$ the-wire · showcase
Codex raises the prompt cap, Gemini CLI guards shell expansions
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
Codex drops its 4,000-byte ceiling on thread/sample prompts and reworks how sandbox proxy settings and status-line config are resolved, while Gemini CLI v0.69.0 plus two policy fixes tighten and correct what your shell commands are allowed to do.
Remove the 4,000-byte prompt limit for `thread/sample` (#62378) openai/codex
thread/sample no longer rejects prompts over 4,000 bytes at request validation; expectedTurnId and prompt still have to be nonblank. If you script long prompts against the app server, the old size ceiling that forced you to truncate is gone.
fix(core): keep ASK_USER for shell commands with unresolved expansion google-gemini/gemini-cli
PolicyEngine.applyShellHeuristics() used to classify a command from its unexpanded text and upgrade read-only defaults to ALLOW, even when bash rewrites the arguments via brace expansion, parameter expansion, ANSI-C or locale quoting, arithmetic expansion, substitutions, or unquoted globs. Those commands now keep the ASK_USER decision instead of being auto-approved on a classification that did ...
Release v0.69.0 google-gemini/gemini-cli
v0.69.0 fixes the core,cli path that mangled literal $ characters during template placeholder replacement and prevents __GCLI_EXPAND_TARGET__ env collisions in expandEnvVars, and adds a /stats heatmap subcommand for daily token usage. Worth taking if any of your prompts or templates contain literal dollar signs.
Expose custom status lines in `/experimental` openai/codex
Custom status lines move out of a TUI client feature flag and into a Custom status line opt-in under /experimental, saved on the connected app server, with the provider picked via /statusline and read from tui_extensions_statusline. That fixes the case where the client flag and the server's active profile disagreed about whether the feature was on.
Prepare managed proxy settings once per sandbox command openai/codex
SandboxManager now prepares the managed proxy environment and sandbox network context together once per command, rather than re-preparing at launch and overwriting settings already bound to the command, including credentials set up for its shell. If you route sandboxed commands through a proxy, per-command credentials should stop getting clobbered.
docs: deprecate GLM-5.3-Flash Free (#64600) anomalyco/opencode
by opencode-agent[bot]
Elsewhere: GLM-5.3-Flash Free is deprecated in opencode's docs, OpenHands shipped 1.42.0 with connected-git-provider selection for repositories, and Goose added llmman as a declarative Ollama-compatible provider defaulting to http://localhost:17434 with dynamic model discovery.
Action items
- → Upgrade gemini-cli to v0.69.0 and move any custom scripts off /experimental ordering assumptions google-gemini/gemini-cli [plan]
- → Whitelist a replacement model in opencode now that GLM-5.3-Flash Free is deprecated anomalyco/opencode [plan]