fix(warp): skip synthetic transcript entries when picking the notification title

[bobrnor]

Problem

on-stop.sh (and, since #73, on-stop-failure.sh) titles the Warp notification with "the user's last prompt", scraped from the session transcript. The filter's only exclusion is tool-result entries — the comment at on-stop.sh:33-38 states the assumption that every other type: "user" entry is human-authored.

That assumption doesn't hold. Claude Code writes a number of its own injections into the transcript as user-type entries with ordinary string content, so they sail through the filter and become the notification title:

<bash-stdout>arn:aws:ecs:us-east-1:123456789012:task/prod-cluster/a4d9c84a37984f9d8f48b62f…
Task completed.

The most visible source is local ! bash mode — running !aws ecs list-tasks … in Claude Code records <bash-input>…</bash-input> and <bash-stdout>…</bash-stdout><bash-stderr></bash-stderr> as two plain-string user entries. If the turn ends after that, the ARN dump is the title.

Sampling 25 of my own recent transcripts, these user-type entries are all currently eligible to become a title:

entry	count
<task-notification>	739
<local-command-stdout>	19
<command-name>	19
<system-reminder>	11
<bash-stdout> / <bash-input>	4 / 4

Root cause

on-stop.sh:39-51 and on-stop-failure.sh:28-40 (identical copies of the same jq filter) select the last type == "user" entry whose content is a plain string or an array containing a text block. Tool results are arrays of tool_result blocks and drop out; every synthetic injection with text content does not.

Fix

Move the transcript scraping into a sourceable scripts/extract-transcript.sh — which also removes the duplicated filter on-stop-failure.sh had to carry — and skip entries positively recognized as Claude Code internals, falling back to the previous real prompt.

Following the approach argued in #72:

• Anchor on the ASCII envelope tags, not the surrounding English framing. <teammate-message> / <cross-session-message> arrive behind a "Another Claude session sent a message:" line that a localized build may translate, so those two are matched anywhere in the entry rather than as a prefix.
• Only rewrite what is confidently recognized. No generic XML stripping and no rejection of odd-looking input — an unusual title beats a missing notification.
• English-only markers ([Request interrupted by user…], Stop hook feedback:) are best-effort: a localized build falls through to today's behaviour rather than breaking.

When every candidate is filtered out, QUERY is empty and Warp falls back to its own "<agent> completed" title, which is still an improvement on raw XML.

legacy/on-stop.sh is untouched — it uses the first user entry, which is the human's opening prompt, so it never hit this.

Tests

18 cases added to tests/test-hooks.sh covering each synthetic form, plus the cases that guard against over-filtering: a human prompt that merely mentions a tag mid-sentence is kept, tool results are still skipped, text blocks are still joined, and degenerate transcripts (all-synthetic, no user entries, missing path) return empty rather than erroring.

=== Results: 74 passed, 0 failed ===

shellcheck -S warning is clean on all three touched scripts.

Verified end-to-end against the real transcript that produced the screenshot above — replaying it through the patched hook, query goes from the <bash-stdout>arn:aws:ecs:… dump to the user's actual prompt.

Relationship to existing issues and PRs

• Supersedes fix(warp): skip synthetic system messages when picking last user prompt #59, which fixes the same class but whose tag list omits bash-*; I replayed my reproduction through that PR's filter and <bash-stdout> still comes through. This PR keeps every tag fix(warp): skip synthetic system messages when picking last user prompt #59 covers, adds the missing ones, extends the fix to on-stop-failure.sh (which didn't exist when fix(warp): skip synthetic system messages when picking last user prompt #59 was opened), and adds test coverage. Happy to close this in favour of fix(warp): skip synthetic system messages when picking last user prompt #59 with the extra tags folded in, if the author prefers.
• Fixes Stop hook notifies on every sub-agent completion (<task-notification> re-invocations), spamming a banner per sub-agent #67 (<task-notification> from background subagents) and Stop hook surfaces raw <teammate-message> XML as the notification title when an Agent Teams message is the last user-type transcript entry #72 (<teammate-message> envelopes).
• Does not touch on-stop.sh: byte-level truncation splits multibyte UTF-8, Stop hook fails Claude Code JSON validation #70 — the byte-level ${QUERY:0:197} truncation that splits multibyte UTF-8 is in the same block and worth fixing, but it's an independent bug; happy to send it separately.