[BUG] Destructive and cost-affecting defaults are undisclosed: the metered user is the discovery mechanism for the vendor’s own behaviour

[pmcnary]

Preflight Checklist

• I have searched existing issues and this hasn't been reported yet
• This is a single bug report
• I am using the latest version of Claude Code

What's Wrong?

Several Claude Code behaviours that materially affect data retention and cost are not disclosed at install, not surfaced in /config, and not documented where a user would look. The only way to find them is to investigate — and on a token-metered product, that investigation is billed to the user.

This issue is about the disclosure gap itself, not any one of the behaviours. Each is filed separately; the pattern is what needs an owner.

The undisclosed behaviours, and how each had to be discovered

Behaviour	How a user finds out	Filed as
Transcripts are deleted after 30 days by default	Notice history is gone, or audit the store against history.jsonl	#62476, #59248, #79758
cleanupPeriodDays may be ignored even when set	Compare git-tracked settings against surviving sessions	#41458 (data-loss, regression)
/bug opens a public GitHub form and attaches no transcript	Extract the URL from the CLI binary	filed alongside this
Bash timeout sometimes hard-kills (exit 143) with no partial output, sometimes auto-backgrounds; criteria undocumented	Lose a long command and diff the two behaviours	#79879, #48861, #77645
Session history is re-materialised up to ~11× by compaction	Parse the transcript store and count duplicate uuids	filed alongside this
Server-gated system-prompt sections change model behaviour, enabled by capability rather than user config, with no opt-out	Notice the model stop honouring your own configuration	#80988
Default reasoning effort was reduced 2026-03-04 → 2026-04-07	A postmortem published after the window closed	—

Why the disclosure gap is the bug

Taken individually each of these is a papercut with a workaround. Taken together they describe a product where the user is the discovery mechanism for the vendor's own defaults, and pays per token for the privilege.

Concretely, in my case: establishing the retention behaviour, the write amplification, and the /bug destination required a full-corpus audit — 3,146 transcript files, 5.39 GB, multi-agent analysis costing millions of tokens, plus a working day of back-and-forth. Every one of those facts is something the vendor already knows and could have stated in one line of documentation or one first-run prompt.

The retention case is the sharpest. On this machine, history.jsonl references 13 sessions and 11 no longer exist on disk — nine from 2026-03-19 → 2026-04-08, one of them 117 prompts long. That data was destroyed by a default I was never shown, and I discovered it only because I happened to audit the store for an unrelated reason, one day before eight more sessions crossed the same threshold.

There is also no in-product path to report any of this privately, and no route to billing or support, so the cost of discovery cannot be recovered or even raised.

Expected Behavior

1. First-run disclosure of destructive defaults. Retention in particular: state it, and surface cleanupPeriodDays in /config where it can be seen and changed.
2. A behavioural changelog. Not version numbers — the actual changes to defaults, system prompts, reasoning effort and tool behaviour, published when they ship, not in a postmortem afterwards.
3. Document the undocumented tool behaviour already identified in the issues above, rather than leaving users to reverse-engineer it from failures.
4. A private support/report path with a route to billing, so a user who has spent real money discovering vendor defaults can raise it somewhere other than a public issue tracker.

Environment

• Claude Code CLI 2.1.220
• VS Code extension 2.1.220 (auto-updated; 19 client versions in 5 weeks, no pinning)
• Linux (Debian)
• 67 sessions / 37 days of continuous use analysed

[pmcnary]

Sharpening the timeline, because it changes what this issue is about.

Anthropic's published postmortem covers 2026-03-04 → 2026-04-20 — default reasoning effort reduced high→medium, a 2026-03-26 bug clearing older thinking each turn that made the model "seem forgetful and repetitive", and a system-prompt verbosity cap. Resolved in v2.1.116 on 04-20, with usage limits reset as compensation on 04-23. (Dates as published; correct me if they're off — you hold the record.)

Every measurement in this report is from 2026-06-22 → 2026-07-29. Entirely after that window closed.

Measured, all post-fix
Output tokens	23.9M across 125 hours, 65 sessions
Sessions that open on context recovery	53 of 65 — carrying 87% of all output tokens
Distinct compaction events	193 across 28 sessions
Transcript write amplification	5.71× corpus-wide, 10.89× in heavy sessions
Distinct crash events	19 in nine days

So this is not a late report of the March regression. It is a second degradation, in the same symptom classes, beginning after a postmortem was published and closed — and this time with no acknowledgement, no postmortem, and no changelog covering the period.

That is the disclosure gap this issue is about, stated more precisely than the original text: the problem is not only that defaults are undisclosed at install. It is that a regression was disclosed once, retrospectively, and the pattern then repeated without disclosure. A user who experienced both has no way to distinguish "fixed" from "fixed, then broken again differently" — except by measuring it themselves, at their own cost, which is what produced the numbers above.

Ask, unchanged but sharper: a behavioural changelog covering system-prompt, effort, and default changes, published when they ship. The March postmortem proves you can write one. The June–July silence is the gap.