# Per-Commit Codex Loops, DeepSeek V4 Flash, and Script-Writing Search Agents *By Coding Agents Alpha Tracker β€’ April 29, 2026* Peter Steinberger's live setup puts Codex on every commit, with auto-fix PRs and review loops catching regressions fast. Also here: DeepSeek V4's price/performance story, local Flash inference on 128GB Macs, Sourcegraph's script-executing Deep Search, and a few operator-grade workflow lessons. ## πŸ”₯ TOP SIGNAL The biggest practical shift today is **agentic CI/CD**. Peter Steinberger says Codex now reviews every landed commit, spawns a fresh Codex to open a fix PR when it finds a bug, then hands that PR to a review/fix loop that can run up to five times; in a separate per-commit-to-main setup, it found one of his own regressions within 10 minutes. [^1][^2] Timeless takeaway: put the agent on the commit path, keep the loop bounded, and use PRs as the handoff + audit boundary. [^1][^2] ## πŸ› οΈ TOOLS & MODELS - **DeepSeek V4 Flash** β€” Salvatore Sanfilippo says Flash is the real local-agent story in the V4 release: he implemented **2-bit asymmetric quantization**, runs it on a **128GB MacBook**, and says **tool calling works perfectly**. He compares it to recent Sonnet-level performance, while being more cautious than frontier-model claims. [^3] - **Flash vs. Pro** β€” Same source argues **V4 Flash**, not Pro, is ready now for local inference, and says Pro's training is not finished yet. He also says Flash beats **Kimi 2.6** at roughly **1/3 the size** and has more disciplined thinking behavior than Qwen 3.6. [^3] - **Local throughput reality** β€” On his MacBook, Sanfilippo reports about **120-130 TPS prefill** and **21+ TPS generation**, with the warning that **prefill is the real bottleneck** for coding agents. [^3] - **Sourcegraph Deep Search** β€” now writes and executes scripts to analyze codebases, then feeds results back into the agent. Sourcegraph frames this as **custom tools on demand**; example query: β€œTop Files in VS Code Bugfix Commits (Last 6 Months) and Contributors.” [changelog](https://sourcegraph.com/changelog/deep-search-evaluator) [^4][^5][^4] - **Cursor / EndorLabs benchmark signal** β€” Jediah Katz shared EndorLabs' latest correctness-and-security benchmark, which he says had **Cursor's optimized harness** on top. Useful if you're tracking harness quality, not just model choice. [^6][^7] - **Codex endurance** β€” Tibo says that with some small tweaks, **Codex can work for days on hard tasks**, and that changes are coming to make this easier to use. [^8] ## πŸ’‘ WORKFLOWS & TRICKS - **Per-commit auto-fix loop** 1. Run Codex on every landed or main-branch commit. 2. If it finds a regression or security issue, spawn a new Codex instance to open a fix PR. 3. Hand that PR to a review agent. 4. If review finds problems, spawn another fix agent and loop again β€” Steinberger caps it at **5 passes**. 5. Use the PR as the audit trail. [Example PR](https://github.com/openclaw/openclaw/pull/74065) and [example commit record](https://github.com/openclaw/clawsweeper/blob/main/records/openclaw-openclaw/commits/212a32648fe70e9f8088d8145736a0e31e6ba0b3.md) [^1][^2][^1][^2] - **Script-escape pattern for repo analysis** β€” When the native agent loop gets stuck, let the agent write and run a one-off script, then feed the result back into the loop. That's the core pattern in Sourcegraph's Deep Search update, and it generalizes to churn analysis, migration prep, and codebase archaeology. [^4] - **Measure the latency that actually hurts** β€” For local coding agents, Sanfilippo says **prefill**, not generation, is the real constraint. If you're evaluating quantization or model swaps, benchmark the part that blocks long-context coding work. [^3] - **Expect API exhaustion once agents scale** β€” Steinberger says agents can hit GitHub rate limits even after a move to Enterprise. In his sessions, Codex worked around GitHub limits via the browser, typed into a comment box to close an issue, and opened Cloudflare to create a new API key when permissions were missing; he also plans to test [ghx](https://github.com/brunoborges/ghx). [^9][^10][^11][^12][^9] - **Claude Code config check** β€” AI Builder Club warns that many setup guides still recommend deprecated `npm install`, and Jason Zhou called out the overlooked `.claude/rules/` directory. Small detail, real leverage. [^13][^14] ## πŸ‘€ PEOPLE TO WATCH - **Peter Steinberger** β€” strongest operator signal in today's notes: per-commit Codex reviewers, live fix PRs, browser fallbacks, and the ugly reality of GitHub API saturation. [^1][^9][^10] - **Salvatore Sanfilippo** β€” brings the local-inference details most model chatter skips: quantization method, RAM target, throughput numbers, and a clear argument that **Flash**, not Pro, is the local-agent story right now. [^3] - **Daniel Neal Adler / Sourcegraph** β€” high signal if you care about agents that inspect large codebases, because he's shipping the β€œwrite code to understand code” pattern into product. [^4] - **Jason Zhou** β€” useful for catching low-drama but high-leverage setup details like deprecated Claude Code install paths and hidden config surfaces. [^13][^14] - **Tibo** β€” short post, strong implication: long-running Codex sessions are getting easier, which matters if your hardest tasks aren't one-shot edits. [^8] ## 🎬 WATCH & LISTEN - **0:02-2:32 β€” Why DeepSeek V4 Flash matters more than Pro for local agents.** Sanfilippo walks through the practical case: 2-bit asymmetric quantization, 128GB Mac target, working tool calls, and why Flash is the interesting part of the release for local inference. [^3] [![La perla di casa DeepSeek Γ¨ il modello Flash, non il Pro, almeno per ora](https://img.youtube.com/vi/2gOhmng0LFI/hqdefault.jpg)](https://youtube.com/watch?v=2gOhmng0LFI&t=1) *La perla di casa DeepSeek Γ¨ il modello Flash, non il Pro, almeno per ora (0:01)* - **21:02-23:58 β€” Codex using the browser as a test harness.** Riley Brown shows Codex turning an HTML file into an app and then validating buttons, navigation, and quiz flows by controlling the browser itself. [^15] [![Learn 95% of Codex in 30 minutes](https://img.youtube.com/vi/474wZZHoWN4/hqdefault.jpg)](https://youtube.com/watch?v=474wZZHoWN4&t=1262) *Learn 95% of Codex in 30 minutes (21:02)* ## πŸ“Š PROJECTS & REPOS - **[openclaw/openclaw PR example](https://github.com/openclaw/openclaw/pull/74065)** + **[clawsweeper commit record](https://github.com/openclaw/clawsweeper/blob/main/records/openclaw-openclaw/commits/212a32648fe70e9f8088d8145736a0e31e6ba0b3.md)** β€” best repo-level signal today because it's live evidence of agentic CI, not a concept deck. One link shows the fix-PR loop; the other shows a regression caught almost immediately after launch. [^1][^2] - **[ghx](https://github.com/brunoborges/ghx)** β€” niche, but relevant if your bottleneck is GitHub API exhaustion rather than model quality; Steinberger singled it out while troubleshooting agent-heavy workflows. [^9] *Editorial take: today's durable edge is orchestration β€” commit hooks, bounded review/fix loops, script escapes, and browser fallbacks matter as much as the base model.* [^1][^4][^10] --- ### Sources [^1]: [𝕏 post by @steipete](https://x.com/steipete/status/2049356949523730699) [^2]: [𝕏 post by @steipete](https://x.com/steipete/status/2049290741013262522) [^3]: [La perla di casa DeepSeek Γ¨ il modello Flash, non il Pro, almeno per ora](https://www.youtube.com/watch?v=2gOhmng0LFI) [^4]: [𝕏 post by @DanielNealAdler](https://x.com/DanielNealAdler/status/2049157567365189706) [^5]: [𝕏 post by @Sourcegraph](https://x.com/Sourcegraph/status/2049187660208693460) [^6]: [𝕏 post by @jediahkatz](https://x.com/jediahkatz/status/2049212952381731188) [^7]: [𝕏 post by @jediahkatz](https://x.com/jediahkatz/status/2049213025069015285) [^8]: [𝕏 post by @thsottiaux](https://x.com/thsottiaux/status/2049362299392241784) [^9]: [𝕏 post by @steipete](https://x.com/steipete/status/2049244352057094645) [^10]: [𝕏 post by @steipete](https://x.com/steipete/status/2049035044702843053) [^11]: [𝕏 post by @steipete](https://x.com/steipete/status/2049035851493953750) [^12]: [𝕏 post by @steipete](https://x.com/steipete/status/2049046713357254796) [^13]: [𝕏 post by @aibuilderclub_](https://x.com/aibuilderclub_/status/2049277582713999468) [^14]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2049311955975086455) [^15]: [Learn 95% of Codex in 30 minutes](https://www.youtube.com/watch?v=474wZZHoWN4)