no-human-ai/no_human

Problem _DIFF_CAP = 60_000 (review/reviewer.py) applies to the diff in the main review, not only the angle passes. Git's path order means tests/ is usually what gets cut. 2026-09-09 to 2026-09-15 Reviews gate reviews checked 269 diff truncated (median full size 91k chars) 141 a changed file with neither patch nor full text in the prompt 121 of those, no tool call naming that file 43 of those 43, PASSED 26 Proposal Budget the diff per file, so every changed file's patch appears up to its share. List every file whose patch was cut, by path, in the prompt. If a non-excluded changed file was never inspected, route the result through the existing ReviewerUnavailable path (re-run, then escalate). Never PASS, and never spend one of the coder's attempts on a reviewer omission. Exclusions come from a trusted, documented list, not from files the coder can edit. Acceptance criteria With a diff over the cap, every changed file's patch appears up to a per-file budget and cut fi

Found while unblocking PR #308 (an external contributor's PR, red on this). Reproduced on that branch with the tool from main. What happens tests/test_readme_claims.py fails on citation drift and the failure message says: Run uv run python scripts/reanchor_citations.py --apply to re-anchor it; the symbol resolves, so the rewrite is exact Running exactly that leaves the tree worse than before and exits FAIL. A citation is pinned in two places — the prose in docs/*.md and the CITATION_TABLE row in tests/test_readme_claims.py. The tool's own module docstring says it rewrites both ("CITATION_TABLE row to the line the content now lives on"). Measured, it rewrites only the doc: before --apply after --apply docs/eval.md 7736 / 7765 / 7887 7745 / 7774 / 7896 <- correct CITATION_TABLE 7736 / 7765 / 7887 7736 / 7765 / 7887 <- untouched They agreed before and disagree after. Test count went from 3 failures to 4, because test_the_citation_table_co

no_human's review gate runs on machine-checkable signals: tests, held-out tests, the tamper guard, the repro gate, ruff lint evidence, wiring evidence (see docs/verification.md). One signal that every serious agent harness has converged on is missing from that list: type/compile diagnostics. This issue is the plan for adding it, written after a deep dive into what actually works, because "integrate LSP" hides two very different features with very different evidence behind them. Diagnostics as a verification signal is the proven half. SWE-agent's ablation credits its edit-loop linter (rejecting syntax-breaking edits) with roughly +3 points on a 12.47% baseline — the single most-cited agent-computer-interface feature of that paper (NeurIPS 2024). Aider lints after every LLM edit and feeds errors back for a fix round, by default. Claude Code's native LSP support (CLI v2.0.74) documents automatic post-edit diagnostics as the headline benefit — the model notices the error it just introd

The web board (web/src, React + node --test unit tests) works, but it was built by an engineer optimizing for correctness, and it shows. This is an umbrella invitation for UI/UX contributions — frontend and design skills welcome, no Python needed for most of it. Concrete rough edges observed on a fresh install (2026-09-01, v0.1.8 board against demo data — pick one, or report your own): First-run noise — READ THE CONSTRAINT FIRST (amended 2026-09-11). The ! badge on Settings is not noise and must not be hidden, dimmed or deferred while onboarding is incomplete: it is the only persistent signal that the install is not finished, and a board whose AI configuration is half-done is a board that cannot do its job. Proposals that suppress it will be declined, however elegant. What is open: the "Complete AI configuration" popup overlaps the sidebar stats card, and the header can say "nothing needs you" on a screen that plainly needs something. Fix the overlap, or the contradictory header

KI-1 in docs/KNOWN_ISSUES.md records a real product defect: with two tasks running concurrently against one Store (concurrency.enabled: true, max_workers > 1), an attempt could die on sqlite3.OperationalError: cannot commit transaction - SQL statements in progress with the traceback entirely in shipped code (Orchestrator._run_attempt → Store.update_attempt). The only test that drives that configuration, tests/test_scheduler.py::test_two_repos_run_concurrently_in_worktrees, measured a 3/8 serial failure rate on 2026-07-30 and is deselected in CI to keep the badge truthful. Since then, a mitigation landed: every Store write goes through the serialized_write lock in src/no_human/core/db.py. The failure rate has not been re-measured since. KI-1's own status line says: re-measure, then re-enable or close. The task — measurement first, code second (possibly no code at all): Re-measure: run uv run pytest -q tests/test_scheduler.py::test_two_repos_run_concurrently_in_worktrees serially,