technical-skill-finder

Technical Skill Finder

Purpose

Find recurring pain points from local agent logs and convert them into actionable skill candidates, reuse opportunities, or existing skill updates.

When to use

• You want to discover missing technical skills from historical agent activity.
• You want reproducible criteria before creating a new skill.
• You want to validate whether an existing skill already covers the pattern.
• You want to include optional personal-signal sources (when authorized).

Inputs

• SCOPE (required): repository paths, workspace, or tool domains to inspect.
• SOURCES (required): ordered source list to mine.
• TIMEFRAME (optional): default all unless constrained by user.
• PRIVACY_POLICY (required): explicit user direction for personal logs.
• TOP_N (optional): number of highest-priority candidates to return.

Workflow

1. Initialize source set
   • ~/.codex/history.jsonl
   • ~/.codex/archived_sessions/*.jsonl
   • ~/.codex/sessions/*.jsonl and ~/.codex/log/* if present
   • Repository-specific telemetry in AGENTS.md/local docs when available
   • Cursor / Codex agent logs detected under known dotfiles directories
2. Normalize extraction signals
   • Parse stack traces and classify failure type (auth, type-check, llm-error, git/ci, runtime, refactor-merge, test)
   • Parse recurring command phrases (rg, mypy, pytest, gh, git, package-manager failures)
   • Record frequency, recency, and affected project context
3. Cluster signals
   • Group by: domain (python/js/rust/docs/tooling), command lineage, and error signature.
   • Deprioritize one-off sessions with low recurrence.
4. Map to existing skills
   • Compare candidate clusters with available skills by name and description.
   • If overlap is high, propose skill update path.
   • If no overlap, propose new skill.
5. Emit ranking output
   • Provide impact, frequency, confidence, skill-fit, and first-apply command set.
6. Produce minimal first-iteration output for high-priority candidates
   • Candidate title + scope
   • Trigger phrase examples
   • Required inputs
   • Suggested workflow summary
   • Evidence snippets (line/file-level)
   • Suggested dependencies/tools (e.g., jq, rg, shell utilities, MCP resources)
   • Return this through chat/stdout by default. Create a persistent artifact root only when the user selects one or another required workflow declares it, with file/byte budgets and source/input identity.
7. Optional extension to personal-signal sources
   • Only after explicit approval to read personal channels.
   • If MCP is available and user has granted access, run MCP resource discovery and include message-signal-derived patterns.
   • Keep this opt-in and isolated from coding-signal output unless user requests a merged plan.

Guardrails

• Never infer or emit private content from message logs unless explicitly permitted.
• Skip binary/corrupt files and summarize only parseable text sources.
• Prefer deterministic commands and small scripts over ad-hoc manual parsing.
• Always avoid proposing skills with unresolved operational context (credentials, environment, private URLs).
• If evidence is ambiguous, return confidence: low and request one more session sample.
• Reuse one canonical identity-matched inventory instead of materializing duplicate large extracts. Apply the $operations-worktree task artifact contract when retention or resumable phase state is required.

Outputs

• skill_candidates.md-style report in chat:
  • reuse candidates (existing skill can be extended)
  • new skill candidates (not yet covered)
  • top source anchors with references
  • recommended next action (create/update)

Read references/sources.md for source precedence. Read references/scorecard.md for prioritization rules.

Flow

stateDiagram-v2
    [*] --> SelectAuthorizedSources
    SelectAuthorizedSources --> NormalizeAndCluster
    NormalizeAndCluster --> CompareExistingSkills
    CompareExistingSkills --> ProposeUpdate: substantial overlap
    CompareExistingSkills --> ProposeNewSkill: no existing coverage
    CompareExistingSkills --> ReportUncertainty: insufficient evidence
    ProposeUpdate --> RankAndReport
    ProposeNewSkill --> RankAndReport
    RankAndReport --> [*]
    ReportUncertainty --> [*]
    note right of SelectAuthorizedSources
        Personal channels require explicit approval.
        Return findings inline unless retention is selected.
    end note