Add revision and body excerpts to agent-mode search and list rows

[naveenspark]

Agents reading kata search --agent now get enough context to choose a result without running kata show on every hit. Each search row appends owner, priority, revision, and an excerpt from the issue body. kata list --agent rows append revision, and kata show --agent prints a Revision: line.

Before this change, a search row carried only the id, score, status, matched fields, and title. An agent had to open each candidate to learn who owned it, how urgent it was, or what the body said. No agent-mode read exposed the revision that --if-match writes need, so agents had to fall back to --json for it.

The excerpt is at most 160 characters. It takes up to 24 words of the body around the first word that matches a query term, with … marking trimmed ends. If that window is still over budget, it is re-centered on the matched term so long lead-in text cannot push the match out of view. Term matching folds case and diacritics and splits on the same letter/number boundaries the lexical search backends use, so cafe finds café, foo-bar finds foo bar, and log does not match inside catalog. When no term matches, such as a title-only or semantic hit, the excerpt is the start of the body. Empty bodies emit no excerpt= field.

All new fields are appended after the existing ones, so agent_format stays 1 and current parsers keep working. docs/reference/agent-output.md documents the new row shapes.

[roborev-ci]

roborev: Combined Review (87fd0cf)

Verdict: Code is clean with no Medium-or-higher severity findings.

No Critical, High, or Medium issues were identified.

---

Reviewers: 2 done | Synthesis: codex, 6s | Total: 3m16s

[roborev-ci]

roborev: Combined Review (e145ca8)

Verdict: High-severity prompt-injection risk identified; no Medium or Critical findings reported.

High

• cmd/kata/search.go:205-215 — kata search --agent automatically includes attacker-controlled issue-body text in coding-agent output. A lower-privileged issue author could inject instructions that an agent interprets as commands, potentially causing privileged mutations or repository actions. agentValue prevents formatting/terminal injection but not semantic prompt injection. Keep body content out of automatic agent output or expose it through an explicitly untrusted data channel.

---

Reviewers: 2 done | Synthesis: codex, 6s | Total: 3m52s

[wesm]

looking

[roborev-ci]

roborev: Combined Review (6003fd7)

No issues found.

---

Reviewers: 2 done | Synthesis: codex | Total: 6m5s