Skip to content

Known issues

Defects, gaps and unproven claims in the 0.11.0 line, kept here so every maintainer and reviewer sees the same list. Confirmed means reproduced or read in the code at this commit. Reported means found in an earlier session and not re-checked at this commit. A fix removes its entry in the same commit and names it in the changelog.

For 0.11.0-rc.1 every Confirmed entry was re-checked on 2026-10-28 against the stack top 2bffac4, which has every follow-up PR through #289 merged. Entries fixed since they were recorded are listed under Fixed before 0.11.0-rc.1 with the PR that fixed them. The two review-speed entries are proposals rather than defects and moved to the roadmap’s medium-term proposals.

What goes wrong Status
review-prepare diffs against the branch’s recorded base (git config branch.<name>.base). After a branch is rebased, the recorded base must be updated, or the review captures unrelated files from the new base. Confirmed at 2bffac4 (src/harness/independent-review.ts). #275 made a sandbox record the commit it forked from; nothing updates the recorded base after a rebase.
What goes wrong Status
uh queue has only been tested with fake launchers; no live queue has run. Confirmed. #273 changed how the queue launches missions (like uh mission run), again tested with fake launchers only.
The hive has not run live: no real uh land, uh queue or uh verify has recorded a fact yet; every hive behaviour is proven only by the test suite. Confirmed at 2bffac4.
The hive’s last changes (project-owned hive and review root, evidence resolved against the owning project, facts recorded by uh verify) were not independently reviewed: the last review, with no contradicted claim, covered the slice before them. Each change has a regression test that fails without it. Confirmed. The hive follow-ups merged since (#274, #282, #289) were not independently reviewed either.
What goes wrong Status
uh steer on an orchestrator stops its session and leaves the workers it launched orphaned; they stop mid-work without committing. Steer an orchestrator only between its steps. Confirmed (an orchestrator steered mid-wave orphaned a worker at turn 42); src/harness/steer.ts still does not stop or hand over an orchestrator’s workers at 2bffac4.
Real settled runs have not produced deliveries to configured notification sinks; uh notify test delivers. Reported.
A Windows toast sent through the windows-toast preset is reported as a handoff: PowerShell accepted it, but whether Windows displays it (registered app id, notification settings) cannot be confirmed by UH and has not been checked by eye since the app id changed. Confirmed (by design, unverified live). #284 hardened the app id quoting; nothing was checked by eye since.
The toast and notification-report change was recovered from a worker stopped before it finished and was not independently reviewed. Confirmed. Its follow-up #284 adds tests but was not independently reviewed either.
uh ps turn counts for Claude Code (distinct assistant message ids, not counting a sub-agent’s messages, which carry a parent_tool_use_id) and ACP (agent activity until a tool call completes) are tested on recorded streams only; the ACP rule is a heuristic that has not been checked against a live run. Confirmed. Since #283 a Claude Code run’s count is the runtime’s own num_turns once its final record arrives; the ACP rule is unchanged.
Supervision now counts Claude Code turns, so limits.max_turns can stop a Claude Code run where it previously never counted; not yet observed on a live run. Confirmed (behaviour change). #283 stopped counting a sub-agent’s messages; still not observed live.
When neither the run’s control file nor its digest has a turn count, uh ps reads the run’s whole events.ndjson to count turns, which is slow for very long runs. Confirmed in src/harness/live-runs.ts, mitigated by #283: only for runtimes whose turns can be counted from events, and cached by file size and modification time, so the file is read again only after it changes.
The turn-count change was recovered from a worker stopped before it finished and was not independently reviewed. Confirmed. Its follow-up #283 was not independently reviewed either.
uh wait returned orphaned for a run that passed. Reported.
A blocked run can leave a second run id inside a sandbox, and UH_RESULT can name a run directory that was never created. Reported.
Eight concurrent sandbox creations failed once on Linux CI: one git worktree add failed under contention (tests/sandbox.test.ts, “eight concurrent createSandbox calls all register”). The same test passed on the other Linux CI runs of this line; it has not been reproduced. Confirmed once (Linux CI). The test passed in Linux CI on 2bffac4.
On macOS, uh mission run-team can wait up to admission_timeout_ms (20 minutes by default) before admitting a single worker on a machine with plenty of usable memory: os.freemem() counts only free pages there, not inactive or purgeable memory, so the default 700 MB worker plus 1024 MB reserve rarely fits. tests/cli-notification-exits.test.ts (“run-team announces team.settled”) times out on a busy Mac for the same reason. #291. Confirmed on macOS at 2bffac4; Linux reads reclaimable memory and is not affected.
What goes wrong Status
Package installs issued through Command Code’s PowerShell tool are not denied as package_install. Reported.
A .harness path inside a worker’s own scratch directory is treated as guard_tamper and hard-stops the run. Reported.
PowerShell targets built with (Join-Path ...) are denied as literal paths; nohup cmd 2>&1 without a trailing & is denied as containment_escape. Reported.
oh-my-pi’s eval and python tools bypass path checks. runtime_config.tools can withhold them; the default tool set still allows them. Reported.
The oh-my-pi guard logs only denials, so a run with no denials cannot prove its guard loaded. Reported.
What goes wrong Status
uh verify and review-prepare fall back to the main checkout without saying so when a mission’s worktree no longer exists. Reported.
A run stopped by its turn cap can settle passed instead of turn_limit (acceptance R7-turn-cap, R7-turn-cap-cmdc). Reported (last acceptance campaign).
S3-unknown-cost-cmdc: both workers fail with unknown cost on Command Code. Reported (last acceptance campaign).
Every JEV decision receipt observed so far is uncertain with confidence 0. Reported.
What goes wrong Status
Command Code 1.62.1 exits with an uncaught write EOF after a single tool result of roughly 70K characters or more (CommandCodeAI/command-code#859). The guard forces large file reads into windows; large shell output is not limited. Confirmed upstream; reads mitigated only. The upstream issue was closed as completed on 2026-10-06; which Command Code release carries the fix has not been checked, so the read windows stay.
Claim Why it is not proven
oh-my-pi runtime_config.tools and ACP runs in sandboxes behave as tested Unit-tested only.
Prompts over 32K reach Command Code and oh-my-pi No live run since prompts moved to stdin and files.
Codex route attestation No live run.
Command Code cost reporting Streams report tokens but no price.
uh mcp serve in a real client Tested over stdio in the test suite; not used from a live client.
What goes wrong Status
The docs site at the 0.11.0-rc.1 tag (apps/docs) synced only the roadmap from docs/; its other pages were maintained separately and did not include the handbook or the newer runbooks. Fixed after 0.11.0-rc.1 by #250: apps/site (uh.agenticeng.app) mirrors all of docs/, specs/ and CHANGELOG.md on every build; apps/docs and its sync script are removed.
docs/verification/capability-inventory.md predates this development line; capabilities added since are listed in its section 5 without per-row test and evidence detail. Confirmed at 2bffac4.
What went wrong Fixed by
On Linux, a shell delete of the hive (rm -rf "<project>/.harness/hive") was refused as delete_outside instead of stopping the run as guard_tamper (tests/hive-integrity.test.ts, 4 failing tests on Linux CI). The guard read the /-prefixed path as a cmd.exe switch and skipped it. #240 (layer 01, f75a076): /x tokens are switches only for cmd.exe verbs. The 4 tests pass in Linux CI on 2bffac4.
--root on uh notify detect, list and test had no CLI test. #284 (tests/cli-notify.test.ts, “uh notify with –root”).
Team memory reservations were released by controller pid, so workers sharing one controller pid could release each other’s reservations and over-admit (recorded as a hypothesis). #280: a reporting worker releases one reservation of its controller, not all.
  • uh mission run-team takes a mission id, not a path; a path fails with Invalid mission id.
  • uh mission put of a team packet needs its worker packets in the same call.
  • uh report and uh steer take a run id (or a unique prefix), not a mission id.
  • uh land requires a clean target worktree and a retained worktree for each worker branch (run-team --retain).