$ the-wire · showcase
Codex tames cloud-thread replays, opencode makes failed reads recoverable
By RepoJournal · Filed · About Agentic Coding · Composed from the cited sources · methodology
![opencode-agent[bot]](https://avatars.githubusercontent.com/in/1549082?v=4)
11 people shipped this
copyberry
@copyberry
2 cited
kitlangton
@kitlangton
1 cited
tlongwell-block
@tlongwell-block
1 cited
Abhijay007
@Abhijay007
1 cited
adamdotdevin
@adamdotdevin
opencode-agent[bot]
@opencode-agent[bot]
juanmichelini
@juanmichelini
phatloi0010
@phatloi0010
rajshah4
@rajshah4
Engel Nyst
@enyst
NoX1De
@NoX1De
The day's operator-relevant work is mostly about making agents stop wasting turns: opencode turns bad file reads into useful ones, Codex fixes the noise around resumed cloud threads, and goose opens live-voice instructions to per-session configuration.
fix(core): tail reads from the end of a file and suggest close file names anomalyco/opencode
A read past the end of a file now returns its last lines instead of an error, a negative offset reads from the end, and a missing file suggests close names from its own folder. The PR cites a week of local sessions in which 583 reads failed on an out-of-range offset and 1,526 on a missing file, each costing the agent another call or two to recover.
Treat goal objectives as untrusted context openai/codex
Goal objectives now reach the model as untrusted `goal_objective` context rather than being embedded in continuation instructions, and the model is directed to call `get_goal` when the objective is missing or truncated. Newly created objectives are injected into their originating active turn without dropping your client context, rendered as tool outputs when `UntrustedAsToolOutput` is enabled.
Avoid duplicate assistant messages when resuming Cloud threads openai/codex
Resuming a Cloud thread in the TUI no longer re-renders assistant messages that were already replayed from history: completed items are tracked by turn and item ID, so later starts, text deltas and completions for them are ignored. Unfinished items and new turns still stream normally.
feat(live-voice): configurable and per-start voice instructions aaif-goose/goose
goose takes voice instructions at two layers: `GOOSE_LIVE_VOICE_INSTRUCTIONS` in config applies to every voice session on the install, while an optional `instructions` field on `_goose/unstable/session/live-voice/start` applies to that interaction only, so a client hosting several agents can pass each agent's name, speaking style and context per start without disturbing goose's own live voice i...
fix(docs): correct SECURITY_PROMPT_THRESHOLD sensitivity direction aaif-goose/goose
Also worth knowing: the prompt-injection guide had the `SECURITY_PROMPT_THRESHOLD` table backwards. The scanner flags a tool call when `final_confidence >= threshold`, so lower thresholds are more sensitive, and the docs now say so, with 0.01 to 0.30 as Maximum and 0.80 to 1.00 as Very lenient. If you tuned that value from the old table, re-check it.
Action items
