$ the-wire · showcase
Codex hardens git remote parsing, opencode stops mangling clipped text
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
Most of today's work sits under the hood, but two changes reach operators directly: Codex now reads git remotes in a way that can't misparse clone filters or blow the 8,096-character metadata limit, and opencode stops cutting multi-byte characters in half when it truncates text.
Harden Git remote collection and limit remote URL metadata (#61488) openai/codex
Codex used to read remote URLs by parsing `git remote -v` output, which could mistake partial clone filter text for a remote entry; it now reads fetch URLs from NUL-delimited Git configuration, sanitizes credentials, and lets Git resolve `insteadOf` rewrites without contacting remotes. Ambiguous URLs resolve with bounded concurrency, which also keeps turn metadata under the Responses API's 8,09...
fix: keep text cuts from splitting characters anomalyco/opencode
Several places clipped text at a fixed count of UTF-16 code units or UTF-8 bytes, so a cut landing inside a character left a lone surrogate or a `�` at the edge of the string. The clipping now respects character boundaries, which matters wherever opencode truncates tool output, logs, or pasted content.
Describe context and permissions per environment behind a feature flag (#61494) openai/codex
Selected environments can carry different filesystem and network permissions, but context only ever described the primary environment's rules. The new `per_environment_context` feature, disabled by default, groups filesystem and network access under each environment ID and labels shared network requirements separately.
Report Cloud stream recovery failures and cancel abandoned reads (#61599) openai/codex
Cloud stream failures used to surface a generic reopen message even when the conversation was missing, access was denied, or sign-in had changed; they now report the specific reason, including failures during idle reconnection. Abandoned turn and item list requests are now cancelled instead of holding a native worker.
Allow Ctrl+C to quit during Cloud session switches openai/codex
Ctrl+C now quits while Codex is preparing a Cloud session: it clears buffered draft input first, then requests shutdown and abandons the pending switch. Abandoned thread-start requests are cancelled so they can't replace the active attachment or block later worker commands.
feat: reuse profile secret consent for automatic managed App enablement OpenHands/OpenHands
OpenHands reuses the Agent Profiles `ProfileScopeList` in the shared confirmation modal, so granting app-scoped managed-backend credentials and selecting secrets both run through one consent prompt. Less obvious from the headline: expect fewer separate confirmation dialogs when enabling a managed App.
fix: generate UUIDs without crypto.randomUUID on plain-HTTP non-loopback origins OpenHands/OpenHands
`crypto.randomUUID()` is a SecureContext-only API, so opening Agent Canvas over plain HTTP on a non-loopback host (LAN IP, VM, WSL2, Tailscale IP) left it undefined and starting a conversation failed with `crypto.randomUUID is not a function`. The fix generates UUIDs without that API, so those origins start conversations again.
fix(app): never count subagents in the new timeline anomalyco/opencode
The long tail: opencode's new timeline now gives each visible subagent event its own row with an agent chip and a link to the child session, with only background shells counting toward event counts, and dialog sizes are now consistent rather than a scatter of per-dialog overrides.