Duplicate skills override installed ones in one run of five, and the agent's reply names the swap 0.9% of the time

In 0.9% of the runs where a coding agent swapped one skill for a near-duplicate, the final reply said which skill it had used. That figure comes from One Skill Too Many, posted October 8, which mined 20,947 repositories for skill directories, found 822,109 candidate similar-skill pairs, and ran 312 confirmed pairs across three models for 6,368 runs and 542 agent-hours. Nearly one in four installed skills sits beside another that does the same job, and 37% of judged skills live inside copied collections. The similar skill takes one run in five from the installed one without lowering task completion, and runs that open it first lose over a third of the core functions only the installed skill provides, a ban on touching git being the paper’s example. Install location decides which skill runs; listing order barely matters. The conflict is settled at the first skill read, almost always before any file changes, and a pre-tool hook at that read restores fidelity. Anyone maintaining a shared skills directory should treat that hook as cheap insurance and should stop assuming a passing task means the intended instructions ran.

The same theme, a check that reports success without having checked, runs through Cross-Provider Review as a Runtime Contract for Coding Agents, dated October 7 and ingested today. The authors wrap a second-provider reviewer in a contract: separate resource pools, bounded execution, read-only capabilities, complete input delivery, typed failures, per-attempt evidence. In a 20-turn pilot the reviewer produced a material finding on eight turns (95% interval 19.1 to 63.9%). The more useful result is the list of ways their own harness lied. A reviewer that received 1% of its input still returned a clean verdict and was counted as a completed review; so was one cancelled during process reaping, and one that exited nonzero with a well-formed verdict. The third defect surfaced after a preregistered study had already logged 25 observations, and because exit status was never recorded per attempt, they restarted the cohort at zero. In read-only probes Codex attempted writes in five of five trials; every write tool failed and no repository changed.

Cadence, dated October 8, moves the supervision inside the run. Its monitor gives advisory guidance when an agent’s trajectory looks healthy and replaces the agent’s next step when it has gone badly wrong, then tightens inspection frequency after a replacement and relaxes it after an advisory. On 300 SWE-bench Lite tasks it adds 76 resolved tasks to mini-swe-agent (25.33 points) and 47 to Moatless (15.67 points), with token cost the authors describe only as competitive with other monitors. Fixed-interval inspection is what most homegrown watchdogs do, so the scheduling result is the part to copy.

Two papers from October 8 attack evaluation from the user side and the judge side. SWE-Journey synthesizes long-horizon tasks and simulates four user personas mined from real interaction data. Models pass over 75% of tests for requested functionality when the simulated user is a software architect and under 25% when the user is a non-coder, on the same tasks. AgentHorizon tests the judges themselves: 1,373 instruction-trajectory pairs from 166 hours of human-recorded computer use, with negatives built by swapping in a closely related instruction. The best agentic judge, GPT-5.5, reaches 80.9% balanced accuracy on the frontier split, and giving judges tools hurt the open-weight models. A judge that is wrong one time in five on near-miss trajectories is a weak reward signal for training and a weak gate for evaluation.

Runnable Commit Untangling for Coding Agents asks whether cleaning up agent patches pays off downstream. RucTangle splits 131 agent-generated patches into ordered commits that all leave the code runnable, where four baselines produce unrunnable histories 20.6% to 37.4% of the time. Feeding those histories to other agents repairing 453 regression-introducing patches lifts pass@1 by 5.2 points. It is a modest number, and it is the first direct measurement I have seen that commit hygiene helps the next agent.

On retrieval, Finding the Right Balance settles a disagreement about diversification in RAG. On clean candidate pools, diversifying hurts relevance, evidence coverage, and answer quality. It helps on multi-evidence questions once overlapping chunks or near-duplicates cause nearest-neighbor retrieval to return the same passage repeatedly. Their rule diversifies only when the effective number of distinct documents in the top-k falls below what the query needs, computed from embeddings you already have.

Nothing in the last day and a half landed squarely on semantic layers or governed analytics. The thing to watch is whether skill platforms start exposing which skill ran, and whether the cross-provider study’s restarted cohort produces a gate result.

Feed note: the mirror’s status endpoint reported a last sync of September 1, 2026, although items dated through October 10 were present; treat coverage of today as possibly incomplete.