We only highlight major and minor versions here. Patches and hotfixes ship to GitHub Releases — the source of truth for the full changelog.
2026-09-15
- browser — visible hand-off for user logins, attach to the user’s own Chrome and second-tier verbs
- browser — report a closed page before reattaching, and a quit browser as its own result
- browser — survive a closed tab, pass Google’s automation check and tear the browser down after one-shot runs
2026-09-14
- cost — attribute every lost prefix to the request region that changed
- memory — cache-shaped extraction request and the memory balance in /cost
- prompts — bind specific claims to tool evidence in every mode prompt
- skills — generalize learned skills at authoring time and measure tool use per run
- cost — keep background and worker traffic out of the prefix telemetry and count the orchestrator’s own progress
2026-09-14
- cache — refresh the cache entry through long tool calls on every adapter that can
- cost — persist the prompt-cache telemetry with the session and expose the write/read ratio
- prompts — size the reasoning block to the task in every mode prompt
- skills — measure each skill against the baseline of every run
2026-09-13
- agent — keep the coder system prompt byte-stable across runs
- cache — scope the cache lifetime per tenant and keep Fable entries warm
- cost — honest cache accounting in /cost and a per-request cache log
- agent — stop rewriting the cached prefix on every turn of the coder loop
2026-09-10
- config — register the compression knobs in the defaults registry and name them in /config
- storage — inventory and on-demand curation of every local store
- cost — price Devin turns without a listed rate and bill enterprise cache reads once
- storage — complete only what the command still accepts
2026-09-09
- agent — split the workspace block by lifetime and de-duplicate skills across runs
- chat — curate the prefix behind a recovery tool, and cache what is stable
- usage — normalize input token accounting across every provider schema
2026-09-07
- catalog — o valor GITHUB_MODELS sai do enum de provider do CRD Instance e do schema do chart Helm. ChatCLIInstance que ainda aponte para ele passa a ser rejeitado na admissao e precisa migrar de provider antes do upgrade. As env vars GITHUB_MODELS_MODEL, GITHUB_MODELS_TOKEN, GITHUB_MODELS_MAX_TOKENS e GITHUB_MODELS_API_URL deixam de ter efeito, e /auth login github-models deixa de existir.
- catalog — add GPT-6 Astra and drop retired GitHub Models
- readme — document the GitHub Models removal and pin the next version
2026-09-05
- auth — the keychain setting starts doing something, on every platform including Windows
- deploy — the chart can configure what the security page says it can, and the CORS setting reaches the server
- plugins — the signing toolchain and quarantine the docs described
- server — RS256, mutual TLS, real roles, and an audit trail that names the caller
- agent — the allowlist covers the whole line, and coder test stops being the way around it
2026-09-05
- agent — a run under a spending ceiling tells the model how much room is left
- knowledge — a passage carries where it came from into the index
- llm — per-turn context as a turn-scoped system message
- llm — the OAuth path carries per-turn context like the others
- llm — xAI and MiniMax count their own tokens
- agent — tool schemas stop travelling once they crowd out the conversation
- cache — an advisable compaction waits while the prefix cache is warm
- cli — per-run state that is per run, and a seal that counts once stored
- cli — per-turn context follows what the model still reads
- cli — reasoning blocks are part of a journaled message identity
- coder — a write may not leave the workspace, and the denylist works on macOS
- config — every variable the code reads is visible in /config, enforced by a test
- context — the per-turn block stops repeating itself into the window
- context — the provider engine reaches Bedrock and can summarize server-side
- cost — the calibration sample measures what the request actually sent
- ctxmgr — refresh adopts the current segmenter and situating seals
- gateway — an inbound message with no valid signature is refused
- llm — a turn-scoped system message goes where the provider accepts it
- llm — cache decisions hold for the conversation, thresholds follow the window
- llm — reasoning blocks travel back with the turn that produced them
- llm — the cache breakpoint survives a trailing system message
- llm — the effort a turn asks for reaches the request on every provider
- llm — the task budget travels, on both providers that read it
- server — refuse to serve an unauthenticated API on a reachable address
2026-09-04
- audit — record every LLM request on every surface in the audit trail
- cache — explicit Gemini cache resources, storage pricing and auto TTL
- cache — rolling conversation breakpoint, extended TTL and cache telemetry across providers
- compaction — session threshold, dedicated summarizer, tool-catalog reserve and repeated-read dedup
- config — organization-managed defaults and locked policies
- context — compaction hooks, verbatim-safe levels, summarizer accounting, 1h cache TTL on Bedrock/OpenRouter
- context — incremental re-index with file stamps, /context refresh and /context watch
- context — live context inspector and learned chars-per-token calibration
- context — persist the token calibrator per state root and route every estimate through it
- context — prefix budget allocator, projected ctx%, /clear conversation semantics
- gateway — per-principal store sets under hub isolation
- gateway — tenant-scoped park and task-graph roots and a per-tenant daily budget
- knowledge — opt-in lexical normalization, fact IDF, local embeddings with cost, retrieval budget and citations
- memory — export/import, multi-process merge for profile/topics/projects, recall reasons and project labels
- memory — pluggable memory provider and context engine through MCP
- memory — vector-aware auto-recall with a relevance floor and session-persisted context attachments
- retention — expire park and cost snapshots on the session TTL and show every store’s policy
- retrieval — rerank stage, corpus weights, structure-aware passages and evidence-based recall
- security — encryption at rest for memory, contexts, CCR and costs, key rotation and a hash-chained audit trail
- session — append-only transcript journal and memory flush before compaction
- session — undo compaction, persisted rewind checkpoints and transcript export/search
- telemetry — OpenTelemetry metrics export over OTLP/HTTP and a provider-side context engine
- tokens — provider token counting on every surface, exact calibration and budgeted summaries
- workspace — nested AGENTS.md hierarchy with imports, nested watch dirs and legacy source paths
- agent — follow the @model route in delegation and never report an empty worker as success
- agents — workers share the orchestrator’s compactor, overflow recovery and journal
- cache — provider cache parity — Bedrock system marker TTL, GPT-5.5 retention, cache key on OpenAI-compatible upstreams, OpenRouter marker cap
- claudeai — surface SSE error events and decode br/zstd responses
- config — registry, reload and surface truth — missing defaults, reloadable variables, /config memory rows, /help lines, completer entries, sections hint from the router, dead code
- context — byte-stable system prompt with per-turn context as a flagged user message, stable cache key, frozen prefix ratio
- context — cache-stable prompt prefix, reversible guided compact and fsync durability
- context — close the P0 gaps of the second context audit
- context — mirror provider context edits locally — stub cleared tool results, skip calibration, note the rebuild, surface freed tokens
- context — recover from context overflow on chat, RPC, one-shot and MoA; classify every provider’s overflow body; per-session recovery budgets
- context — single context estimate and window-scaled retrieval budget
- context — verbatim results survive pairing, verbatim cap for convergence, skipped-compaction bookkeeping, shared restore helper
- cost — attribute usage to the call that produced it — summarizer, memory worker, streaming, Gemini thinking, provider resets
- cost — cache telemetry truth — subset-schema total misses, per-provider first and hit ratio, expected rebuilds on every prefix change, Bedrock TTL as sent
- cost — long-context tiers per call, Amazon Nova rows, Copilot as a subscription
- cost — reset clears every aggregate, daily-aware budget notices, summarizer cache keyed by env, locked audit path fails closed
- cost — telemetry semconv and opt-in session series, embedding and daily-budget metrics, CSV export, spend finalized on every surface, park on hard stop
- gateway — swap every per-conversation state with the tenant, guard the base memory worker, tag external memory calls, keep the workspace on queued segments
- knowledge — batched, single-flight embedding with skip-and-continue; binary/minified skip; passage hard cap; partial upsert persistence; rune-safe cuts
- knowledge — hybrid retrieval as a vector-BM25 union fused by RRF, query embedded once per turn, system-only tagged reranker
- memory — cross-process lock around every merge-and-write of the shared stores
- memory — drain the queued backlog at boot and on every tick, bounded shutdown wait, exit-tail queueing, race-free watermark
- memory — fence recalled and retrieved text as data, keep injected blocks out of extraction and search, sanitize and validate stored facts
- memory — hygiene batch — rune-safe cuts, test-only helpers out of production, CCR skip surfaced, coder read line cap, transcript cap and plaintext notice, isolation refusal, external-only memory provider, worktree-aware projects
- memory — lock sealed stores read-only when the key is missing, keep daily notes plain, seal and redact the pending queue
- security — at-rest v2 — payloads bound to their store and tenant, passphrases stretched, fsync on reseal, 16-byte tenant digest, 0700 dirs
- security — encrypt session stores at rest and redact secrets on the LLM path
- security — never hand ChatCLI’s own secrets to MCP servers and executable plugins
- security — persisted-data hygiene — strict-policy redaction on every store, sealed REPL history, missing secret shapes, periodic tenant retention
- security — share one locked, rotating, sealable audit chain across the REPL, gateway and gRPC server
- session — exact schema gate, fork on its own journal with attachments, cost and CCR references, atomic state writes
- session — journal integrity — full-prefix hash scan, tool-call-aware hashes, torn-line tolerance, no tail dedup
- ui — footer shows used context apart from the answer reserve
- workspace — instruction-file parity — import boundary and per-file cap, global always merged, local and override files, .claude/rules, watched-dir cap with .gitignore
2026-09-02
- devin — list the models the account can invoke via the Devin CLI
- chat — count cached input tokens in the ctx% footer
2026-09-02
- models — refresh provider catalogs for September 2026
- operator — pin libexpat 2.8.4-r0 in the runtime image to clear CVE-2026-66046, CVE-2026-76641, CVE-2026-76956 and CVE-2026-76957
2026-08-30
- quality — route high-complexity auto plan-first to @taskgraph
- taskgraph — @taskgraph DAG orchestrator with executor-reviewer separation
- taskgraph — live browser dashboard for graph runs
- workers — capability parity — plugin grants, per-task skills, learning digest
- taskgraph — field-test hardening — plan-first deferral, prompt-safe streaming, plan-from-file
2026-08-29
- agent — doom-loop guard for identical repeated successful tool calls
- browser — @browser — Chrome real via DevTools protocol, zero deps novas
- coder — outline e map — orientação estrutural sem ler arquivos inteiros
- coder — post-edit diagnostics injected into the tool result
- coder — sandbox opt-in de exec no OS — macOS, Linux, degrade nos demais
- coder — shadow-git checkpoints with automatic snapshots and restore
- forge — @forge — PRs, issues e CI via gh/glab autenticados
- proc — interactive pseudo-terminal sessions with stdin driving
- skills — activation and memory parity across every surface + parser describe-leak fix
- view — @view — o agente enxerga imagens locais no meio da tarefa
- tools — @tools describe now renders @browser/@forge/@view subcommands
- tools — honor flattened —flag args in @browser/@forge/@view parsers
2026-08-27
- update — one-shot chatcli update subcommand for scripts and automation
2026-08-27
- zai — first-class GLM Coding Plan support and custom endpoint routing fix
2026-08-24
- cost — real usage across all providers, /cost subcommands, budget gates and session persistence
2026-08-18
- agent — add @embed tool for semantic text operations
- models — refresh image and embedding catalogs for Aug 2026
2026-08-16
- models — refresh provider catalogs and add Bedrock Mantle-to-runtime fallback
- deps — batch bump AWS SDK, x libraries and protobuf across both modules
2026-08-15
- agents — squad worker context navigation, recall and policy parity
2026-08-13
- commands — auto-route coder-mode slash commands from chat to a coder one-shot run
2026-08-13
- commands — interop matrix v2 — Codex, Gemini, Qwen, Cursor, Windsurf, opencode and Copilot
- commands — universal slash-command templates across every surface
- commands — tolerate real-world frontmatter, surface skipped files, refresh ACP command list mid-session
2026-08-11
- memory — make cross-session memory visible and recallable from turn one
2026-08-10
- context — leaner context window across chat and coder modes
- memory — persisted knowledge graph participating in recall
2026-08-06
- acp — terminal parity in client permission dialogs and session policy automode (c693e18)
2026-08-05
- cli — tool subcommand runs built-in tools directly without an LLM turn
- security — route unattended ask policy through client permission dialogs
2026-08-04
- catalog — add Kimi K3 to the Moonshot catalog with its own pricing tier
- catalog — refresh Devin CLI model roster
- session — cross-surface session continuity for REPL, ACP, MCP and gateway
2026-08-03
- agent — live type-ahead preview while the coder spinner runs
- agent — keep the orchestrator in the squad flow and make dispatch progress live
- agent — render the turn-spinner type-ahead preview below the line
2026-08-03
- agent — cross-process /agents via hub, mid-run side commands and skills in headless runs
- agent — resolve model inherit hint to the session model in squad dispatch
2026-08-03
- agent — agent squad with live run registry, work board, squad mail and autonomous delivery playbook (793cd44)
2026-07-30
- cli — announce new release at boot and resolve Devin CLI from known install dirs (15f8981)
2026-07-30
- acp — structured session updates, slash commands and IDE permission dialogs
- cli — complete /update check and close completer and palette gaps
2026-07-28
- update — announce new releases in-session and revamp version UX
2026-07-27
- providers — list all gateway models when endpoint URL is customized
2026-07-24
- models — add Claude Opus 5 to the Anthropic, Bedrock and OpenRouter catalogs
- update — channel-aware auto-update with /update command and staged background mode
2026-07-21
- bedrock — BEDROCK_MODEL selects the default model
2026-07-21
- bedrock — custom endpoint overrides for corporate DNS and VPC
2026-07-20
- session — proactive cross-session recall with tolerant search and titles
2026-07-19
- park — capture user messages typed while parked as resume directives
2026-07-17
- session — MCP session autosave by default and bounded machine-session lifecycle
2026-07-17
- memory — derive rollups from the episodic record
- memory — episodic timeline with temporal recall
- memory — proactive auto-recall in index mode
- session — sessions become retrievable memory
- devin — drop identity coercion and accept tool tag alias
2026-07-16
- mcp-server — full ChatCLI experience parity over MCP
2026-07-14
- agent — @model tool — AI-driven model/provider routing with tiered catalog and delegate
2026-07-13
- mcp — re-export connected MCP servers’ tools on the mcp-server surface
2026-07-12
- ui — add ui/kit presentation components and fix help header styling
- ui — adopt the sóbrio borderless treatment across chat, agent and coder
- ui — align command surfaces on kit padding, rules and notices
- ui — make the welcome screen responsive and theme-aware
- ui — sweep agent timeline glyphs and rules onto the kit vocabulary
- cli — restore spinner, type-ahead queue and AltGr input in the Windows chat REPL
- i18n — honor CHATCLI_LANG from .env and detect the Windows locale
- i18n — sentence-case display titles and notice-based session status
- i18n — strip variation selectors and embedded rules from the catalogs
- ui — give warnings their own theme role
2026-07-11
- mcp — OAuth 2.1 authorization for remote MCP servers
2026-07-10
- DEVIN provider — local Devin CLI as LLM transport
- full MCP + ACP server surface — every tool, harness routing, skills as prompts
- embedding — circuit-break Bedrock credential failures instead of hammering AWS
2026-07-10
- GPT-5.6 family (Sol/Terra/Luna) + Codex client headers fix
2026-07-10
- mcp — dynamic tool discovery on tools/list_changed notifications
2026-07-09
2026-07-07
- skills — match multi-word triggers in any word order
- skills — re-activate skills mid-loop in agent and coder modes
2026-07-03
- tools — expose MCP tools through the @tools catalog
2026-07-02
- agent — defer the tool catalog — index plus on-demand @tools describe
- compress — harden CCR reversibility and add runtime profiles
- http — structured HTTP client for the agent as the @http tool
- lsp — expose semantic code navigation to the agent as the @lsp tool
- proc — background-process supervision for the agent as the @proc tool
- ctxmgr — harden knowledge-mode durability, UTF-8 boundaries and retrieval cost
- lint — zero both module baselines and enforce full-module linting
- memory — eliminate silent memory-loss vectors and harden retrieval
2026-07-01
- models — Claude Sonnet 5 + GLM-5.2 + Bedrock Fable 5/Sonnet 5 support
2026-06-28
- graphview — interactive grap-style, graph in agent, coder and chat
2026-06-26
- live telemetry in agent/coder + on-demand CCR curation
2026-06-25
- self-evolution engine, in-core knowledge graph and memory quality
2026-06-22
- native context compression, image compression and output-token reduction
2026-06-21
- plugins — add @diagram builtin for embedded Graphviz rendering
- plugins — add selectable system/embedded Graphviz backend to @diagram (3ea68b5)
2026-06-18
- multimodal vision input + multi-provider image editing
2026-06-17
- docs-flatten — ingest code and infra repos into knowledge bases
- coder — recover stdin after Ctrl+C at a security confirmation
2026-06-15
- full @context parity + @registry-tags sort/last
2026-06-14
- context — autonomous knowledge-base management via @context
- plugins — add @registry-tags and @wikipedia builtins
2026-06-13
- transcription — pure-Go OGG/Opus decoding and actionable ffmpeg guidance
2026-06-12
- plugins — bring @docs-flatten in as a builtin tool
2026-06-11
- tls — global corporate TLS trust overrides for all outbound clients
2026-06-10
- context — knowledge mode: index-card attach + keyless hybrid retrieval (BM25 + embeddings) + @knowledge pull tool (agent/coder/chat)
- i18n — stop x/text from stamping MISSING into argless lookups
2026-06-10
- transcription — embedded offline Whisper STT via sherpa-onnx
2026-06-09
- webfetch — render JS pages via headless browser escalation
- websearch — keyless fallback providers + Claude Fable 5 in catalog
- embedding — rebuild embedding provider on /reload
- operator — resolve all 82 open gosec findings
2026-06-09
- context — semantic passage retrieval for /context attach
- memory — blend cosine, lexical and temporal signals in retrieval ranking
2026-06-06
- gateway — voice replies, embedded TTS and memory resilience
2026-06-06
- imagegen — Z.AI + MiniMax image backends, model-aware auto, and /model-image
2026-06-04
- proactive messaging, productivity skills pack, and multimodal agent tools
2026-06-03
- gateway — voice messages (transcription) + conversational replies
2026-06-01
- add interactive ask_user tool for agent, coder and chat
2026-05-31
- cli — interactive command palette overlay
2026-05-30
- cli — gateway runtime model, /model + /max-tokens, 9 new themes
- gateway — list stop in the /gateway usage hint
2026-05-29
- ui — unified color theme system + indentation-safe rendering
2026-05-29
- catalog — add Claude Opus 4.8 with adaptive thinking and fast mode
2026-05-27
- gateway — messaging gateway — Telegram, Slack, Discord, WhatsApp, Webhook
- hub — cross-channel conversation continuity via shared hub
- lsp — language server diagnostics via /lsp
- memory hybrid capture, security & multi-provider robustness (Tiers A+B+partial C)
- rpcserve — MCP server + ACP agent over stdio
2026-05-23
- aiops — close 5 gaps surfaced by 2026-05-23 chaos test
2026-05-23
- auth — TokenProvider with proactive OAuth refresh and retry-on-401
2026-05-21
- mcp — complete channel push delivery, reactive triggers and persistence
2026-05-20
- ui — cross-mode UI polish — i18n, red errors, lipgloss cards, unified banners
2026-05-17
- Claude Code parity tool layer (atomic tools, @todo, multipatch, schema validation, per-tool policies)
- tool layer enterprise refactor (security UX + parallel orchestration + provider-agnostic tool_result)
2026-05-15
- mcp — add Streamable HTTP transport (MCP 2025-03-26)
2026-05-15
- moonshot — add Moonshot (Kimi) provider with thinking-mode support
- qg — 15-floor Quality Gate overhaul — fix Floor 3, add 6 floors, redesign table
2026-05-12
- mcp — tier 1-3 server config extensions + autoApprove audit
2026-05-11
- skills — session-scoped pin for multi-turn auto-injection
2026-05-08
- mcp — per-server commands + logs ring + manual reload
- coder — use go-prompt for iteration input (paste + full width)
- mcp — inherit parent env, expand $, and hot-reload from empty state
2026-05-05
- Agent Park & Resume —
@parktool that suspends the agent loop, frees the terminal, and auto-resumes itself when the timer/probe completes. Four modes:delay,until,for_url,for_cmd— plus/park-noteand@park list/noteto send directives to a parked agent, delivered at resume. See Agent Park & Resume. - Foreground auto-resume — TIOCSTI on Unix and
WriteConsoleInputWon Windows automatically inject/resume <token>into the TTY; transparent fallback when the kernel restricts it (Linux 6.x+, macOS Ventura+). - Management slash commands —
/parked(list),/resume <token>,/cancel-park <token>with prefix-match and idempotency against the auto-resume race. - Operator SchemeBuilder migration — controller-runtime 0.24 deprecated
sigs.k8s.io/controller-runtime/pkg/scheme.Builder; migrated to apimachinery’sruntime.NewSchemeBuilder(–43 net lines).
2026-04-30
- Scheduler (Chronos) — cron, wait-until, DAG, and daemon mode with durable WAL
- Bedrock provider — Claude/Llama/Nova via AWS Bedrock with IAM/profile/credentials chain
- Token waste audit — cache in agent mode, cached tool defs, early-exit/routing
- Anthropic cache coalescer — local planner respecting the 4-marker cap without losing cache
- /config sections — became hierarchical (
/config <section>) with runtime mutation
2026-04-15
- MCP client + hooks + web tools — parity with Claude Code
- Cost tracking with per-provider price tables
- Plan mode + worktrees integrated
2026-03-29
- Harness/Quality Pipeline (7 patterns) — ReAct + Plan-and-Solve + Reflexion + RAG+HyDE + Self-Refine + CoVe + Reasoning Backbone
- Context injection via system prompt with Anthropic caching
- Bubble Tea TUI as default (go-prompt removed)
Track all releases on GitHub Releases or ArtifactHub.