they inform the LLM without cluttering the visible conversation.
  • agent_end tracking — the watchdog records the latest context usage percent after each agent run, available for diagnostic use.

TUI and Status Indicators

  • Context usage in status bar — a live ctx 65% indicator in the TUI status bar, color-coded: green (<30%), yellow (30-50%), orange (50-70%), red (≥70%).
  • Ledger count indicator — a 📒 N badge in the status bar showing the number of ledger entries, hidden when the ledger is empty.
  • /ledger command — shows all ledger entries with name, line count, and first-line preview in a TUI notification overlay.
  • Live spawn rendering — child agent sessions render inside the parent's TUI with:
    • Collapsed view: model name, thinking level, real-time "last action" summary (thinking, tool execution, or assistant text preview), 5-line preview of last assistant output, and token/cost statistics.
    • Expanded view: full chat history with 4-space indent, streaming assistant messages, and tool execution components.
    • Ownership lifecycle: session ownership transfers from state to the renderer on first render, with proper cleanup on dispose, reset, or reattachment.
  • Render caching — rendered output is cached by width/expanded/showImages, avoiding unnecessary re-renders when inputs haven't changed.
  • Graceful degradation — the spawn renderer recovers from malformed events, missing tool definitions, subscribe failures, and null results without crashing the parent TUI.

Extension Wiring and Lifecycle

  • Single state object — all primitives read/write through a single AgenticodingState instance, shared across extension hooks, tool executions, and the TUI renderer.
  • Session lifecycle hooks:
    • before_agent_start — injects context primer and ledger listing.
    • context — injects watchdog reminders before LLM calls.
    • session_start — rehydrates ledger on continuation; resets state on /new.
    • turn_end — updates TUI indicators.
    • agent_end — records context usage and manages handoff enforcement.
    • session_before_compact — consumes pending handoff task as the compaction summary.
  • One-line install — install via pi install npm:pi-agenticoding and disable platform compaction in settings.
  • Comprehensive test suite — 50+ tests covering spawn execution and rendering (concurrency, cancellation, truncation, stale detection, ownership lifecycle, microtask batching), ledger tools (add/get/list, staleness, rehydration, empty states, prompt hints), handoff (tool, command, compaction), watchdog (nudge injection, enforcement), and extension lifecycle.
  • MIT licensed — open-source permissive license.