[FEATURE]: Remove or truncate <location> path in system prompt available_skills — wastes ~3.3K tokens/turn

[ghostbody]

Description

The system prompt's <available_skills> block includes a <location> field with the full absolute file path for every skill. This field is:

1. Unused by the LLM — skills are loaded via skill(name="foo"), not by path. The tool's internal registry already maps name → location during discovery, so the LLM never needs to know the path.
2. Extremely verbose — macOS paths like <location>/Users/yejq/Library/Application Support/com.xxx.zzz/mmmm/.opencode-config/skills/art-requirement/SKILL.md</location> average ~136 characters each.
3. Made worse by pathToFileURL — spaces become %20, # becomes %23, and a file:// prefix adds 7 chars on top (as documented in fix(skill): skill directory context missing or URL-encoded in system prompt and slash-command paths #33786).

Token waste measurement (real-world data)

Measured from an actual system prompt dump with 84 skills:

Total <location> blocks:        84
Total characters in locations:  11,382
Average per location:           ~136 chars
Estimated tokens wasted:        ~3,252 tokens per turn

3.2K tokens burned on every API call on paths the LLM never reads, never uses, and doesn't need.

Single sample entry:

<location>/Users/yejq/Library/Application Support/com.xxx.zzz/mmmm/.opencode-config/skills/art-requirement/SKILL.md</location>

= 141 characters, for a skill whose name (art-requirement) is the only thing the LLM needs to invoke it.

Why <location> is unnecessary

The skill loading flow:

1. System prompt tells the LLM: "these skills exist, here are their names + descriptions"
2. LLM decides to use a skill → calls skill(name="ppt-master")
3. The tool internally maps name → location (already known from discovery)
4. The tool reads SKILL.md and returns the content + a base directory hint

At no point does the LLM need the file path. The tool's internal registry already knows where every skill lives. The <location> field in the system prompt is pure context pollution.

For skills that reference relative resources (scripts/, references/), the base directory is already provided in the tool output when the skill is loaded — again, no need for the LLM to pre-know it.

Where the waste happens

In packages/opencode/src/tool/skill.ts (as documented in #13188):

...accessibleSkills.flatMap((skill) => [
  `  <skill>`,
  `    <name>${skill.name}</name>`,
  `    <description>${skill.description}</description>`,
  `    <location>${pathToFileURL(skill.location).href}</location>`,  // ← waste
  `  </skill>`,
])

And in the system prompt via Skill.fmt(list, { verbose: true }) (as documented in #22236):

Injection point	Format	Includes <location>?
System prompt body (verbose: true)	XML	Yes
Skill tool description (verbose: false)	Compact markdown	No

The verbose: false path already excludes <location>, proving it's not essential for the tool description. The system prompt just inherited it.

Proposed solutions

Option A (minimal, safe): Strip <location> from the system prompt only

• Keep the verbose: true XML format but omit the <location> tag
• The tool description already omits it (verbose: false)
• The full location is still returned in the tool output when a skill is actually loaded (where it's genuinely useful for resolving relative resource paths)

Option B (configurable): Add an includeLocation flag to Skill.fmt()

Skill.fmt(list, { verbose: true, includeLocation: false })

Option C (root cause): Use the skill name as the identifier instead of the full path

<location>art-requirement</location>

This preserves the field for any consumer that depends on its presence, without the token bloat.

Related issues

• [FEATURE]: Lazy-load agent/skill lists instead of inlining into tool descriptions #13188 — Lazy-load agent/skill lists (identifies location as part of the waste but focuses on broader lazy-loading)
• Skills descriptions bloat system prompt (~600KB / ~150K tokens), breaking Bedrock and small-context models #20647 — 600KB system prompt bloat from skill descriptions (includes <location> in the bloat total)
• Skill list injected twice per request wastes ~5K–7K tokens of context budget #22236 — Skill list injected twice per request (notes verbose: true includes <location>)
• fix(skill): skill directory context missing or URL-encoded in system prompt and slash-command paths #33786 — Location URL-encoding bug via pathToFileURL (confirms the encoding problem)
• Skill descriptions silently dropped from system prompt above some cumulative-size threshold anthropics/claude-code#59921 — Claude Code silently drops descriptions above a threshold (downstream symptom of the same budget pressure)

Impact

For users with many skills (common in plugin-heavy setups), this is pure waste — 3K+ tokens that could be used for actual conversation context, tool results, or just cost savings. The fix is effectively a one-line change with zero functional impact, since the skill tool resolves name → location internally regardless of what's in the system prompt.