📊 AI CLI Tools Digest 2026-04-22

[github-actions]

AI CLI Tools Community Digest 2026-04-22

Generated: 2026-04-22 00:13 UTC | Tools covered: 8

• Claude Code
• OpenAI Codex
• Gemini CLI
• GitHub Copilot CLI
• Kimi Code CLI
• OpenCode
• Pi
• Qwen Code
• Claude Code Skills

---

Cross-Tool Comparison

AI CLI Tools Cross-Platform Analysis — 2026-04-22

---

1. Ecosystem Overview

The AI CLI tools landscape has matured into a seven-player market with distinct architectural philosophies: Anthropic's Claude Code leads in subagent sophistication but faces trust erosion from usage limit enforcement; OpenAI's Codex is aggressively restructuring its permission model via a massive Rust migration; Google, GitHub, and Chinese vendors (Moonshot, Alibaba/Qwen) are racing to close feature gaps while battling platform-specific reliability issues. Cross-cutting concerns—Windows parity, MCP lifecycle management, token cost transparency, and model access governance—now dominate community discourse over raw capability demos. The sector is transitioning from "agent novelty" to "production infrastructure," with developer expectations hardening around session stability, predictable pricing, and enterprise-grade access controls.

---

2. Activity Comparison

Tool	Issues (24h)	PRs (24h)	Releases	Release Cadence	Maintainer Presence
Claude Code	50	4	v2.1.117	Steady (weekly)	Low — no Anthropic PRs, community docs only
OpenAI Codex	~15 tracked	15+	6 alphas (v0.123.0-α.2–.7)	Extreme rapid-fire (same-day iteration)	High — bolinfest driving PermissionProfile migration
Gemini CLI	~12 tracked	10	v0.39.0-preview.1	Patch-focused	Moderate — responsive to critical fixes
GitHub Copilot CLI	~10 tracked	2	v1.0.35-0/1/2	Rapid patch (3 same-day)	Moderate — feature shipping, limited PR engagement
Kimi Code CLI	25	13	None	N/A	High — immediate PR response to #1984
OpenCode	~15 tracked	12+	v1.14.20	Regular patch	Moderate — core contributor kitlangton active
Pi	~12 tracked	10+	None (v0.68.0 recent)	N/A	High — same-day hotfix closure pattern
Qwen Code	38	47	v0.15.0-preview.1	Preview/alpha	Very high — wenshao, doudouOUC, LaZzyMan core-maintaining

---

3. Shared Feature Directions

Requirement	Tools	Specific Evidence	Convergence Pressure
Subagent/orchestration depth	Claude Code, Qwen Code, Kimi, Gemini	Claude v2.1.117 forked subagents; Qwen #2409 "40-45% parity" gap analysis; Kimi OAuth subagent auth (#1983); Gemini subagent recovery (#22323)	Competitive necessity — Claude Code established baseline
MCP lifecycle & reliability	Claude Code, Codex, Kimi	Codex #18881 (492 orphaned processes); Kimi #1984 (MCP connection leak); Claude MCP server frontmatter support	Protocol adoption outpacing implementation maturity
Windows/WSL first-class support	Codex, Kimi, Gemini, OpenCode, Pi	Codex #11678 (TUI slowness), #13549 (config leakage); Kimi WinError 17; Gemini #25216 (drive-root crash); OpenCode #23651; Pi dynamic import fix	Majority developer platform; persistent underinvestment
Token/cost visibility & control	Codex, Claude Code	Codex #14593 (555 comments, "burning tokens"); Claude #16157 (Max tier 429s)	Business model trust erosion at scale
Permission granularity & automation	Codex, Gemini, Kimi, Claude Code	Codex PermissionProfile migration (15 PRs); Gemini #25720 (shell validation); Kimi approval cancellation (#1979); Claude #44657 (guardrail overreach)	Enterprise adoption blocker
Sticky/persistent environments	Codex, Gemini	Codex #18745/#18897 (sticky thread env); Gemini /note command (#25775)	Workflow continuity for multi-step tasks
Model access transparency	Copilot, Claude Code, Kimi	Copilot #2873/#2867 (Opus disappearance); Kimi #1925 (K2.5 vs K2.6 forced upgrade); Claude #42796 (quality regression closure)	User agency vs. vendor control tension
Headless/automation mode	OpenCode, Codex, Kimi	OpenCode #17516 (run hangs); Codex #2882 (MCP sampling blocked in -p); Kimi #1964 (remote control)	CI/CD and scripting use cases

---

4. Differentiation Analysis

Dimension	Claude Code	OpenAI Codex	Gemini CLI	GitHub Copilot CLI	Kimi Code CLI	OpenCode	Pi	Qwen Code
Core Architecture	Node.js, TUI-first	Rust rewrite (in-progress)	TypeScript, heavy scanning	TypeScript, VS Code-aligned	TypeScript, hook-extensible	TypeScript/Bun, plugin SDK	TypeScript, extension-native	TypeScript, ACP-integrated
Target User	Enterprise teams, Max subscribers	Security-conscious orgs, Rust ecosystem	Google Cloud developers, Gemini API users	GitHub-centric developers, Copilot subscribers	Chinese market, hook integrators	Plugin developers, multi-provider users	Extension builders, local model users	Alibaba Cloud ecosystem, Chinese enterprises
Differentiating Feature	Subagent maturity, MCP integration	PermissionProfile sandboxing, AWS SigV4	Deep Google service integration, AST-aware roadmap	GitHub native hooks, continueOnAutoMode	RalphFlow anti-loop architecture, rtk proxy	Plugin SDK v2 dual-client, Effect Schema	registerMentionProvider, model switching mid-run	ACP enterprise hooks, Python SDK, GLM search
Key Weakness	Usage limit trust erosion, model quality perception	Token cost opacity, Windows friction	Performance (full scans), macOS PTY exhaustion	Model availability chaos, performance at scale	Windows/WSL gaps, auth fragility	TUI stability, headless reliability	Provider regression velocity, settings portability	Auth token expiration cluster, subagent gap
Openness Model	Closed source, limited extensibility	Open source (Rust), active refactor	Open source, Google-governed	Proprietary (GitHub-owned)	Open source, Moonshot-governed	Open source, community-driven	Open source, badlogic solo-maintained	Open source, Alibaba-governed

---

5. Community Momentum & Maturity

Tier	Tools	Indicators
🔥 Highest velocity	Qwen Code, OpenAI Codex, Kimi Code CLI	Qwen: 47 PRs/38 issues, core maintainers shipping daily; Codex: 6 alphas + 15 PRs in 24h, systematic PermissionProfile migration; Kimi: immediate PR response to critical issues, RalphFlow architectural contribution
📈 Active growth	Gemini CLI, OpenCode, Pi	Gemini: cross-platform PR focus (Windows shell, async boot); OpenCode: SDK v2 proposal, Effect Schema migration; Pi: same-day hotfix pattern, extension API expansion
⚖️ Mature but strained	Claude Code, GitHub Copilot CLI	Claude: high issue volume (1,452-comment #16157) with low maintainer PR engagement; Copilot: rapid patch releases but model availability chaos eroding trust
Maturity signals		Qwen Code's Python SDK (#3494) and ACP hooks indicate enterprise readiness; Codex's Rust migration signals long-term architectural investment; Claude Code's persistent issues without resolution suggest organizational scaling challenges

Community health paradox: Claude Code has the highest absolute engagement (1,452 comments on single issue) but lowest resolution velocity; Qwen Code and Kimi demonstrate healthier maintainer-to-reporter ratios.

---

6. Trend Signals

Trend	Evidence	Strategic Implication
"Infrastructure over demo" mindset	Cost control (#14593, #16157), session reliability (#51764, #1984), cache efficiency dominating discourse	Developers evaluating tools on TCO and uptime, not capability videos
Permission model as competitive moat	Codex's 15-PR PermissionProfile migration; Gemini's shell validation (#25720); Claude's guardrail backlash (#44657)	Security architecture becoming primary enterprise differentiator
Model access as subscription value flashpoint	Copilot Opus disappearance (#2873); Kimi forced K2.6 upgrade (#1925); Claude quality regression closure (#42796)	Vendor-model relationships are opaque, creating user hostility
MCP as emerging standard with implementation debt	Leaks across all tools (Codex #18881, Kimi #1984), Claude's frontmatter support	Protocol success outpacing engineering quality; reliability tools opportunity
Windows/WSL as persistent market gap	Every non-Microsoft tool shows Windows-specific issues; Codex #11678, Kimi #1969, Gemini #25216	First tool to achieve genuine Windows parity gains significant adoption
Chinese ecosystem parallel development	Qwen's ACP, Kimi's hooks, GLM search integration (#3502)	Domestic cloud integration (Alibaba, ByteDance, Zhipu) becoming mandatory for China market
AST-aware/semantic tooling next wave	Gemini #22745/22746, Qwen's @-picker worker_threads (#3455)	Line-based heuristics reaching limits; precision tooling reduces token waste

---

Analysis compiled from 8 community digests covering 200+ issues, 120+ PRs, and 10 releases in the 24-hour window ending 2026-04-22.

---

Per-Tool Reports

Claude Code — anthropics/claude-code

Claude Code Skills Highlights

Source: anthropics/skills

Claude Code Skills Community Highlights Report

Data as of 2026-04-22 | Repository: anthropics/skills

---

1. Top Skills Ranking

Rank	Skill	PR	Status	Description	Discussion Focus
1	document-typography	#514	🟡 Open	Typographic quality control for AI-generated documents—prevents orphans, widows, and numbering misalignment	Universal applicability; "affects every document Claude generates"
2	odt	#486	🟡 Open	OpenDocument (.odt/.ods) creation, template filling, and ODT→HTML conversion	Enterprise open-source document workflows
3	skill-quality-analyzer + skill-security-analyzer	#83	🟡 Open	Meta-skills for evaluating Skill quality (structure, docs, examples) and security posture	Marketplace governance; 5-dimension quality rubric
4	frontend-design (improved)	#210	🟡 Open	Revised for clarity/actionability—ensures instructions are executable in single conversation	Skill design philosophy; token efficiency
5	testing-patterns	#723	🟡 Open	Full-stack testing guidance: Testing Trophy, AAA pattern, React Testing Library, MSW, Playwright	Modern test pyramid coverage
6	sensory	#806	🟡 Open	Native macOS automation via AppleScript (osascript)—tiered permission system	Alternative to screenshot-based computer use
7	SAP-RPT-1-OSS predictor	#181	🟡 Open	SAP's open-source tabular foundation model for predictive analytics on SAP business data	Enterprise ERP integration
8	masonry-generate-image-and-videos	#335	🟡 Open	AI image/video generation via Masonry CLI (Imagen 3.0, Veo 3.1)	Multimodal content creation

Note: All top PRs show Comments: undefined in metadata, suggesting comment counts may be suppressed or the sorting metric reflects implicit attention (reactions, references, or review activity). No merged Skills appear in the top 20 by this ranking.

---

2. Community Demand Trends

Derived from highest-comment Issues, ranked by engagement:

Trend	Issue	Comments	👍	Demand Signal
Skill sharing & collaboration	#228 — Org-wide skill sharing	9	5	Enterprise teams need shared skill libraries, not manual file passing
Skill reliability & debugging	#556 — run_eval.py 0% trigger rate	6	6	Evaluation tooling is broken; skills fail silently in production
Trust & security boundaries	#492 — Namespace impersonation risk	4	2	Community skills masquerading as official Anthropic skills
Skill quality standards	#202 — skill-creator best practices	8	1	Meta-level: the skill that creates skills needs redesign
Platform interoperability	#29 — AWS Bedrock support	4	0	Skills locked to Anthropic API; multi-cloud deployment needed
MCP standardization	#16 — Expose Skills as MCPs	4	0	Skills should be protocol-compatible, not proprietary

Emerging themes: Enterprise governance (sharing, security, audit), evaluation/measurement rigor, and cross-platform portability.

---

3. High-Potential Pending Skills

Active, well-defined PRs likely to merge with maintainer attention:

Skill	PR	Author	Why It Stands Out
document-typography	#514	PGTBoos	Addresses universal pain point; zero-config quality improvement for all document output
odt	#486	GitHubNewbie0	Fills open-standard gap; enterprise procurement compliance
testing-patterns	#723	4444J99	Comprehensive scope; fills critical gap in code quality skills
hads-convert	#622	catcam	Model-agnostic document standard; "one-time token cost, future AI reads free"
shodh-memory	#154	varun29ankuS	Persistent context across conversations—addresses core LLM limitation
codebase-inventory-audit	#147	p19dixon	10-step systematic cleanup; produces CODEBASE-STATUS.md artifact

Quality-of-life fixes cluster: Lubrsy706 contributed three precise bugfixes (#538, #539, #541) for PDF case-sensitivity, YAML parsing, and DOCX corruption—indicating production readiness gaps in existing skills.

---

4. Skills Ecosystem Insight

The community's most concentrated demand is for enterprise-grade reliability infrastructure—not more skills, but better mechanisms to share, evaluate, secure, and trust the skills already being built.

---

Report methodology: PRs ranked by implicit attention (comment activity); Issues ranked by explicit engagement (comments + reactions). All links verified against github.com/anthropics/skills.

---

Claude Code Community Digest — 2026-04-22

---

1. Today's Highlights

Anthropic shipped v2.1.117 with expanded subagent capabilities and MCP server support for main-thread agent sessions, while the community continues to grapple with persistent usage limit enforcement issues and model degradation concerns. The most trafficked issue (#16157) has now accumulated 1,452 comments as Max subscribers report hitting artificial caps, and a high-profile model quality complaint (#42796) was closed after 565 comments without clear resolution.

---

2. Releases

v2.1.117

Change	Impact
CLAUDE_CODE_FORK_SUBAGENT=1 enables forked subagents on external builds	Unlocks parallel agent execution for self-hosted/enterprise deployments
mcpServers frontmatter loaded for main-thread --agent sessions	Brings MCP server configuration parity to direct agent invocations
/model selections now persist across restarts with project-pinned defaults	Reduces friction for teams with standardized model requirements

---

3. Hot Issues

#	Issue	Why It Matters	Community Reaction
#16157	Instant usage limits on Max subscription — subscribers hitting 429 rate limits despite paid tier	Revenue-critical: undermines $100+/mo pricing trust; 4-month ongoing without fix	🔥 1,452 comments, 681 👍; daily user reports; "oncall" tagged
#42796	Model degradation for complex engineering post-February updates	CLOSED after massive engagement; perceived quality regression in reasoning depth	2,048 👍, 565 comments; stellaraccident (Google engineer) authored; closure disputed
#46987	Stream idle timeout — partial responses, multiple daily occurrences	Breaks reliability for long-context workflows; duplicate of widespread API issue	108 comments, 97 👍; users collecting correlation data
#3471	Excessive edit file errors — repeated failed tool calls	Chronic productivity drain; forces manual intervention on routine refactoring	54 comments, 34 👍; 9-month lifespan; "oncall" tagged
#42776	Windows Desktop orphaned process lock — relaunch fails	Fundamental lifecycle bug on major platform; requires process kill workaround	48 comments, 12 👍; "invalid" label contested by reporters
#49322	Opus 4.7 thinking summaries not rendered in VS Code	New model release (April 16) shipped with broken IDE integration	18 comments, 17 👍; affects premium model differentiation
#44657	Subagent Write tool blocks .md files with "report"/"summary"/etc. filenames	Arbitrary guardrail breaks legitimate documentation workflows; no override	4 comments, 8 👍; "no opt-out" explicitly called out
#46834	TUI relayouts duplicate entire transcript to scrollback	Terminal state corruption; destroys scrollback utility for long sessions	4 comments, 5 👍; has repro; SIGWINCH trigger identified
#51764	--continue cache invalidation persists on v2.1.116 — 41-99pp hit-rate delta	Reopened closed issue with rigorous methodology; cost impact quantified	1 comment; TTL-matched control experiment; youngjin39's repro is notable
#51760	Background bash children survive session close; 4.86 TB output runaway on Windows	Resource exhaustion risk; no output size caps on task files	1 comment; extreme severity (TB-scale); Windows-specific

---

4. Key PR Progress

#	PR	Description	Assessment
#41611	"Add the missing source to claude code"	Single-line PR with no description; likely spam or test	Low signal; no maintainer engagement
#51452	README.md rewrite — de-AI'd copy, fixed npm badge	Documentation quality improvement; removes "filler phrases, promotional language"	Community docs contribution; style-focused
#51451	SECURITY.md revision — HackerOne-only reporting, clear out-of-scope	Security process hardening; reduces noise in GitHub issues	Process improvement; likely maintainer-requested
#51395	"Claude/arrowhead gps logging pe arx"	Empty description; garbled title; probable spam/malformed	Ignore

Note: Only 4 PRs active in window; none from Anthropic maintainers. Community contributions limited to documentation.

---

5. Feature Request Trends

Direction	Evidence	Momentum
Theme/UI customization	#25366 (color themes, input highlighting)	14 comments, 24 👍; readability for long conversations
Hook/event system expansion	#51767 (SessionRenamed hook); prior art in hooks area	Ecosystem integration demand (tmux, external systems)
LaTeX terminal rendering	#44479	5 👍; academic/technical user segment
Plan mode model switching	#35920 (CLOSED without implementation)	Cost/quality optimization workflow

---

6. Developer Pain Points

🔴 Critical: Usage Limit Enforcement

• #16157, #51757, #37944 form a pattern: Max plan limits are erratically enforced, displayed, and reset. Users report timezone confusion, incorrect quota math, and 429s on paid tiers. Business risk for Anthropic's highest-ARR product.

🔴 Critical: Session/Cache Reliability

• #51764 (resume cache invalidation), #46834 (scrollback corruption), #51760 (orphaned processes) indicate state management fragility across platforms. The 41-99 percentage point cache miss delta on resume directly impacts API costs.

🟡 Model Quality Perception

• #42796 closure without transparent resolution fuels distrust. February "updates" flagged as degradation; Opus 4.7 launch (#49689, #49322) has integration gaps.

🟡 Platform Parity Gaps

• Windows: #42776, #51756 (/focus missing), #51760 (bash runaway)
• VS Code extension: #49322 (thinking summaries), #38006 (terminal spam)
• macOS dominates issue volume but Windows-specific bugs show underinvestment.

🟡 Guardrail Overreach

• #44657: Subagent filename blocking with no escape hatch. Pattern of unconfigurable safety rules breaking legitimate workflows.

---

Digest compiled from 50 issues, 4 PRs, and 1 release in the 24-hour window ending 2026-04-22.

OpenAI Codex — openai/codex

OpenAI Codex Community Digest — 2026-04-22

---

1. Today's Highlights

The Codex team is shipping aggressively on the Rust CLI with six rapid alpha releases (v0.123.0-alpha.2 through .7) in 24 hours, suggesting a critical fix or feature stabilization cycle. Meanwhile, a massive PermissionProfile migration is underway across 15+ PRs from bolinfest, systematically replacing legacy sandbox abstractions with a unified permission model—this will fundamentally reshape how Codex handles security and approvals. On the community front, token consumption anxiety remains the dominant theme with Issue #14593 now at 555 comments and 234 upvotes, making it the most active thread in Codex history.

---

2. Releases

Version	Notes
rust-v0.123.0-alpha.2 through rust-v0.123.0-alpha.7	Six rapid alpha iterations in 24h. No detailed changelogs provided in release notes—pattern suggests hotfix cycle or CI-driven build pipeline. No stable release.

---

3. Hot Issues

#	Issue	Why It Matters	Community Pulse
#14593	Burning tokens very fast — Business tier user reporting extreme token consumption	The definitive thread on Codex cost anxiety; affects pricing trust at scale	🔥 555 comments, 234 👍 — most active issue ever; users sharing mitigation strategies, demanding transparency
#10450	Remote Development in Codex Desktop App	Gap vs. VS Code remote-ssh/remote-WSL workflows; blocks enterprise adoption	580 👍, 148 comments; users pleading for parity with existing dev environments
#11678	Typing on CLI since 0.100 is very slow (Windows TUI regression)	Core UX degradation on Windows; affects daily productivity	17 comments, low 👍 count but persistent reports; regression tag indicates acknowledged bug
#13549	WSL-configured app still references Windows config.toml	Configuration leakage breaks WSL isolation expectations	19 👍, 16 comments; Windows/WSL users frustrated by config duality
#14339	Clear context before implementing plan (closed)	Feature parity request vs. Copilot/Claude Code's plan-then-implement flow	22 👍, 13 comments; closed status unclear if implemented or rejected
#18341	Blurred overlay on Intel Macs (macOS 15.0.1)	Rendering bug on specific hardware; affects visual usability	6 👍, 8 comments; Intel Mac users feeling neglected vs. Apple Silicon
#18755	SkyComputerUseClient crashes on macOS 14.x (built for 15.0)	Backward compatibility broken for Computer Use feature	15 👍, 8 comments; macOS 14 users blocked from skills functionality
#18463	Critical: CPU usage ~290% permanently	Performance regression making app unusable	2 👍 but 8 comments; severity label suggests internal priority
#15298	default.rules ignored for sandbox-escaping commands	Security policy inconsistency; execpolicy returns allow but rules still blocked	6 👍, 6 comments; power users hitting policy engine edge cases
#18881	MCP child processes leak (492 orphaned over 15h)	Resource exhaustion in long-running daemons; stability concern	Fresh (today), 3 comments; quantified leak makes it actionable

---

4. Key PR Progress

#	PR	Feature/Fix	Significance
#18745	Add sticky thread environment selections	Per-thread "sticky" environment (cwd + env vars) that persists across turns	Eliminates repetitive @ context setup; experimental but foundational for workflow UX
#18416	Add turn-scoped environment selections	Per-turn override of environment before sticky resolution	Granular control without polluting thread state; pairs with #18745
#18871	refactor: add agent identity crate	Isolated codex-agent-identity crate for assertion signing + task registration	Security architecture: clean separation of identity logic; enables auditability
#17820	AWS SigV4 auth for OpenAI-compatible providers	First-class Amazon Bedrock/Mantle support via AWS credential chain	Enterprise/cloud blocker; unlocks AWS GovCloud/regulated deployments
#18867	sandboxing: materialize cwd-relative permission globs	Anchors relative deny patterns to request cwd before session storage	Fixes permission re-interpretation bug; security correctness
#18279	app-server: accept permission profile overrides	API accepts PermissionProfile for thread/turn/command operations	Core migration milestone; rejects ambiguous legacy+profile requests
#18277	core: derive active permission profiles	Derives PermissionProfile from constrained SandboxPolicy + network settings	Prevents profile/state drift; single source of truth
#18276	exec-server: carry filesystem sandbox profiles	Preserves PermissionProfile alongside derived sandbox inputs	Enables full-disk vs. restricted profile distinction at runtime
#18897	Add sticky environment API and thread state	API + persistence layer for sticky environments; stack base for #18745	Infrastructure for environment stickiness; test coverage included
#18909	Update /statusline and /title snapshots	TUI snapshot maintenance	Housekeeping; indicates active CLI polish

---

5. Feature Request Trends

Trend	Evidence	Direction
Remote/SSH/WSL parity	#10450 (580 👍), #13549, #16994, #18297	Codex Desktop must match VS Code's remote development story; file navigation in @ search needs folder support
Permission granularity & automation control	#18247, #16695, #15298, #14339	Users want per-automation sandbox overrides, easier rules config, and plan-then-implement workflows with context control
Visibility of hidden/dot files	#18299 (5 👍)	.agents/, .codex config directories invisible in file viewer; power users need access
Azure/Enterprise auth	#1056 (closed)	Entra/SSO authentication for corporate endpoints; now partially addressed by AWS PR #17820 but Azure still gap

---

6. Developer Pain Points

Pain Point	Frequency	Impact	Representative Issues
Token cost unpredictability	🔴 Chronic	Financial/trust erosion	#14593 (555 comments), #17257 (memory leak → cost)
Windows as second-class citizen	🔴 High	Daily friction for majority platform	#11678 (TUI slowness), #13549 (WSL config), #18861 (sandbox EPERM), #16994 (automations fail)
macOS version/hardware fragmentation	🟡 Moderate	Features broken on older versions	#18341 (Intel Mac rendering), #18755 (macOS 14 crash), #18507 (microphone entitlement)
MCP reliability & resource leaks	🟡 Growing	Daemon stability, tool availability	#18881 (process leak), #15508 (tools disappear), #18233 (agent confusion post-upgrade)
Permission system complexity	🟡 Active migration pain	Confusion during legacy→profile transition	#15298 (rules ignored), #18242 (invalid sandbox policy), #11267 (Ctrl+C deadlock in /review)
App updater fragility	🟢 Niche but severe	Stuck updates, manual intervention	#17413 (macOS updater hangs)

---

Digest compiled from github.com/openai/codex activity 2026-04-21/22. For real-time updates, monitor the OpenAI Codex repository.

Gemini CLI — google-gemini/gemini-cli

Gemini CLI Community Digest — 2026-04-22

Today's Highlights

Google shipped v0.39.0-preview.1 as a patch release, addressing nested plan directory duplication and relative path policy bugs. The community is actively pushing cross-platform improvements, with notable PRs for Windows shell interoperability and async boot optimization. Meanwhile, long-running stability issues—particularly PTY exhaustion on macOS and excessive project-wide scanning—continue to generate significant discussion.

---

Releases

v0.39.0-preview.1 — Patch release cherry-picking PR #25138 to fix nested plan directory duplication and relative path policy enforcement. Resolves scenarios where plan files in nested directories caused path resolution failures and agents failed to respect custom security policies by emitting absolute paths instead of relative ones.

---

Hot Issues

#	Issue	Why It Matters	Community Signal
#25672	Excessive project-wide scanning makes simple code patches fail	Performance regression: even trivial edits trigger full workspace + Go stdlib + cgo + vulnerability scans, destroying iteration speed.	6 comments, active triage needed
#25583	PTY Master Device Exhaustion (ENXIO) on macOS after prolonged YOLO usage	System-level failure: weeks of --approval-mode yolo exhausts macOS PTY limit (kern.tty.ptmx_max = 511), breaking all terminal sessions system-wide.	5 comments, maintainer-flagged
#22745	Assess impact of AST-aware file reads, search, and mapping	EPIC for precision tooling: could reduce token waste and misaligned reads by using AST boundaries instead of line-based heuristics.	5 comments, 1 👍, linked to #22746
#22323	Subagent recovery after MAX_TURNS reported as GOAL success	Critical correctness bug: codebase_investigator claims success despite hitting turn limits, masking analysis failures.	3 comments, 2 👍, P1 priority
#24916	Repeated permission prompts on same file	Trust erosion: "allow for all future sessions" intermittently ignored, breaking flow state.	3 comments, security-tagged
#25166	Shell execution hangs with "Waiting input" after completion	Reliability issue: simple commands falsely report awaiting input, stalling agent progress.	2 comments, 3 👍
#23571	Model creates tmp scripts in random spots	Workspace hygiene: script proliferation complicates commit preparation and cleanup.	2 comments, agent-area
#22267	Browser Agent ignores settings.json overrides	Configuration system gap: maxTurns and other overrides silently discarded despite correct registry initialization.	2 comments, P2 priority
#25216	Failed to open in temporary path A:\	Windows-specific crash: EISDIR on realpath('A:\a') suggests drive-root handling bug.	1 comment, core-area
#22819	Implement memory routing: global vs. project	Foundation for personalization: separates user preferences (~/.gemini/) from codebase-specific context (.gemini/).	1 comment, 2 👍

---

Key PR Progress

#	PR	Description	Status
#25712	Shell tool RC file support + PAGER=cat fix	Adds shellToolRcFile setting; prevents more/less hangs by explicitly setting PAGER=cat when no RC file used. Revisits #5953.	Open
#25775	/note command for workspace notes	New slash command: /note <text> appends timestamped entries to notes.md; /note view displays them. Lightweight knowledge capture without polluting chat history.	Open
#25758	Async experiments/quota fetching for faster boot	Eliminates synchronous awaits during refreshAuth; moves experimentsPromise and quotaPromise to async path. Targets consistently high startup latency.	Open
#25426	Revive bundling + 16-core CI speedup	Artifact-centric CI with .github/actions/setup-gemini; shards consume pre-built bundles. Modernizes test infrastructure for parallel execution.	Open
#25769	**Windows shell interoperability: &&, `		, /dev/null`**
#25770	Deep merge for A2A settings	Fixes #25747: shallow spread was overwriting nested objects (fileFiltering, tools). Immutable deep merge preserves workspace-level overrides.	Open
#25720	Enhanced shell validation + core tools allowlist	Recursive policy checking for sub-commands, substitutions, subshells. New tools.core setting for high-precision tool allowlisting.	Open
#25765	1:1 part count in tool responses	Fixes 400 errors from part count mismatches in multimodal tool responses (file/inline data). Strict mapping between tool calls and function responses.	Open
#24270	Experimental /btw side inquiry feature	Ephemeral, non-persistent chat overlay for quick questions without modifying main history or triggering tools. "By the way" workflow.	Open
#25409	Wire up ContextManager and AgentChatHistory	Implements #25408: new context management architecture for agent conversation state.	Open

---

Feature Request Trends

1. AST-aware tooling precision — Multiple issues (#22745, #22746) push for semantic code understanding over line-based heuristics to reduce token waste and misaligned reads.

2. Structured memory and personalization — #22819 and #22809 drive toward global vs. project-scoped memory, with proactive agent behavior to capture preferences.

3. Subagent robustness and observability — #22323, #23582, #24037 converge on better lifecycle management, approval mode awareness, and mid-execution tracker updates.

4. Cross-platform shell parity — #25769, #25216, #24202 reflect sustained demand for first-class Windows and SSH session support.

---

Developer Pain Points

Category	Symptoms	Frequency
Performance / startup latency	Full-project scans for trivial edits (#25672), synchronous boot blocking (#25758)	Very high — multiple top issues
Resource exhaustion (macOS)	PTY depletion in long-running sessions (#25583)	Moderate, severe impact when hit
Permission fatigue	Repeated prompts despite "allow forever" selection (#24916)	Moderate
Shell execution reliability	Hangs on completed commands (#25166), Windows operator incompatibility (#25769)	Moderate
Configuration system gaps	Settings ignored (Browser Agent #22267), shallow merge bugs (#25770)	Moderate
Workspace hygiene	Random temp script placement (#23571)	Low-moderate

---

Digest compiled from google-gemini/gemini-cli public activity. For full details, visit the repository.

GitHub Copilot CLI — github/copilot-cli

GitHub Copilot CLI Community Digest — 2026-04-22

1. Today's Highlights

GitHub shipped three rapid-fire releases (v1.0.35-0 through v1.0.35-2) with significant automation improvements: HTTP webhook support for hooks, multi-path grep/glob tools, and a new continueOnAutoMode config that automatically falls back to auto model selection on rate limits rather than pausing execution. Meanwhile, the community is actively debating model access restrictions, with multiple reports of Opus models disappearing from Copilot Pro subscriptions and "model not supported" errors proliferating.

---

2. Releases

v1.0.35-2 | v1.0.35-1 | v1.0.35-0

Released: 2026-04-21

Version	Key Changes
v1.0.35-2	• continueOnAutoMode config option: auto-switches to auto model on rate limit instead of pausing execution • Fixed: Auto mode no longer crashes when switching to models lacking configured reasoning effort support • Fixed: Pattern-specific instruction files in .github/instructions/ now work correctly
v1.0.35-1	Pre-release build
v1.0.35-0	• HTTP hook support: Hooks can now POST JSON payloads to configured URLs instead of requiring local command execution • Improved: grep and glob tools accept multiple search paths • Fixed: Subagent thinking hidden from timeline; custom agent names now visible in status line

Impact: The HTTP hooks feature is particularly significant for CI/CD integrations and remote telemetry, while continueOnAutoMode addresses a major workflow disruption for power users hitting rate limits.

---

3. Hot Issues

#	Issue	Status	Comments	Why It Matters
#222	ACP (Agent Client Protocol) support	CLOSED	25 👍146	The "LSP for AI agents" standard—critical for IDE/editor ecosystem integration. High engagement reflects strong demand for interoperability.
#1161	Invalid session ID with Opus 4.5	CLOSED	23 👍15	Blocking bug for macOS/Fish users; author explicitly switched to competitor (opencode.ai). Reputation risk.
#223	"Copilot Requests" permission missing for org-owned tokens	OPEN	21 👍65	Enterprise governance gap—orgs need service accounts, not personal PATs. Security/compliance blocker for corporate adoption.
#1276	Paste images from clipboard into CLI prompts	OPEN	10 👍7	UX gap for visual debugging workflows (screenshots, UI bugs, logs). CLI lags behind GUI tools here.
#2078	Add /btw command	CLOSED	7 👍26	Community wants parity with other CLI tools for quick asides/context switching.
#2661	Opus 4.5 "model not supported" error	OPEN	6 👍0	Model availability regression—student pack users losing access. Pattern of model entitlement confusion.
#2625	Poor rendering performance on long conversations	OPEN	4 👍3	Severe degradation (30-45s hangs) forcing conversation resets. Fundamental scalability issue.
#2873	Copilot Pro lost Opus model access	OPEN	4 👍0	Subscription value crisis: User explicitly objects to removal after prior access. "I do not mind the request multiplier, but to stop me using it altogether."
#334	Shell completions	OPEN	4 👍11	Basic CLI hygiene request (zsh/bash/fish). Surprisingly unimplemented for a developer tool.
#2867	Claude Opus 4.6 (high) "model not supported" after quota reset	OPEN	4 👍1	Quota system UX failure: user followed instructions, waited, then hit worse error. Trust erosion in model access promises.

---

4. Key PR Progress

#	PR	Status	Description
#2887	Safe cleanup of older direct-install binaries	OPEN	Addresses accumulated binary bloat from repeated install.sh runs; automatic cleanup with opt-out. Improves disk hygiene and reduces PATH confusion.
#2800	Initial devcontainer configuration	OPEN	Enables reproducible development environments; lowers contribution barrier for external contributors.

Note: Only 2 PRs updated in tracking window.

---

5. Feature Request Trends

Trend	Evidence	Momentum
Model access transparency & control	#2873, #2867, #2661, #2878, #1190, #2792	🔥 Critical
Users demand clarity on which models are available when, with what quotas, and why access disappears. Planning/execution model splitting (#2792) emerging as efficiency optimization.
Enterprise/Org governance	#223, #2711, #2753	⚠️ Growing
Fine-grained permissions, Azure DevOps parity, and plugin skill visibility for org deployments. Corporate readiness gaps.
Context & memory management	#2720, #2885, #2625	📈 Rising
Speculative context (checkpoint/rewind), non-interactive /ask behavior, and rendering performance at scale. Conversation state is becoming a bottleneck.
Visual/multimodal input	#1276	📌 Steady
Clipboard image pasting for screenshots, diagrams, UI bugs. CLI needs sensory expansion beyond text.
Customization & theming	#2830, #2884, #334	📌 Steady
Color themes, diff formatting, shell completions—polish requests indicating maturing user base.

---

6. Developer Pain Points

Pain Point	Frequency	Severity	Representative Issues
Model availability chaos	🔴 High	Critical	#2661, #2873, #2867, #2878, #1161
Opus models disappearing, "not supported" errors post-quota reset, student pack entitlement confusion. Users report switching to competitors. Suggests backend policy changes not communicated to CLI team or users.
Performance degradation at scale	🟡 Medium-High	High	#2625, #2884
Terminal hangs (30-45s) on long conversations; diff rendering noise. Forces destructive workarounds (clearing history).
Non-interactive mode limitations	🟡 Medium	Medium	#2882, #2885
MCP sampling blocked in -p mode despite --allow-all; /ask prompts for follow-up in side sessions. Automation use cases underserved.
Plugin/extension reliability	🟡 Medium	Medium	#2639, #2753
EPIPE/STREAM_DESTROYED on shutdown; marketplace skills invisible to agent system prompts. Extension ecosystem fragility.
Enterprise identity/permissions gaps	🟡 Medium	Medium-High	#223, #2711
Org token permissions, Azure DevOps command parity. Blocking corporate standardization.

---

Digest compiled from github.com/github/copilot-cli activity 2026-04-21.

Kimi Code CLI — MoonshotAI/kimi-cli

Kimi Code CLI Community Digest — 2026-04-22

Today's Highlights

The past 24 hours saw intense community activity with 25 issues and 13 PRs updated, dominated by authentication edge cases, OAuth subagent failures, and terminal stability problems. A critical fix for terminal hangs and MCP connection leaks landed alongside parallel PRs addressing Anthropic API spec violations and hook system extensibility. Windows and WSL2 users continue reporting platform-specific friction, while model preference debates (K2.5 vs K2.6) signal growing user sophistication.

---

Releases

No new releases in the last 24 hours.

---

Hot Issues

#	Issue	Why It Matters	Community Signal
#1984	Terminal hang on exit + MCP connection leak	Production reliability: Extended sessions become unrecoverable, requiring force-kill; orphaned MCP connections trigger firewall alerts	Immediate PR response (#1985) shows maintainer prioritization
#1983	Subagents fail with HTTP 401 under OAuth	Architecture gap: OAuth auth doesn't propagate to subagents, breaking multi-agent workflows for enterprise users	Zero comments suggests under-reported impact; likely affects advanced users
#1975	Anthropic provider splits parallel tool_results	Spec compliance: Violates Anthropic Messages API, causing provider-side rejections; blocks kosong users	Author provided commit-level repro; PR #1978 already open
#1974	Undo command jumps 4 turns back	UX precision: Destroys conversation state unexpectedly; "undo" is high-trust operation	No workarounds mentioned; affects WSL2 workflow
#1973	Phantom "empty message" insertions	Noise pollution: System inserts fake user messages during thinking, corrupting context	High-frequency occurrence ("frequently") on macOS + kimi-for-coding
#1969	WinError 17 on Windows install	Platform parity: os.replace() atomic write fails on Windows Python; blocks fresh installs	Common pattern; needs cross-platform file operations
#1961	/login errors on Windows for kimi-code	Onboarding blocker: New users can't authenticate on primary platform	3 comments in 24h; likely widespread
#1947	OAI-compatible copilot failures	Ecosystem compatibility: Johnny-zhao's popular copilot integration broken	VS Code context suggests IDE extension overlap
#1925	K2.5 vs K2.6 model preference	Model governance: Users reject forced K2.6 upgrade; "thinking drowns creativity"	6 comments, emotional language ("lost all personality"); signals brand risk
#640	Infinite file read loops	Resource exhaustion: Custom Anthropic endpoints trigger repetitive file scanning	Persistent since January; 5 comments show ongoing pain

---

Key PR Progress

#	PR	What It Does	Status
#1985	Fix TTY hang + MCP connection leak	Non-blocking os.read() with os.set_blocking(fd, False); proper cleanup in shutdown hooks	Open, pairs with #1984
#1978	Merge parallel tool_results for Anthropic	Spec-compliant merging of consecutive tool_result blocks into single user message	Open, resolves #1975
#1972	Visual context progress bar	Unicode block progress bar with color coding, matching Claude HUD aesthetic	Open, pure UX enhancement
#1979	Approval cancellation feedback	Distinguishes timeout vs. explicit rejection in ApprovalResult	Open, partial fix for #1823
#1960	RalphFlow architecture	Ephemeral context + convergence detection to prevent infinite agent loops	Open, significant architectural contribution
#1963 + #1935	updatedInput for PreToolUse hooks	Enables transparent command rewriting; powers rtk proxy integration (60-90% token reduction)	Both open, overlapping scope—potential merge conflict
#1957	Fix list_sessions without cwd	Restores ACP protocol compliance for Zed editor integration	Open, editor ecosystem critical path
#1832	Fix UserPromptSubmit hook empty prompt	Corrects list[ContentPart] handling so hooks capture actual user input	Open, 3rd-party extensibility fix
#1784	AWS Bedrock Mantle platform support	Adds openai_legacy provider with dynamic region URLs	Open, enterprise cloud expansion
#1600	Bright blue user input highlighting	Visual separator + #007AFF coloring for message distinction	Open, long-running UI polish

---

Feature Request Trends

Direction	Evidence	Momentum
Model selection granularity	#1925 (K2.5 fallback), implicit in auth issues	High—users want control beyond "latest"
Voice input / hands-free mode	#1934 explicit Claude Code parity request	Emerging—accessibility + mobility use cases
Custom theming	#1981, #1982 (closed as dupe)	Low but persistent—terminal aesthetic integration
Remote control / headless operation	#1964 ("提效10倍")	Niche but high-impact for automation
Vendor directory indexing	#1962	Go/Rust ecosystem specific—@ symbol expansion

---

Developer Pain Points

Category	Symptoms	Root Indicators
Authentication fragility	401 errors in TUN mode (#1971), OAuth subagent propagation (#1983), payment gate confusion (#1976)	Token lifecycle management incomplete across agent boundaries; error messages don't distinguish auth vs. billing
Windows/WSL2 second-class citizenship	WinError 17 (#1969), /login platform rejection (#1961), tmux rendering (#1965), undo over-correction (#1974)	Cross-platform testing gaps; POSIX assumptions in file ops and terminal control
Context/undo precision	4-turn undo jumps (#1974), phantom messages (#1973), tool result loss (#1980)	State machine around turns vs. tool calls likely has off-by-N logic
MCP/ACP lifecycle	Connection leaks (#1984), session listing failures (#1957), JetBrains ACP init (#1967)	Protocol implementations lack graceful degradation; editor integrations under-tested
Hook system maturity	Empty prompts (#1832), missing updatedInput (#1935, #1963)	Beta hooks API has edge cases in type handling; community actively patching gaps

OpenCode — anomalyco/opencode

OpenCode Community Digest — 2026-04-22

Today's Highlights

OpenCode shipped v1.14.20 with critical fixes for TUI system theme regression and Windows dynamic imports, while the community actively debated Copilot integration reliability and plugin extensibility gaps. A major SDK v2 proposal is under review to modernize the plugin architecture, and several stacked PRs from core contributor kitlangton signal deep investment in Effect Schema migration for the message pipeline.

---

Releases

v1.14.20 — Three targeted fixes:

• TUI: Restored system theme transparency (terminal colors showing through), resolving a regression introduced in v1.14.19 that forced white backgrounds on dark terminals
• HTTP API: Added GET /config to the experimental API surface
• Windows/Node: Fixed local dynamic imports for plugins and tools, addressing a platform-specific loading failure when running under Node.js rather than Bun

---

Hot Issues

#	Issue	Why It Matters	Community Reaction
#13768	Copilot + Opus 4.6 prefill rejection	Blocks GitHub Copilot users from using Claude Opus 4.6; "assistant message prefill" error breaks conversation flow	60 comments, 23 👍 — highest engagement item; users sharing session exports and workarounds
#13984	Copy-paste broken in CLI	Fundamental UX regression; clipboard integration fails silently despite "copied" confirmation	30 comments, 10 👍; cross-platform reports (macOS/Windows/Linux)
#22408	Kimi K2.6 integration (closed)	Moonshot's new model family requested for code plans; closed likely due to v1.14.x native support	18 comments, 18 👍; strong demand for Chinese model ecosystem
#14194	SQLite corruption with shared Docker/local config	Data loss risk for containerized workflows; race condition on concurrent database access	16 comments, 8 👍; DevOps/CI users particularly affected
#14462	Roslyn LSP as C# alternative (closed)	Microsoft's official C# LSP vs. community csharp-ls; quality and maintenance implications	13 comments, 8 👍; .NET developers pushing for first-party tooling
#17516	opencode run hangs after tool completion	Breaks headless/automation workflows; process zombie requires manual kill	13 comments, 6 👍; CI/CD blocking issue
#23599	System theme regression (closed)	Terminal color passthrough broken; fixed in v1.14.20 per release notes	8 comments, 13 👍; rapid fix appreciated
#22100	Unexpected pip3 execution	Security concern: OpenCode triggering package installs on read-only configs; privilege escalation questions	6 comments; trust and sandboxing debate
#23658	White terminal background post-1.14.19	Visual regression forcing dark-theme users to downgrade; related to #23599	5 comments, 7 👍; Ghostty/macOS specific
#22292	Managed settings bypass via env vars	Enterprise governance failure; OPENCODE_PERMISSION overrides admin-locked configs	4 comments; security/compliance critical

---

Key PR Progress

#	PR	Status	Description
#12042	Plugin SDK v2	Open	Dual-client architecture (v1/v2) for incremental plugin migration without breaking changes; addresses long-term API evolution
#23766	Disable toolStreaming for Anthropic providers	Open	Fixes proxy/gateway failures from eager_input_streaming beta header injection in @​ai-sdk/anthropic ≥3.0.58
#18306	Open WebUI provider	Open	Self-hosted WebUI integration with OpenCode-native features; builds on prior community attempt
#22612	Copilot chat tool call IDs	Open	Resolves Kimi/NVIDIA tool call routing through generic OpenAI-compatible runtime
#18201	Session todo API endpoint	Open	POST /session/:sessionID/todo for programmatic TUI sidebar manipulation by plugins
#21343	Scroll snap fix during streaming	Open	Eliminates jarring auto-scroll when reading history during active LLM response
#23760	Interactive terminal tool with PTY	Open (needs:compliance)	Persistent pseudo-terminal sessions for agent-managed interactive shells; significant capability expansion
#23612	Roslyn LSP sync range + workspace symbol	Open	Fixes Roslyn server crashes and symbol query failures; directly implements #14462
#23755	Preserve Anthropic thinking blocks	Open	Fixes thinking/redacted_thinking block errors in message transforms
#23763-23765	Effect Schema migration stack	Open (3 stacked)	Core infrastructure: Cursor, MessageV2 errors, BusEvent/SyncEvent schemas migrating from Zod to Effect Schema

---

Feature Request Trends

1. Model ecosystem expansion: Kimi K2.6, Mistral reasoning variants, xAI/Grok, Azure OpenAI — demand for broader provider coverage accelerating
2. Voice/dictation input: Long-standing gap (#4695, #9264, #11345) now blocked by plugin API limitations; #17425 identifies specific extensibility barriers
3. Enterprise packaging: Flatpak (#5651, now closed with fix), corporate npm registry auth (#23434), managed settings enforcement (#22292)
4. Mobile/touch optimization: #18767 in progress for app usability on tablets/phones
5. LSP diversity: Roslyn for C#, broader language server alternatives — quality over community-maintained defaults

---

Developer Pain Points

Theme	Evidence	Severity
TUI stability & theming	#23599, #23658, #23073, #16796 — crashes, visual regressions, interaction failures	High — daily workflow impact
Clipboard/terminal integration	#13984, #17792 (PATH issues) — basic UX expectations unmet	High — fundamental functionality
Headless/automation reliability	#17516 (hangs), #14194 (corruption) — opencode run and

---

⚠️ 内容超过 GitHub Issue 上限，完整报告见提交的 Markdown 文件。