Sponsored tool discovery, inline todo card, and cache-aware notices

Sponsored tool discovery, inline todo card, and cache-aware notices

Highlights

• New discover_tools capability lets the agent browse and select sponsored third-party tools in two phases (browse a category, then fetch setup instructions), with clear (sponsored discovery) provenance tags, substantive selection reasons, and an opt-out config
• A live todo card can be pinned inline in chat via /todos or Alt+P, and todos now track confidence history with a spike-aware validation gate
• Claude Sonnet 5 supports reasoning effort selection and is classified as a native 1M-context model

Improvements

• Named provider models from your config now appear in the model picker (issue #444)
• Endorsed-but-not-installed skills are explained instead of reported as 'Unknown skill' (issue #445)
• Cold prompt-cache notices are shorter, show how long ago the cache went cold, report effective prompt tokens, and the expiry countdown window scales with the cache TTL
• Documented harness cache invalidations are attributed in the TUI so cache misses are explainable
• Swarm plans render as a reliable, legible mermaid task graph and are labeled Plan in the todo widget
• The initiative tool no longer auto-opens the side panel

Fixes

• Explicit /login credentials now win over stale API-key environment variables inherited from the parent shell
• Setting prevent_sleep_while_streaming=false is honored by the TUI sleep guard
• Long silent reasoning phases no longer trip the client stall guard (issue #451)
• OpenRouter: local llama.cpp context windows are read from meta.n_ctx (issue #447), and tool schemas are sanitized for strict OpenAI-compatible endpoints (issue #446)
• default_provider pointing at an OpenAI-compatible catalog profile is honored (issue #448)
• Server-path message hashing uses a cache-relevant projection, avoiding false KV-cache misses
• Mermaid aspect-profile diagrams no longer draw outside their render scope, streaming previews clear on empty-buffer commits, and /clear fully re-scopes active diagrams
• Inline-image marker text is never drawn to the terminal
• Rewinding a session mid-stream drops stale streaming state instead of replaying it

Full changelog: v0.36.0...v0.37.0

TB2.1 confstep bench build (943eae93)

jcode v0.36.15-dev 943eae9 with spike-aware confidence gate, for benchmark harness in-container install

v0.36.0

Image pipeline hardening, big memory cuts, and swarm orchestration upgrades

Highlights

• Inline images are hardened end to end: payloads are released after decode (~8MB back per image-heavy session), evicted cache entries self-heal from disk, and undecodable images can no longer spin the render loop
• Client memory drops sharply: syntect switches to the onig backend (~17MB), the session-search index is bloom-filtered (-24MB server-side), retained heap is trimmed on idle, and glibc mmap thresholds are pinned so large transients return to the OS
• Swarm plans render as a live inline mermaid task graph in chat, with per-agent task labels, honest run_plan progress, and per-spawn model/effort selection

Improvements

• Swarm: broadcasts and shared context are scoped to the sender's spawned subtree
• Swarm: tasks stranded on dead workers are salvaged and reported to the coordinator; orphaned children are reparented when a mid-tree member leaves
• Swarm: run_plan pauses on credential-failure waves and surfaces the fix instead of burning retries
• Swarm: state persistence moved to a durable directory so it survives reboots
• Swarm: long messages and reports require a tldr and collapse to it with an expand badge
• Composed input text is mouse-selectable and copyable (issue #430)
• Interrupts fan out to all active turn signals so cancel is immediate (issue #428)
• Fable 5 defaults to low reasoning effort; effort=none now fully suppresses thinking
• Provider runtimes (Anthropic, OpenAI, OpenRouter, Copilot, Cursor, Gemini, Antigravity) moved out of the base crate, so provider edits rebuild far less
• Resumed sessions survive corrupt journal lines and keep the last prompt
• agentgrep fixes common agent failure modes and truncates huge trace lines

Fixes

• Scroll no longer hangs when older history is prepended to a live transcript (issue #344)
• Sidebar context figure stays consistent with compaction accounting (issue #441)
• Auto provider init registers external runtimes in every entry path, so OpenAI-compatible profile routes always reach the model picker
• Resume no longer deadlocks when the live-attach target agent is busy
• Auto-poke continuations no longer render as user prompts after reload
• Copy badges no longer truncate blockquote content

Full changelog: v0.35.1...v0.36.0

v0.35.1

Complete the 0.35.0 release for macOS and FreeBSD

Highlights

• Restores macOS and FreeBSD release binaries, which were missing from 0.35.0 due to a platform-specific build error

Improvements

• All 0.35.0 changes apply: inline mermaid diagrams, swarm deep-mode task graphs, and one-keypress model reroute after refusals

Full changelog: v0.35.0...v0.35.1

v0.35.0

Superseded by v0.35.1, which adds the macOS and FreeBSD binaries missing here due to a build error.

Highlights

• Mermaid diagrams now render inline in chat like images, on by default, with aspect-aware sizing and cached re-rasterization on resize
• Swarm deep mode grows a live task graph with gate audits and renders it as a live mermaid plan diagram, plus a swarm dock widget with a live agent list
• After provider guardrail refusals or terminal errors, jcode offers a one-keypress reroute to a fallback model instead of silently ending the turn

Improvements

• Anthropic models gain a 'max' reasoning effort rung above xhigh
• GitHub release notes are now generated from the human-readable changelog
• Skills installed via Claude Code plugins and marketplaces are picked up automatically
• MCP servers can be individually disabled with a per-server enabled flag
• FreeBSD x86_64 release binaries are now shipped
• /login guides you through provider filtering instead of instant-launching a login
• Onboarding imports logins with a summary-first flow and a one-time resume picker
• Session search got faster with incremental per-source indexes, and startup connect bursts use much less CPU
• Rendered markdown blockquotes show a copy badge
• New scripts/uninstall.sh performs a clean wipe
• Swarm agent lifecycle transitions are surfaced as status notices, and swarm activity tails interleave live tool output

Fixes

• Rapid Esc presses are no longer swallowed by a stale cancel reset
• Project-local MCP config loads from the session working directory instead of the server's
• Reasoning effort picked in the model picker is forwarded correctly in remote sessions
• Real-world network outage errors are retried as transient instead of failing the turn
• Stream stall detection now works across all providers, not just one
• Scrolling no longer flips the scrollbar layout or causes flicker during overscroll
• Webfetch URLs and search queries survive display compaction
• Swarm spinners keep animating while the coordinator processes without streaming
• The pinned diagram pane is capped so it never crushes the transcript
• /rewind N now matches the numbered transcript entries shown in the TUI
• Fixed border shear from ambiguous-width glyphs in info widgets

Full changelog: v0.34.0...v0.35.0

v0.34.0

Full Changelog: v0.33.0...v0.34.0

Full Changelog: v0.33.0...v0.34.0

v0.33.0

What's Changed

• Triage fixes: install.sh, /config edit, FreeBSD, .agents skills, MCP compat, retry fail-fast, terminal config by @1jehuang in #417
• Make master CI green: fmt, clippy, e2e transport flake, budget baselines by @1jehuang in #419

Full Changelog: v0.32.0...v0.33.0

v0.32.0

Full Changelog: v0.31.2...v0.32.0

Full Changelog: v0.31.2...v0.32.0

v0.31.2

Full Changelog: v0.31.1...v0.31.2

v0.31.1

Full Changelog: v0.30.3...v0.31.1

Full Changelog: v0.30.3...v0.31.1