dsh-tool-failure-journal

The hardest thing about a coding agent is not a single failure but the same failure repeating over and over, with each piece of evidence folded, compacted, and buried by the next attempt before anyone can spot the pattern. The session log is the model's own context—it is rewritten, pruned, and rewritten again, so it can never be the place a failure history lives.

dsh-tool-failure-journal keeps a second record outside the context window. Every abnormal tool call—thrown exceptions, schema violations, unknown tool names, pre-dispatch denials, caller cancellations—is synchronously appended as one line to a durable JSONL file on disk, surviving even a process crash mid-turn. Near-identical failures are normalized (paths, timestamps, UUIDs replaced by placeholders) and hashed into stable signatures, so forty repeated errors become a single row with a count and a recurring flag. A tool that succeeded earlier and then fails is tagged as a regression; caller-initured cancellations are kept separate so they never poison the recurrence numbers. For high-frequency failure classes like EDIT_NO_MATCH, COMMAND_NOT_FOUND, and TIMEOUT, each record carries the root cause and a next action that is genuinely different from the one that just failed.

All of this is exposed to the agent through a single failure_journal tool: one stats call turns five identical stack traces into one actionable signal. It is built for any developer running coding agents on DeepSeek Harness who wants failure patterns to become trackable, countable evidence instead of noise lost in the context window.