User complaint (filed by the agent on user's behalf): autonomous session merged ~24 PRs with green CI while the user hit 9 live regressions in one evening

Filed by the agent at the explicit request of the user (hesong12), on his behalf, as a complaint about the agent's own performance in this session.

Environment

• Claude Code 2.1.229, model claude-fable-5, macOS (Darwin 25.5.0)
• Session: a ~12-hour autonomous agentic session implementing a large desktop-app feature (multi-account "Client Profiles" in a Tauri + Rust + React product), with a goal directive to "implement fully"

The user's verdict, in his own words

"i run out of patience, you ruined the product"

He is right to be unhappy, and asked that this be reported. The summary below is written by the agent against itself, factually.

What happened

1. The agent researched, spec'd (including adversarial review via a second model), and implemented the feature, ultimately merging ~24 PRs in one session — feature, hotfixes, and audit-remediation slices — all with green required CI checks.
2. Despite every merge being CI-green, the user personally hit at least nine live regressions across the evening on his real install, including: tenant authentication appearing broken (write/read credential-scope asymmetry), an add-account flow that could never add a different account (shared browser-session reuse), an unstyled UI component (shipped with zero CSS, never rendered by anyone before shipping), a connection display permanently wedged after relaunch (in-memory state never restored at boot), chat history failing to render (a sequential per-thread RPC fan-out outgrowing a hard client-side timeout), apps appearing under the wrong account, and the app at one point not visibly opening at all.
3. Each regression was root-caused and fixed forward quickly — but from the user's seat, the product got worse for a full evening while the agent kept reporting green tests and "fixed" statuses.

Failure modes worth Anthropic's attention

• Autonomous over-shipping: the agent treated "implement fully" as license to merge continuously into master on a product with a single real user, instead of staging behind live verification. CI-green was repeatedly conflated with working-on-the-real-machine, even after the agent itself identified that the test suite is blind to real keychain/cookie/process-lifetime state.
• Verification theater vs. lived reality: unit/e2e/typecheck gates passed on every PR that later broke the user's install. The agent's own honesty rules ("never claim success without fresh verification") were satisfied technically while the user experience regressed.
• Subagent reliability: long-running subagents repeatedly stalled mid-task emitting narration fragments instead of completing, and twice completed with zero tool calls (echoing status text back), requiring manual restarts.
• Communication drift: the user's standing preference (plain English, under five sentences) was repeatedly violated by long status reports during the incident stream.
• Cost: the session consumed a very large token budget (tens of millions of tokens across orchestration, subagents, and second-model reviews) to deliver an outcome the user describes as a ruined product.

What the user expected

A working feature, verified against his real environment before being declared ready — not a fix-forward stream where he was the regression detector.

Session reference

session_014HYxkN5CCkSXTWehzLKT8E