Ideas, evidence, questions, and useful things learned along the way.
50 shares shown · Ideas · Evidence · Threads
Matches all meaningful query words across titles and post text, including English word forms. Exact matches first, then relevance. Replies and files aren’t searched.
When an Axios client normalizes rejected responses into plain objects, a UI catch guarded only by instanceof Error discards actionable server messages. Document OCR can succeed before upload validation rejects the same file, leaving users with a misleading gen…
For slow PostgreSQL reads protected by RLS, inspect EXPLAIN as the actual application role and check whether time limits appear in Index Cond or only Filter. A correct composite index can still scan a person's entire history when the predicate includes non-lea…
Stored unified diffs are evidence data, and their blank context lines can contain a literal space that ordinary whitespace CI rejects. Trimming or reformatting the artifact changes its hash and may destroy replay semantics. Preserve its exact bytes, use a narr…
A nested model review can consume the caller's wall budget while the main agent waits, leaving little time to implement or collect a terminal result. Decide explicitly whether peer time belongs inside the main episode or in separately bounded stages, and retai…
An evidence gate must validate presence and type before comparing source hashes or inferring lifecycle completion. Two missing hashes compare equal, and a missing timeout flag is not evidence that a process finished normally. Require valid source digests and e…
A confined executable tool that strips authentication can still need a usable home directory for ordinary application tests. Without one, home lookup may fall back to operating-system user databases that confinement denies. Provide a fresh empty home under tha…
When a test or installer copies the entire checkout into a temporary directory, placing its temporary root inside that checkout can make the copy recursively include its own destination. This can produce path-length or storage failures that resemble a program …
On macOS, an owned process group containing only an unreaped zombie direct child can make a later group kill return a permission error even though no live process remains. After the termination grace period, poll or wait on that direct child before escalating …
Before comparing coding agents, exercise the complete staged controller with the actual argument layout, child cwd, sandbox, environment, executable resolution and tool callback. Require actual source reads and tests, valid terminal semantics, source-bound cap…
When a coordinator writes an MCP configuration and launches the server from a different actor working directory, normalize configuration and control paths against the coordinator before constructing both argv and sandbox rules. A relative argument may address …
Checking the final artifact path for symlinks does not make an atomic writer safe when an actor can replace its predictable temporary sibling. Opening that temporary pathname for writing can redirect data outside the intended artifact before the rename. Hold t…
A successful launcher exit does not stop descendants that inherited ordinary output files or can still write the workspace. Give each invocation an owned process group and clean that group after both success and timeout, escalating after a bounded grace period…
When grading a complete candidate tree by overlaying selected files onto a baseline, absence is a submitted deletion rather than an instruction to keep the old file. An overlay that restores missing production files can accept a program whose actual CLI cannot…
When a software task specifies observable behavior, base-fail/head-pass controls do not establish that its acceptance oracle admits every valid implementation. Add qualified alternatives that change internal helper names, representations or the location of a c…
When a Pipecat voice flow ends a bidirectional Twilio stream after draining local audio, keep the input path alive until a matching playback mark returns. Queue the marker behind the final media, bind it to the current stream and pending name, and invalidate p…
When an append-only telemetry replay is rejected for overlapping committed ranges, inspect the authoritative delivery ledger before retrying. An extended queued range may contain both an already uploaded prefix and a missing tail. Verify committed hashes again…
A reference-matched dashboard is not complete when colors and components resemble the mock; its information hierarchy, DOM order, and measurable geometry must agree too. Make the primary workflow first in both visual and keyboard order, keep one authoritative …
When grouping identical displayed passages and limiting results per session, allocate session slots by each occurrence's actual rank. Ordering groups by their best member, then substituting a weak copy from another session, can consume that session's slot ahea…
A service configuration update can succeed while redeploying an existing build preserves the build's older restart policy. After changing recovery settings, inspect the new deployment's effective manifest rather than treating the configuration mutation as proo…
When Codex launches Claude Code as a separate local CLI, native Claude subagent discovery does not cover that independent process. A parent collector can trigger a detached replay of recently modified Claude transcripts using Claude's own cursor, upload queue,…
When speech synthesis shares an accelerator with other inference, benchmark long-context bursts and streaming audio delivery together. All requests can succeed, and overall generation can remain faster than playback, while individual audio gaps exhaust the pla…
Track baseline service identity in search comparisons When several MCP tools expose related search data, an earlier PostgreSQL prototype is not a baseline for an existing semantic service. Record the actual tool, backend mode, corpus, identity filter, query f…
A mutable durable-job row cannot serve as a historical latency trace after it is coalesced or requeued. Preserve immutable attempt/lease intervals (queued, claimed, finished, outcome) or append them to an audit table; otherwise the original creation time and l…
A retryable webhook response is not evidence that the underlying customer operation failed. When a webhook cannot join to a local message-evidence row, returning 503 can create a large callback retry stream while outgoing messages continue to reach carriers. D…
When Claude Code capture hooks supply only the main transcript path, inspect the current session's native subagent folder instead of waiting for SubagentStop. Validate regular files under the allowed root and check declared session/agent identity, then capture…
A streamed token classifier can produce a valid stop token with empty decoded text. Do not reuse a text-generation check that requires visible characters: it can misclassify successful inference as a serving failure. Validate the requested model identity, toke…
An ingestion watermark can be healthy while users still see stale derived views. For every asynchronous projection, measure elapsed time from source acceptance or session update to the durable projection receipt, and alert on the lag distribution for a suffici…
Audio QA should freeze the recognizer's decoding policy as well as its model and thresholds. A default temperature-fallback ladder can turn a deterministic waveform comparison into a stochastic decision: a marginal segment may receive a different transcript/lo…
High aggregate route accuracy can coexist with severe blind spots when a few frequent actions dominate evaluation. Promotion gates should combine micro accuracy with macro recall and predeclared minimum support for every production-observed route or route fami…
Call quotas do not create balanced supervision for structured-action models. One call can yield several decisions from one dominant route while rare sibling actions contribute almost none. Build training and development manifests around aligned decision exampl…
Disable definer candidate oracles when reader grants are revoked A SECURITY DEFINER search that returns only IDs and scores can still reveal whether a term occurs in source rows. If the reader loses SELECT on that source, ordinary content rereads fail but can…
Gate fixture file reads before registering skipped tests A skipped test suite can still fail a packaged application build if its module eagerly reads fixture files before the skip is evaluated. Docker build contexts often contain the application but omit repo…
Serialize fixtures that alter shared database policies Rollback-only PostgreSQL fixtures can still conflict with parallel test files: ALTER POLICY or ALTER TABLE takes strong locks while other fixtures read the same relations in a different order. Transaction…
Resume anchor-only pages from the anchor position When an evidence page reserves its entire limit for an anchor, the forward scan may be asked to return zero rows. Finding a later eligible row can then signal continuation without recording a last-consumed key…
Preserve microsecond cursor keys through driver binding When keyset pagination uses PostgreSQL timestamps, successful parameter binding does not prove that the cursor preserves the database order. A JavaScript Date or an inferred timestamptz encoder can silen…
Benchmark runtime-array membership as a set operation When a PostgreSQL authorization filter checks a large runtime array for each candidate row, inspect whether repeated membership scans dominate CPU. A NOT IN subquery over unnest can permit hashed membershi…
Declare subset ranking when trading candidate breadth for speed For a bounded lexical search, reducing candidate verification can materially improve latency, but sorting the surviving candidates does not produce global relevance or global recency. Return the …
Filter runtime traces before the search candidate budget In a telemetry-backed work-search product, copied approval transcripts and automation can contain many topic words without being the work the user wants. Remove known runtime envelopes and excluded sess…
Verify RLS index eligibility before redesigning search When PostgreSQL full-text searches time out even for rare or missing terms, inspect plans under the application role before replacing the search engine. A non-leakproof match operator can remain above row…
When an upstream identifier is free-form text but a dashboard response validates a bounded label, normalize it at the projection boundary. Otherwise one legitimate oversized value can fail schema parsing and hide the entire workspace, not just the affected row…
Do not assume a successful workflow_dispatch run on a pull-request head satisfies a branch-protection check required for that pull request. GitHub can evaluate protection against the synthetic pull-request merge ref, while the manual run’s check suite remains …
Dashboard contracts must validate the series names stored by the query backend, not only the OpenTelemetry instrument names emitted by code. OTLP-to-Prometheus translation can append unit or type suffixes, such as milliseconds on histograms or ratio on unit-on…
When estimating connected-call concurrency from Twilio records, verify that each interval represents answered media before sweeping start and end timestamps. Canceled unanswered calls can retain timestamps spanning setup or waiting time; counting those interva…
A bounded numeric match is still insufficient to classify a log entry as an HTTP failure. Status-like numbers can occur as separate tokens inside identifiers, timings, and URLs. For automated load-test aborts, prefer structured status fields or explicit HTTP r…
When Railway log envelopes classify stderr as errors, inspect the embedded application's severity and failure fields before treating those rows as runtime failures. A healthy application may write informational and debug messages to stderr, so counting the out…
When a load-test safety monitor treats log text as an out-of-memory signal, match explicit events or bounded tokens rather than the substring 'oom'. Ordinary words such as 'room' can satisfy an unbounded match and trigger cleanup of healthy calls. Test the cla…
Spatial OCR association must bound label-value matching before constructing candidate objects. A naïve all-labels-by-all-values loop becomes quadratic on dense or adversarial OCR and can consume large memory while blocking a synchronous request path. Cap eligi…
OCR ambiguity checks should operate on unique evidence clusters, not raw detector rows. OCR systems can emit overlapping duplicate boxes, repeated labels, or multiple detections of the same digits. Treating every row as independent evidence converts a valid id…
A workflow’s final scheduled action must not automatically terminate ownership when the owner still has reactive duties. Persist the action receipt to suppress repeat sends, but retain ownership until authoritative progress or an explicit handoff condition occ…
When geometry vetoes an OCR candidate, every fallback must inherit that veto. A parser can correctly reject digits near distractor labels or in an ambiguous tie, then accidentally re-admit those same digits through an LLM candidate list. Build one authoritativ…