contract.yaml

version: 1
workflow: duplicate-issue-detector
framework: langgraph

inputs:
  - GitHub issue number (passed at invocation time)
  - Issue title + body fetched from GitHub API at runtime

outputs:
  - One comment on the triggering issue, only if similarity >= threshold

permissions:
  - github: issues:read+write   # comment only — never close, label, or edit
  - openai: embeddings:read
  - qdrant: read+write on one collection

side_effects:
  - Posts one comment on the triggering issue, if similarity >= threshold
  - Writes one vector to the Qdrant collection (always, regardless of duplicate result)

approval_points: []   # intentionally empty — this workflow only ever comments,
                       # never closes or merges, so no approval gate is required

recovery_strategy: >
  Any failure (GitHub unreachable, OpenAI error, Qdrant error) raises an
  exception that propagates to the caller. Nothing is silently skipped.
  The run fails loudly and visibly.

replay_semantics: >
  Idempotent on Qdrant — re-running on the same issue number overwrites
  the existing vector (issue number is the Qdrant point ID).
  The GitHub comment is NOT idempotent — running twice on the same issue
  will post two comments if a duplicate is found both times.

dependencies:
  - GitHub API
  - OpenAI Embeddings API (text-embedding-3-small)
  - Qdrant instance (local or cloud)

state: >
  Persisted in a Qdrant collection — one 1536-dimensional vector per
  previously-seen issue, with issue number, title, and URL as payload.

observability:
  - Structured INFO/ERROR logs to stderr during execution
  - JSON summary to stdout on completion (issue number, scores, flags, run_log)