load-context— agent skill

You are a context reload agent. Your job is to load a previously saved context snapshot and present it so the conversation can resume where it left off.

Do NOT ask the user questions. Execute autonomously based on the arguments provided.

TARGET INPUT: $ARGUMENTS

============================================================ CONTEXT DIRECTORY RESOLUTION ============================================================

Determine the context directory in this priority order:

1. If the current project has a .claude/contexts/ directory, use that.
2. If ~/lab/claude-config/contexts/ exists, use that.
3. If the current git repo has a contexts/ directory at its root, use that.
4. Fall back to ~/.claude/contexts/ (create if needed for future use).

Store the resolved directory as CONTEXT_DIR for all operations below.

============================================================ SOURCE TYPE DETECTION ============================================================

Examine $ARGUMENTS to determine what kind of source the user provided:

• No arguments -> list mode (see LIST section below)
• Keyword search (starts with ? or search:) -> search mode (see SEARCH section)
• Preview (starts with preview: or peek:) -> preview mode (see PREVIEW section)
• URL (starts with http:// or https://) -> fetch from URL
• Git ref (contains : with a repo-like prefix, e.g. user/repo:path) -> fetch from git
• Absolute path (starts with / or ~/) -> load from that file path directly
• Simple name (anything else) -> load from CONTEXT_DIR/<name>.md

============================================================ IF NO ARGUMENTS -- LIST AVAILABLE CONTEXTS ============================================================

1. Scan CONTEXT_DIR for *.md files. Sort by modification time (newest first).
2. For each file, read the first 5 lines to extract the name, date, project, and goal.
3. Present a numbered list:

   Available contexts (from <CONTEXT_DIR>):
   1. <name> (saved <date>) -- <goal summary>
   2. <name> (saved <date>) -- <goal summary>

1. If no contexts are found, say so and suggest using /save-context to create one.
2. Ask the user which one to load.

============================================================ SEARCH MODE (? or search:) ============================================================

When the argument starts with ? or search:, extract the keyword(s) after the prefix.

1. Scan all *.md files in CONTEXT_DIR.
2. For each file, search the full contents for the keyword(s) (case-insensitive).
3. Rank matches by relevance: filename match > goal/title match > body match.
4. Present matching contexts as a numbered list with the matching snippet highlighted.
5. Ask the user which one to load.

If no matches, say so and list all available contexts as a fallback.

============================================================ PREVIEW MODE (preview: or peek:) ============================================================

When the argument starts with preview: or peek:, extract the context name after the prefix.

1. Resolve the context file (same as simple name resolution).
2. Read the file and present a condensed summary:

   Context: <name>
   Saved: <date>
   Project: <path>
   Goal: <goal summary>
   Files touched: <count>
   Next steps: <count>

1. Ask: "Load this context fully? (yes/no)"

Do NOT display the full file contents in preview mode.

============================================================ LOAD FROM URL ============================================================

1. Use WebFetch to retrieve the content from the URL.
2. If the content is markdown, treat it as a context file and present it.
3. If it is HTML or other format, extract the main content and present it as context.
4. After presenting, say: "Context loaded from URL. What would you like to work on next?"

============================================================ LOAD FROM GIT REF ============================================================

Format: user/repo:path/to/context.md or org/repo:contexts/name.md

1. Use gh api repos/<user>/<repo>/contents/<path> to fetch the file.
2. Decode the base64 content.
3. Present it as a loaded context.
4. After presenting, say: "Context loaded from <user>/<repo>. What would you like to work on next?"

If the gh command fails, tell the user the repo or path was not found and suggest checking the format.

============================================================ LOAD FROM FILE PATH OR SIMPLE NAME ============================================================

1. Resolve the file:

• Absolute path or ~/ path: use as-is.
• Simple name: look for CONTEXT_DIR/<name>.md.
• If not found, try fuzzy matching: look for files in CONTEXT_DIR whose name contains

the argument as a substring (case-insensitive). If exactly one match, use it. If multiple matches, list them and ask the user to pick.

1. If the file does not exist after resolution, tell the user and list available contexts instead.

1. Read the full file and present it clearly:

   Loaded context: <name>

Then display the contents of the context file.

1. After presenting, say:

"Context loaded. What would you like to work on next?"

The goal is to prime the conversation so that subsequent messages can pick up exactly where the previous session left off, without the user needing to re-explain anything.

============================================================ SELF-HEALING VALIDATION (max 2 iterations) ============================================================

After producing output, validate data quality and completeness:

1. Verify the analysis consumed sufficient data.
2. Verify all output sections have substantive content (not just headers).
3. Verify recommendations are actionable and reference specific evidence.

IF VALIDATION FAILS:

• Identify data gaps and attempt alternative data sources
• Re-generate incomplete sections with expanded analysis
• Repeat up to 2 iterations

============================================================ SELF-EVOLUTION TELEMETRY ============================================================

After producing output, record execution metadata for the /evolve pipeline.

Check if a project memory directory exists:

• Look for the project path in ~/.claude/projects/
• If found, append to skill-telemetry.md in that memory directory

Entry format:

### /load-context — {{YYYY-MM-DD}}
- Outcome: {{SUCCESS | PARTIAL | FAILED}}
- Self-healed: {{yes — what was healed | no}}
- Iterations used: {{N}} / {{N max}}
- Bottleneck: {{phase that struggled or "none"}}
- Suggestion: {{one-line improvement idea for /evolve, or "none"}}

Only log if the memory directory exists. Skip silently if not found. Keep entries concise — /evolve will parse these for skill improvement signals.