fix(exa-agent): reduce SKILL.md size to avoid prompt truncation

[phuongddx]

Problem

skills/exa-agent/SKILL.md is 260 lines / 9.3KB (~2.5K tokens). Agent hosts that inject skills into the main prompt context have per-skill budgets (~2.3K tokens); this file overflows and gets truncated mid-file, silently dropping the end of Batch Script Mode and the entire Failure handling section — exactly the guidance needed when runs misbehave.

Observed in the wild (Codex plugin cache, exa plugin v3.4.1):

⚠ Skill exa:exa-agent exceeded the main prompt context limit and was truncated.

Fix

Condense to 81 lines / 5.2KB (~1.4K tokens) — fits with ~900 tokens of headroom. No behavioral content lost; only redundancy:

Change	Before	After
JSON schema examples	2 full examples (80 lines)	1 minimal example (25 lines)
Batch Script Mode	separate section, duplicated Decision tree branch 1 verbatim	folded into Decision tree branch 1
Continuation patterns	separate section + repeated in workflow + failure handling	folded into Workflow step 4
Exhaustiveness language	separate 27-line section + repeated in workflow	folded into Workflow step 5
Provider list	named in description AND full body list	kept (trigger matching preserved)

Preserved verbatim-in-substance: all 7 Exa Connect providers, request-only-provider warning, decision tree, pre-run checklist, outputSchema rules (maxItems, stable identifiers, formats), ZDR/~750s streaming note, runId/previousRunId continuation rules, input.data/input.exclusion, output.structured/output.grounding, coverage validation, exhaustive-claims guardrails, all four failure-handling modes.

Verification

• Frontmatter parses; trigger keywords intact
• JSON schema example validates (json.loads)
• 31/31 key concepts grep-checked against original
• Diff is single file, 32 insertions / 211 deletions
• ~44% byte reduction, ~69% line reduction

Notes

• Sibling skills/search/SKILL.md (14.3KB) has the same overflow risk — happy to do a follow-up if this approach is accepted.
• Audit report available on request.