lex-assessor

Fleet pipeline intake for LegionIO. Receives normalized work items from source absorbers, deduplicates them atomically, classifies via LLM structured output, and applies configuration templates before routing downstream.

lex-assessor is the first stage of the Fleet Pipeline — the autonomous coding system that takes GitHub issues and pull requests and routes them through planning, implementation, validation, and feedback loops.

When a work item arrives, the assessor:

1. Deduplicates — computes a SHA256 fingerprint of source:source_ref:title and claims it atomically via Redis SETNX. Duplicate submissions are rejected immediately.
2. Classifies — calls an LLM to assess priority, complexity, work type, primary language, and estimated difficulty (0.0–1.0).
3. Templates — selects a config preset based on the complexity classification and deep-merges any source-level overrides on top.
4. Routes — returns the fully assessed work item ready for the next pipeline stage.

When a work item exceeds its maximum iteration count, escalate handles the cleanup: labels the GitHub issue, posts a summary comment, submits to the human approval queue, releases the dedup lock, and clears all Redis refs.

Unknown complexity falls back to moderate_feature.

Legion::Extensions::Assessor::Runners::Assessor.assess(work_item: work_item)
# Success:
# => { success: true, work_item: { ..., pipeline: { stage: 'assessed', trace: [...] } } }
#
# Duplicate:
# => { success: false, reason: :duplicate, work_item_id: "uuid" }

All runner methods accept the standard Legion extraction pattern:

assess(results: nil, work_item: nil, args: nil, **)

work_item can be passed directly or extracted from results[:work_item] or args[:work_item] — whichever is non-nil.

Legion::Extensions::Assessor::Runners::Assessor.escalate(work_item: work_item)
# => { success: true, work_item: { ..., pipeline: { stage: 'escalated' } } }

Escalation performs these steps in order:

1. Labels the GitHub issue fleet:escalated (no-op if lex-github not loaded)
2. Posts an escalation summary comment (no-op if lex-github not loaded)
3. Submits to ApprovalQueue with resume routing key lex.developer.runners.developer.incorporate_feedback (no-op if lex-audit not loaded)
4. Releases the dedup fingerprint so the work item can re-enter the pipeline
5. Clears fleet:payload, fleet:context, and fleet:worktree Redis keys

fp = Legion::Extensions::Assessor::Helpers::Dedup.fingerprint(
  source: 'github', source_ref: 'LegionIO/lex-exec#42', title: 'Fix timeout'
)
# => "a3f4b2c1..." (SHA256 hex, 64 chars)

Dedup.claim!(fingerprint: fp)        # => true (first claim) / false (duplicate)
Dedup.duplicate?(fingerprint: fp)    # => true / false
Dedup.release!(fingerprint: fp)      # removes the lock

Default TTL is Legion::Settings.dig(:fleet, :cache, :dedup_ttl_seconds) || 86_400 (24 hours).

result = Legion::Extensions::Assessor::Helpers::Classifier.classify(work_item: work_item)
# => {
#      priority: :high,
#      complexity: :simple_bug,
#      work_type: :bug_fix,
#      language: :ruby,
#      estimated_difficulty: 0.3
#    }

capability = Classifier.difficulty_to_capability(0.7)  # => :reasoning

template = Legion::Extensions::Assessor::Helpers::ConfigTemplates.for_classification(
  complexity: :moderate_feature
)

merged = ConfigTemplates.merge_with_overrides(
  template: template,
  overrides: { implementation: { solvers: 2 } }
)

All configurable values read from Legion::Settings.dig(:fleet, ...):

Add to your Gemfile:

gem 'lex-assessor'

bundle install
bundle exec rspec      # 43 examples, 0 failures
bundle exec rubocop    # 0 offenses

MIT