Twitter AI Coding - 2026-07-14¶
1. What People Are Talking About¶
1.1 Agent work is gaining live handoffs and in-flight checks (🡒)¶
The most concrete releases concerned supervising work that is already running, rather than merely starting another chat. GitHub connected remote Copilot CLI sessions to Mobile notifications and added an in-progress security review to the Copilot app. Antigravity's team command remained highly visible, but its replies challenged the announcement's novelty.
@github announced (66 likes, 4 replies, 30,371 views) live GitHub Mobile notifications for remote Copilot CLI sessions, including progress, a prompt when input is required, and access back into session logs. The image shows the intended intervention point: a phone notification for a remote “Fix authentication bug” task marked “Needs input.”
@GHchangelog released (12 likes, 1 reply, 2,079 views) /security-review in the Copilot app's public preview. GitHub says the command reviews current changes for high-confidence findings with severity and confidence, covering classes such as injection, XSS, insecure data handling, path traversal, and weak cryptography, while complementing existing code and secret scanning.
@testingcatalog highlighted (274 likes, 13 replies, 23,844 views) Antigravity's /teamwork-preview, whose quoted official post describes parallel planning, building, and verification. The screenshot confirms a first-class team command, but replies from people saying they had used it for months mean the evidence supports current visibility, not a confidently new capability.
Discussion insight: The useful tension is not opposition to agents; it is where people can observe and intervene. Antigravity replies questioned the release framing, while GitHub's Mobile UI makes “needs input” an explicit control state.
Comparison to prior day: The prior day's focus on operating agent teams continued, but the evidence shifted toward user-facing supervision surfaces: phone notifications and an in-editor security check.
1.2 Teams are composing models into roles instead of picking one winner (🡕)¶
Practitioners described model choice as a task-routing problem. One first-hand account favors a smaller model for basic and medium work; another assigns planning, implementation, and review to different agents inside a custom extension.
@burkeholland reported (57 likes, 10 replies, 3,715 views) doing substantial work, including UI work, with GPT-5.6 Luna in GitHub Copilot after expecting to have moved beyond smaller models. The attached Copilot screen shows a theme-change task moving from plan and file search to changed-file counts and a rendered preview; the author later limited the praise to basic and medium tasks. A reply added that higher reasoning effort can erase the speed and cost advantage, while another raised the value of accumulated transcript and project context.
@davis7 described (23 likes, 4 replies, 554 views) a custom Pi extension in which a Claude Code subagent researches API design, a Codex subagent implements, and a Claude Code subagent reviews for simplicity and correctness. The author calls the dynamic-workflows layer unpolished and says a desktop app may be needed, which usefully distinguishes a working personal workflow from a finished product.
Discussion insight: The Luna thread supports a narrower conclusion than “small models replace frontier models”: people value them for routine work, but context continuity, reasoning settings, and price still govern the choice.
Comparison to prior day: Yesterday's discussion emphasized parallel orchestration; today's strongest new detail is role specialization across different model-backed agents.
1.3 Product convergence is creating a navigation and continuity problem (🡕)¶
The day also surfaced a product-design cost of folding chat, coding, and autonomous work into one surface. This was not an abstract brand complaint: replies named different behavior across mobile and desktop and projects or histories that do not move cleanly between modes.
@emollick contrasted (99 likes, 18 replies, 7,321 views) Claude's Home/Code and Chat/Cowork choices with OpenAI's ChatGPT Work/Codex split, noting that the websites differ again. In the replies, @simonw said mobile Chat and Work differences are material while desktop Codex versus Work is cosmetic; another reply said Projects created in one mode do not appear in another. A Claude Cowork web-beta user supplied a counterexample, saying its web and desktop UIs match.
@signulll argued (80 likes, 14 replies, 8,020 views) that folding Codex/“ChatGPT Work” into ChatGPT is a smart distribution move. The replies split: one reader defended the merge as necessary, while another said it loses chat history and makes chat-only use worse.
Discussion insight: Distribution and usability are pulling in opposite directions. The replies support demand for handoff from chat into work, but not for mode boundaries that hide history or assets.
2. What Frustrates People¶
Fragmented work modes and lost continuity¶
Severity: Medium. The main complaint is not that chat and agent work coexist; it is that users cannot predict where a capability or its history belongs. @emollick documented (99 likes, 18 replies, 7,321 views) conflicting Home, Code, Chat, Cowork, Work, and Codex labels across products and surfaces, with replies describing material mobile differences and projects separated by mode. @signulll framed (80 likes, 14 replies, 8,020 views) the Codex/ChatGPT integration as a distribution play, but a reply reported losing chat history and finding chat-only use worse. The coping behavior visible in the discussion is to stay within the surface that preserves the task context; there is no evidence of a complete cross-mode solution. This is worth building for because it affects task discovery and continuity, not only preference.
Powerful browser context comes with an explicit trust boundary¶
Severity: High. @PablitoPool shared (689 views, 3 bookmarks) an OpenCode bridge that can use a real Chrome profile's tabs, sessions, history, bookmarks, screenshots, page text, and Chrome DevTools Protocol. Its README uses a local native-messaging host and a rotating bearer token, but expressly warns users to run it only in trusted local workflows because it reaches the real profile rather than an isolated browser. The practical gap is safe, inspectable delegation over authenticated browser context. This is worth building for because the convenience and the exposure are inseparable in the current design.
Usage limits remain hard to reason about while model routing changes¶
Severity: Medium. A reply under @burkeholland's Luna report (57 likes, 10 replies, 3,715 views) described exhausting resets under another model and then receiving three more, while another respondent argued that higher reasoning effort makes Luna slower and not much cheaper. @jumperz said (12 likes, 5 replies, 695 views) Codex limits had been reset repeatedly since GPT-5.6 launched, which is a single-user observation rather than a verified product policy. The visible workaround is manual monitoring of resets and shifting model or reasoning settings. This is worth building for as transparent, cross-model budget and reset visibility.
3. What People Wish Existed¶
One continuous workspace from chat through autonomous work¶
The threads did not contain a formal feature request, but they show a practical need: start in a conversation, hand off to work, supervise the run, and return without losing history or assets. @emollick surfaced (99 likes, 18 replies, 7,321 views) the confusing overlap among Chat, Work, Codex, Home, Code, and Cowork, while a reply reported Projects split by mode. @github added (66 likes, 4 replies, 30,371 views) mobile intervention for a remote Copilot CLI session, showing one piece of that lifecycle. The need is practical and urgent for users operating long tasks; partial answers exist, but not a single continuity model across surfaces. Opportunity: Direct.
Reusable agent capabilities with scoped permissions¶
The strongest supply-side response is reusable skills, plugins, and bridges, but their permission boundaries are still tool-specific. @koruneko32767 pointed to (215 views, 2 bookmarks) Microsoft's Power Platform skills marketplace, whose plugins cover several development surfaces and document safer allow-list approaches alongside broad auto-approval. @PablitoPool shared (689 views, 3 bookmarks) a browser bridge that preserves session context but warns it should be trusted-local only. The need is for portable capabilities that advertise the files, services, and browser state they can reach before they run. Opportunity: Competitive.
Clear budget and reasoning-mode controls¶
The Luna discussion showed that a cheaper, smaller model can be useful, but that the value changes with reasoning effort, project context, and reset availability. @burkeholland reported (57 likes, 10 replies, 3,715 views) positive results on UI and routine work, and replies added both a cost caveat and a context-continuity caveat. A response to the thread also described unexpected extra resets. The practical need is a pre-task view of model choice, reasoning setting, expected budget, and remaining allowance; the available evidence shows manual observation rather than such a unified control. Opportunity: Direct.
4. Tools and Methods in Use¶
| Tool | Category | Sentiment | Strengths | Limitations |
|---|---|---|---|---|
| Copilot CLI + GitHub Mobile | Agent CLI and mobile supervision | (+) | Live phone notifications, progress, input-needed alerts, and session-log return path | The release addresses supervision, not full cross-surface task continuity |
Copilot app /security-review |
In-editor security review | (+) | High-confidence findings, severity/confidence, actionable fixes for current changes | Public preview; focused on security review rather than broad code quality |
| GPT-5.6 Luna in GitHub Copilot | Coding model | (+/-) | First-hand report of fast basic/medium and UI work | Higher reasoning effort can narrow its speed/cost advantage; existing context can affect a switch |
| Pi extension | Multi-agent orchestration | (+/-) | Routes research, implementation, and review across Pi, Claude Code, and Codex | Personal extension; author says workflow polish is incomplete and TUI limits remain |
| Power Platform skills | Plugin marketplace | (+) | Official reusable skills, agents, and commands across Power Platform workloads | Permission configuration remains consequential; broad auto-approval grants machine-level access |
| OpenCode Chrome Bridge | Browser-control bridge | (+/-) | Real-profile tabs, sessions, history, screenshots, and CDP access via an extension and local bridge | Reaches authenticated browser context; README limits safe use to trusted local workflows |
| Supabase + Vercel + Claude Code | Production app workflow | (+) | Preview PRs, preview databases, shared data, and a repeatable stack for non-engineer builders | Requires an engineered workflow to keep rapid building consistent and safe |
The satisfaction spectrum runs from positive evidence for more observable control surfaces to caution around permissions, cost, and UI fragmentation. The migration pattern is not a clean move from one vendor to another: @davis7 uses (23 likes, 4 replies, 554 views) different agents for research, implementation, and review, while @burkeholland uses (57 likes, 10 replies, 3,715 views) a smaller model for work that had previously implied a larger-model choice. Common workarounds are role routing, manual reset tracking, and keeping work inside the interface that retains its context.
5. What People Are Building¶
| Project | Who built it | What it does | Problem it solves | Stack | Stage | Links |
|---|---|---|---|---|---|---|
| Expert AI Developer Skills | @jggomezt | Reusable skills, behavior rules, SDLC workflows, sidecars, and a backend plugin for coding agents | Makes quality gates, test rules, branch safety, and security checks reusable instead of re-prompted | Markdown skills/rules/workflows; Python backend; MCP integrations | Shipped | Repository |
| Power Platform Skills | Microsoft | Plugin marketplace for agents building Power Platform apps and automations | Gives Copilot and Claude Code task-specific skills and commands for supported enterprise workloads | React, TypeScript, Vite, Expo, Node.js, PAC CLI, MCP | Shipped | Repository |
| OpenCode Chrome Bridge | @PablitoPool | Lets OpenCode control a real Chrome profile through a local bridge | Preserves authenticated browser context instead of recreating it in a headless session | Chrome MV3 extension, Node.js native host, local HTTP bridge, OpenCode plugin, CDP | Shipped | Repository |
| Pi multi-agent extension | @davis7 | A Pi-led system that invokes Claude Code and Codex subagents for different roles | Coordinates research, implementation, and review inside one workflow | Pi, Claude Code, Codex, Effect v4 | Alpha | Author |
Copilot app /security-review |
GitHub | Reviews uncommitted/in-progress changes for high-confidence security findings | Moves a focused vulnerability check earlier in the coding loop | GitHub Copilot app | Beta | Release notes |
The projects share a build pattern: package an operational capability around an existing agent rather than replace the agent itself. @jggomezt presented (2 likes, 2 replies, 75 views) skills, rules, workflows, and sidecars as a suite, while @koruneko32767 shared (215 views, 2 bookmarks) an official marketplace that maps the same idea to Power Platform tasks. The Chrome bridge is a different but related response: it supplies the stateful external context agents otherwise lack, and its own warning makes trust boundaries part of the product.
The production reference is delight.ai's workflow: its case study says 20–25 Supabase projects support more than 200 apps through Claude Code pull requests, Vercel previews, Supabase preview databases, and a shared Prisma monorepo. @supabase reported (6 likes, 2 replies, 804 views) that the setup lets teams at the 220-person company ship without pulling engineers off the core product; the case study also says an AI engineer assembled the safety and consistency layer.
6. New and Notable¶
Security review inside the active coding surface¶
@GHchangelog announced (12 likes, 1 reply, 2,079 views) that /security-review is in public preview for the GitHub Copilot app. Its release notes specify high-confidence findings with severity and confidence, actionable fixes, and focused checking of current changes. This is notable because it turns a security-focused agent capability into an on-demand check before code lands.
Mobile intervention for remote CLI work¶
@github announced (66 likes, 4 replies, 30,371 views) that GitHub Mobile can notify a user about remote Copilot CLI progress and required input, then return them to logs. The image’s “Needs input” notification makes the product claim concrete: asynchronous work still has deliberate points where human attention is required.
7. Where the Opportunities Are¶
[+++] Cross-surface task continuity — The combined evidence shows chat/work mode confusion, assets that may not cross modes, and a newly available phone-level intervention point. A product that preserves task history, artifacts, permissions, and approvals across chat, desktop, remote session, and mobile would meet a direct operational need.
[++] Permissioned context bridges — OpenCode Chrome Bridge demonstrates demand for using live browser sessions, while its own trusted-local warning shows why authenticated state needs inspectable scopes, visible access logs, and revocation. Reusable skills marketplaces reinforce the same need for explicit tool permissions.
[++] Model-routing and budget telemetry — The Luna thread and Pi role-routing workflow show that teams select models by task role, reasoning level, existing context, speed, and cost, not a single benchmark. A control layer that estimates and enforces those tradeoffs before an agent starts would address the reset and visibility concerns.
[+] Production guardrails for non-engineer builders — delight.ai's case study couples rapid AI-assisted building to PR previews, preview databases, shared data, and engineering-owned workflow design. The opportunity is to package those guardrails for teams that want more builders without uncontrolled production access.
8. Takeaways¶
- The visible frontier is supervising agent work, not merely launching it. GitHub Mobile now signals remote CLI progress and blocked input, while Copilot's app has an in-flight security review. @github announced (66 likes, 4 replies, 30,371 views); @GHchangelog announced (12 likes, 1 reply, 2,079 views).
- Multi-agent practice is becoming role routing. A custom Pi workflow assigns research and review to Claude Code and implementation to Codex, while a Copilot user reports a smaller model handling UI and routine tasks. @davis7 described (23 likes, 4 replies, 554 views); @burkeholland reported (57 likes, 10 replies, 3,715 views).
- Context continuity is now a product constraint. Replies on the ChatGPT/Codex and Claude surface discussion cite materially different mobile behavior, projects split by mode, and lost chat history. @emollick contrasted (99 likes, 18 replies, 7,321 views); @signulll argued (80 likes, 14 replies, 8,020 views).
- The builders are shipping workflow layers around agents. Public projects provide skills, marketplace plugins, browser context, and review checks; the strongest production example adds preview environments and shared data to the same model-assisted workflow. @jggomezt presented (2 likes, 2 replies, 75 views); @supabase reported (6 likes, 2 replies, 804 views).