Skip to content

Releases: Caupulican/pi-adaptative

v0.80.25

14 Jun 23:01

Choose a tag to compare

Fixed

  • Fixed package repository metadata so npm provenance publishing matches the fork repository.

v0.80.24

14 Jun 22:54

Choose a tag to compare

Added

  • Added resource profiles for scoped loading of extensions, skills, prompt templates, themes, context agent files, and tools, including zero-footprint user-level directory overlays and one-shot profile definitions.
  • Added embedded <resource-profile> blocks for extensions, prompt templates, skills, and context agent files; profile blocks are parsed as config and stripped from prompt-facing content.

Changed

  • Updated Pi Adaptative README copy to state the fork credit, purpose, and compatibility mode plainly.

v0.80.23

13 Jun 02:59

Choose a tag to compare

Fixed

  • Fixed JavaScript heap exhaustion in long-running sessions by bounding in-memory retention across the stack:
    • Session load now compacts oversized tool result details, so resuming or branching on top of large session files no longer pins their full payloads in memory.
    • Interactive scrollback components retain tool result details under a dedicated budget (512KB estimate per result), matching resumed-session semantics for pathological payloads while keeping designed display payloads (bash output window, diffs) intact.
    • Scrollback components share one built-in toolset per working directory instead of allocating a full toolset per tool call.
  • Bounded subprocess output retention:
    • pi.exec/execCommand keeps a rolling tail per stream (default 16 MiB, configurable via maxBuffer) and reports stdoutTruncated/stderrTruncated on the result.
    • The git output filter spills oversized git output (default over 48 MiB, PI_GIT_FILTER_MAX_RETAINED_BYTES to tune) to a temp file, discloses the cap in the filtered output, and reuses the spill file as the full-output artifact instead of materializing extra in-memory copies.
    • Package manager command capture is bounded and fails loudly instead of accumulating unbounded output.
  • Hot reload now unsubscribes the replaced extension generation's pi.events handlers from the shared event bus, so repeated reloads no longer pin old extension module graphs or double-process bus events.
  • Fixed abort-listener accumulation in retry backoff sleeps (utils/sleep.ts and the openai-codex provider): each completed sleep now detaches its listener instead of leaving it on the signal for the signal's lifetime.
  • Best-effort temp-file writes (bash full-output capture, git filter overflow spill) now handle stream errors instead of crashing the process with an uncaught error event on disk failures; the git overflow path discloses a failed spill in stderr instead of referencing a broken artifact.
  • Session file rewrites (migration, branch operations) are now atomic (write-then-rename), removing the torn-file window when a crash or a second pi process appending to the same session interleaves with a truncate-in-place rewrite.
  • Codex websocket debug stats and SSE-fallback flags are now cleared with the rest of the session resources on session replacement and dispose, instead of accumulating per session id for the process lifetime.
  • User input during an auto-retry backoff now queues as steering and is incorporated into the retried turn, instead of starting a concurrent run that raced and cancelled the pending retry. isRetrying is true from the moment auto_retry_start is observable, and the interactive editor routes submissions during the retry window through the steering path.
  • Oversized reads no longer spike the heap: the read tool streams line slices for files beyond 16 MiB (any region stays reachable in batches via offset continuation, with true line numbers), images beyond a 128 MiB pathology guard return guidance instead of loading, CLI @file attachments are bounded with a leading window plus a read-tool pointer, oversized SKILL.md files are skipped with a diagnostic, and the session loader skips lines too large to hold in a string instead of failing the whole resume.
  • Delimiter-less streams can no longer grow line-assembly buffers without bound: RPC JSONL input, the Anthropic SSE parser, and the Codex SSE parser each cap their buffers (64 MiB) and discard or fail cleanly.
  • Hot reload no longer freezes the UI: the chat scrollback rebuild is chunked with yields (also applies to resume, tree navigation, and display toggles). Plain Up arrow on an empty editor recalls queued messages for editing, and a >> prefix queues a follow-up message — both work in terminals that swallow the alt-chord bindings.

Pi Adaptative v0.78.4

01 Jun 04:26

Choose a tag to compare

Pi Adaptative v0.78.4

Changed

  • Restores eager startup injection for AGENTS.md, CLAUDE.md, and GEMINI.md context files.
  • Keeps skills and agent resources lazy-loaded to avoid startup prompt pollution.
  • Refreshes generated model metadata.

Fixed

  • Restores extension context handlers in the agent context transform pipeline after auto-compaction support.
  • Updates Pi Adaptative package-command help tests and pending tool-panel regression coverage for the reusable tool-panel registry.

Validation

  • npm run check
  • Focused coding-agent tests for system prompt/resource loading/tool rendering/args/paths/skills/package commands/extension context/pending tool panels
  • Published npm packages and install smoke for @caupulican/pi-adaptative@0.78.4

v0.78.3

01 Jun 03:02

Choose a tag to compare

Pi Adaptative 0.78.3

Highlights:

  • Added context-hygienic lazy startup resources: skills and context files are listed by location and loaded on demand.
  • Added cleaner interactive tool panels with session-scoped reusable/grouped rendering and shortened relative tool paths.
  • Increased Codex SSE response-header timeout and retained the published @caupulican/pi-adaptative package branding.
  • Synced selected upstream provider/session fixes without dropping fork changes:
    • large JSONL session files stream line-by-line for session listing/resume metadata;
    • OpenAI-compatible non-OpenAI providers avoid unsupported developer-role messages;
    • Anthropic Opus 4.7+ omits unsupported temperature and preserves thinking compatibility;
    • GPT-5.5 generated metadata omits unsupported minimal thinking.

Published npm packages:

  • @caupulican/pi-adaptative@0.78.3
  • @caupulican/pi-ai@0.78.3
  • @caupulican/pi-agent-core@0.78.3
  • @caupulican/pi-tui@0.78.3

v0.76.1

27 May 22:15

Choose a tag to compare

Pi Adaptative 0.76.1

  • Rebased the fork on upstream Pi 0.76.0.
  • Fixed Pi Adaptative update checks to use the @caupulican/pi-adaptative npm package metadata.
  • Fixed Bun binary fallback update instructions to point at Caupulican/pi-adaptative releases.

Published npm package: @caupulican/pi-adaptative@0.76.1