kbcode ships continuously — every release lands in your terminal with kb update. This page tracks the latest updates.
📦 v1.73.06-preview
No changelog entries found matching "".
2026-08-14
Permission gate hardened, parallel batches share the token budget
Security: destructive Windows commands wrapped in quotes could slip past the permission gate — the command parser now understands Windows quoting, so the gate can't be bypassed, and legitimately-allowed commands stop re-prompting.
Error-rate feedback is measured per turn, so a thrashing session actually gets warned — even in long sessions.
Concurrent tool calls now share the token budget fairly instead of each claiming the full headroom (a step could previously overshoot by 4×).
Looping models get clear warnings instead of silence.
2026-08-14
Model onboarding probe judges substance, not phrasing
The onboarding honesty probe punished correct-but-explanatory answers, wrongly downgrading capable models. It now reads substance — a verbose correct answer passes, a real fabrication still fails.
2026-08-13
Smarter context handling and review timing
Adaptive compaction backoff: a run of compactions that don't actually help grows the wait before the next one; one useful compaction resets it.
Context-ceiling enforcement now targets a safe margin, so a small window no longer re-fires on the very next tool result.
Decision-reviewer round-trips are timed independently of the main model — a slow main model no longer inflates every review timeout.
Windows destructive commands (remove-item, cmd /c rd, erase) are scored as high-risk exactly like rm -rf.
2026-08-12
Model reasoning is never saved or replayed
Thinking/reasoning blocks exist only while a tool loop is in flight (providers need them for tool-result pairing) and are scrubbed at every persistence and replay surface — transcripts, resumed sessions, and provider history never store them.
The /thoughts command was removed with the accumulation it read from.
2026-08-12
No more duplicate Chrome windows
Browser automation now reuses your running browser — launches are guarded by command-line evidence and session-file proof, with a per-profile cooldown so the agent's retry loop can never storm a pile of new windows.
2026-08-12
Polite edit requests stay in code mode
"Can you add a retry to fetch?" no longer auto-switches to read-only ask mode — requests classified as edits keep their tools, while genuine questions still route to the lightweight mode.
2026-08-12
Decision-review follow-through and honest blast radii
Post-review nudges are consumed only by the call they actually apply to — an unrelated read no longer eats the correction before the real retry.
write_file blast-radius now counts content size: a full rewrite of a large file is flagged as the destructive action it is.
Git commands inside parallel batches run in order — no more .git/index.lock collisions from 12 concurrent git add calls.
2026-08-12
New /refactor command — guided multi-file refactor
/refactor <what> runs the full protocol in one go: blast-radius scan (every caller listed) → plan → atomic edits (callers updated in the same batch) → targeted tests → diff summary.
2026-08-12
Seven debugging improvements — the capture→fix loop is closed
/debug report — a pure-read ledger view: row/occurrence/open counts, recurring hints, last entries.
Self-repair now verifies its own fix by re-running the failing test — green marks it resolved, red puts it back.
Error→fix pairs are captured mechanically and the newest known fixes are injected into future sessions — the same error is recovered by retry, not re-diagnosed.
Identical errors coalesce into one incident with a count, bucketed by cause.
2026-08-12
Compaction pays one model call when the transcript is already under budget
When the first compaction pass already brings the conversation back under the threshold, the second pass is skipped — one summarization instead of two on the common path.