fix(tools): say how much a truncated result dropped

[ericleepi314]

What

A result that overran the row's safety cap ended in a bare …, which reads as "a bit more" whether three lines went or three hundred.

before                          after
  ⎿  Error: Exit code 1           ⎿  Error: Exit code 1
     [stderr]                        [stderr]
     1                               1
     …                               …
     10                              10
     …                               … +30 lines

Claude Code counts what it dropped, and so does this codebase's own verbose block (… +N lines omitted); the compact row now matches both.

Deliberately not (ctrl+o to expand): a result the gateway retained no raw copy of expands to the same truncated text, and pointing at a key that changes nothing is worse than silence.

Gallery

Grows the last tools that had never been driven through it — subagent_fork, job_kill, interrupt_agent, update_goal, and a failure noisy enough to hit the cap — completing coverage of all 26 tools the base bundle registers.

It also learns two things about its own driving:

• a slash prompt needs Esc to dismiss the completion menu before Enter submits (the exit_plan_mode scenario was hanging on this)
• scenarios that deliberately park on a prompt — the approval box, the plan review — are skipped by a no-argument sweep, which otherwise sat through a 90 s idle timeout for each

README

Records what the trail actually does now: the per-tool result summaries, the failed-call message, the running clock, delegation and /agents, and that a resumed transcript carries its results.

Test

• npm run typecheck, npm test — 150 files / 1954 passing
• node scripts/verify-boundary.mjs
• text.test.ts: the marker counts, and reads in the singular when one line went
• live: scripts/tool-gallery.py subagent_fork job_kill interrupt_agent update_goal bash_fail_long todo list_agents job_list mixed_run

🤖 Generated with Claude Code