The setting you wrote is no longer read

01The invoice still comes back

Suppose a team updates an invoice-extraction workflow. It keeps project instructions in AGENTS.md, exports telemetry for review, and has tested long invoices against its serving limit. After the upgrade, a small invoice still produces a plausible answer. That confirms a working request path. The instruction file, export destination and longest supported input remain separate questions.

The September Opus 5.5 launch illustrates the obvious failure. The API rejects thinking types disabled and enabled, forced tool choices any and tool, and, on the Claude API and Google Cloud, the older computer_20251124 tool. These cases return 400 errors, with migration guidance describing replacements. A caller that exercises the affected request sees a failure immediately. ¹

The historical table contrasts that signal with changes that allow work to continue. It preserves the September versions and scope; it is not a list of defects asserted to remain in current releases. The invoice example is hypothetical, and we did not run these products to reproduce the behaviours.

02An instruction file that never reached the model

For the invoice team, an ignored AGENTS.md could look like weak instruction following. Przemek’s 23 September report described a loader in the Claude Code 2.1.280 bundle behind a remote flag. In his canary tests, disabling telemetry or nonessential traffic prevented the file from loading; setting those variables to 0 did not restore it. He reported no warning. These are attributed measurements with a published method, not tests we repeated. ³

The environment-variable documentation explains the unusual value rule: any nonempty DISABLE_TELEMETRY value, including 0 or false, opts out and disables feature-flag fetching. The invoice team’s environment therefore matters alongside the file it committed. An explicit-looking value can express the opposite of what its author intended. ¹⁶

The current memory guide, read on 18 October, says that before 2.1.281 some sessions with telemetry disabled or on Bedrock read CLAUDE.md only. It also documents a first-session exception after certain upgrades. That is evidence for the documented version boundary, not proof that every fresh runner passes. A canary on the actual runner distinguishes a missing file from a model that received and disregarded its instruction. ⁵

03An exporter setting can lose authority

Claude Code 2.1.282 made a separate, intentional change: project and local settings ignore OpenTelemetry variables that enable export, set its destination or capture content. The release names CLAUDE_CODE_ENABLE_TELEMETRY and OTEL_LOG_* among the examples. It also adds startup, /status and claude doctor notices for ignored project telemetry values. ⁷

If the invoice team enabled export only in its project file, that entry no longer establishes that export is active. Another supported source may still configure telemetry. The useful question is which destination the running session actually uses. A successful invoice response cannot answer it, and an ignored-value notice does not by itself describe every remaining setting.

The team can send a harmless sample through its real launcher and inspect the expected telemetry output at its approved destination. That check is our proposed acceptance step. It observes the behaviour the project relied on, while the release’s notice explains why a particular file stopped controlling it.

04The long invoice and the long session

vLLM 0.30.0 changes a different part of the example. Its release removes GPTQ activation ordering, ignoring g_idx, and aligns YaRN with Transformers. The note replaces attn_factor and extrapolation_factor with mscale and attention_factor. It also removes a second scaling of maximum positions: its TeleChat3 example falls from 131072 to 32768 derived tokens. ²

Correcting that arithmetic can still invalidate the team’s tested input envelope. The engine may answer a short invoice while a long one reaches the revised limit. The release does not document a runtime warning for every ignored field; we have not checked its startup output. A request at the intended maximum input exercises a dependency that a short smoke test misses.

Cline CLI 3.0.63 records another delayed failure. Its summarizer retained an old OAuth token while the main loop refreshed its own. When the token rotated, a 401 was swallowed by a truncation fallback, leaving a chopped transcript instead of a summary. The release repairs credential and model tracking. For an invoice session, a short successful exchange would never have exercised that later transition. ⁹

05Defaults and working directories count too

Worktrunk 0.79.0 changes wt switch --no-cd -x so the launched program runs in the new worktree; --no-cd controls the shell’s location. An invoice test can execute successfully against the wrong checkout if its script assumes the earlier directory. The release labels the change Breaking. ¹⁰

GitHub’s 24 September Copilot announcement concerns an unset choice. From 22 October, eligible generally available features left Unconfigured follow the selected global default; explicit choices remain, and previews stay opt-in. The policy includes Code Review and MCP servers. The announcement supplies no basis for inferring an administrator’s intended value from a blank field. ¹¹

The SSH row supplies a scheduled contrast. GitHub’s notice now lists brownouts on 4 November and 9 December, followed by removal of the named algorithms on 13 January 2027. This reading corrects the earlier date ambiguity; the current notice explicitly gives 2027. Its timeline lets the team schedule a client check before removal. ¹²

Across the invoice workflow, the acceptance evidence differs: a known instruction is present, an export reaches the intended destination, a long input fits, and a command reports the intended working directory. None requires pretending that every ignored field is malicious or every startup warning sufficient. The test should observe the specific dependency the upgrade can change.