Task diagnostics and support reports

Koder CLI 0.5.21 adds a diagnostic card for failed or unverified tasks. It shows a report ID, machine-readable codes, a bounded explanation, execution counts and the next step. Unknown causes remain unknown. Report creation does not retry a task or grant permissions.

In chat, use /error to view the most recent diagnostic from that window, /error json for JSON, and /error copy to copy it. These commands do not call a model or send the report. A new chat does not inherit another window's report.

From a system terminal:

koder logs report
koder logs report KR-20261029-012345abcdef --json
koder logs report KR-20261029-012345abcdef --copy

Replace the example ID with the ID on your card. Without an ID, the command reads the latest retained report in the current Koder data directory. This may be from another window. To save a file, redirect --json to koder-error-report.json. koder feedback provides the support address; users choose whether to send it.

Reports distinguish verified operations, returned results, failed calls, unconfirmed effects and calls that were not attempted. They include version, platform, time, codes, bounded counts, hashed task/call identifiers, reported usage, recognizable model metadata and available internal relative source locations. They exclude prompts, arguments, tool output, absolute paths, raw exception messages, raw stacks, email content and credentials. JSON is projected through the same allowlist again when read.

The reports reuse the existing private runtime log and its rotation. Each log generation is bounded at about 4 MiB; current and previous generations are retained. Explicit report lookup reads up to the latest 1,000 reports in those generations. Save important reports before they rotate out. If storage is unavailable, the card says so instead of claiming it was saved.

Local telemetry is on by default. koder setup telemetry off stops automatic report storage for subsequent tasks. The current chat still supports viewing and explicit copying. Existing logs, execution receipts, cancellation, cost limits and permission checks remain intact. Nothing is uploaded automatically.

Clipboard support reuses Windows clip, macOS pbcopy, and Linux wl-copy, xclip or xsel, with bounded waits. An unavailable clipboard produces REPORT_CLIPBOARD_UNAVAILABLE and the report for manual copying. No software is installed automatically.

Common distinctions:

MODEL_INCOMPLETE is a single-response output limit; context limits have separate codes.
MAIL_SMTP_NOT_CONFIGURED concerns sending. Run koder mail doctor --json to inspect receiving and SMTP separately.
MAIL_AUTH_FAILED means authentication was rejected, not that the saved binding disappeared.
TOOL_RESULT_UNCONFIRMED and SHELL_RECONCILIATION_REQUIRED require checking receipts and actual state before another operation.
TASK_DELIVERY_INCOMPLETE is not a verified deliverable. A saved SKIP is not a correct answer.
COMMAND_TYPE_ERROR identifies an internal type exception; an available relative source location helps support locate it.

Background-mail reports are marked background-mail and use existing error deduplication. They do not mark the foreground task as failed. Failure JSON for run and ask gains an additive diagnostic object. Early exceptions also produce a nonempty explanation.

See the Chinese walkthrough for step-by-step instructions. Windows physical checks and the dedicated 163 mailbox remain beta-owned; source and loopback tests do not establish production mailbox or native desktop acceptance.