Before an agent ledger write: open a connection, enable and verify foreign keys before BEGIN, then reject an action whose approval row is missing.

Your agent ledger references an approval. SQLite may not enforce it.

A REFERENCES clause can exist while enforcement is off. This runnable regression separates schema, connection settings and actual approval policy.

November 13, 2026 · 7 min · 1301 words · Thomas De Vos
Read Your agent ledger references an approval. SQLite may not enforce it.
JSON false passes an isinstance integer check and reaches an unlimited branch. Exact type validation rejects false; a valid zero admits no calls.

Your agent's call limit is false. Python accepts it as an integer.

Boolean values, integer limits and zero-as-unlimited defaults can combine into an admission bug. Test the decoded value and the work it permits.

November 12, 2026 · 6 min · 1185 words · Thomas De Vos
Read Your agent's call limit is false. Python accepts it as an integer.
An allowed RAG seed expands into three candidates. Per-chunk tenant and group checks reject two before body loading and admit only a0 to the prompt.

Your RAG search checks permissions. Does neighbour expansion?

An authorised search hit does not authorise its neighbours. A runnable fixture catches cross-tenant expansion and missing chunk permissions before context assembly.

November 11, 2026 · 6 min · 1233 words · Thomas De Vos
Read Your RAG search checks permissions. Does neighbour expansion?
A nullable unique key admits two rows for tenant-a and NULL. Explicit NOT NULL rejects the missing key before admission.

Your agent retry ledger has a unique key. NULL still gets through.

A SQLite UNIQUE constraint can admit repeated missing operation IDs. Run a small fixture, repair the schema and keep duplicate admission separate from remote execution.

November 10, 2026 · 7 min · 1321 words · Thomas De Vos
Read Your agent retry ledger has a unique key. NULL still gets through.
Two wrong answers arrive in reverse case order. Positional grading reports two passes; joining by case ID reports zero.

Your agent eval passed the wrong answers

Parallel evaluation results can arrive out of order. A runnable Python fixture shows why zip(strict=True) cannot protect case identity, and how to validate the join.

November 9, 2026 · 6 min · 1219 words · Thomas De Vos
Read Your agent eval passed the wrong answers
NaN bypasses a reject-below-threshold check. A repaired gate validates the score first and returns INVALID before threshold comparison.

Your agent eval returned NaN. The release gate passed.

A check that rejects scores below a threshold can approve NaN. Validate the score contract before making a release decision.

November 8, 2026 · 6 min · 1275 words · Thomas De Vos
Read Your agent eval returned NaN. The release gate passed.
In a harmless parser fixture, the argument --delete enables an option even with shell=False. An option separator keeps it as data; identifier validation rejects it.

Your agent uses shell=False. The tool still accepts an injected option.

A subprocess argument list preserves argument boundaries. The called program still decides whether each argument is data or an instruction.

November 6, 2026 · 7 min · 1360 words · Thomas De Vos
Read Your agent uses shell=False. The tool still accepts an injected option.
A 2068-byte gzip fixture passes a 64 KiB input check but expands to 2 MiB. A bounded decoder rejects it after detecting that it exceeds a 256 KiB output limit.

A small gzip tool result can exceed your agent's output budget

A compressed-byte check does not bound decoded output. Test the decompressor’s allocation contract before passing tool results to an agent.

November 5, 2026 · 6 min · 1256 words · Thomas De Vos
Read A small gzip tool result can exceed your agent's output budget
A 30-second budget has 10 seconds left after 20 seconds of work. A wall-clock step backwards incorrectly reports 130 seconds. A monotonic deadline retains 10 seconds.

Your agent's 30-second timeout grew when the clock moved backwards

Use one monotonic deadline for local elapsed-time budgets. Queue waits, retries and backoff must spend the same allowance.

November 4, 2026 · 6 min · 1230 words · Thomas De Vos
Read Your agent's 30-second timeout grew when the clock moved backwards
A lexical path check accepts allowed/docs/note.txt, but docs is a symlink outside. A narrower reader accepts one filename and opens it relative to a trusted directory descriptor.

Your agent's file path is allowed. Its symlink points outside.

An allowed path can lead to an unauthorised file. A runnable Python lab demonstrates lexical checks, a deterministic check/use race, and a bounded file reader.

November 3, 2026 · 8 min · 1552 words · Thomas De Vos
Read Your agent's file path is allowed. Its symlink points outside.