Una plataforma SaaS que automatice la corrección de bugs en repositorios de código mediante un agente de IA que lea issues, implemente soluciones, ejecute pruebas y gestione PRs automáticamente.

Repository: fullsend-ai/fullsend Author: ralphbean As a **maintainer**, I want a triaged bug (`ready-to-code`) to result in a PR with a fix and passing tests, opened by the bot, without me writing code. ## Triggers * `ready-to-code` label applied * `/code` command ## Acceptance criteria - The code agent reads the issue (including triage output) and implements a fix following repo conventions - The agent reads general triage comments (including any proposed test case) as context — no structured handoff format - The agent runs the repo's test suite and linters locally, iterating until they pass, up to the configured retry limit (default: 2) - If a PR already exists for this issue (e.g., from a previous run or review rework), the agent updates it; otherwise it opens a new PR linking the issue - When CI checks pass on the PR, the issue advances to `ready-for-review` - If CI checks fail, the agent fetches the failure logs, fixes the issue, and pushes again, up to the configured retry limit (default: 2) - If the retry cap is exceeded, the issue is marked `requires-manual-review` with a comment explaining what failed - The code agent cannot push to protected branches or merge PRs - `docs/architecture.md` is updated to reflect any architectural changes that result from delivering this story, and ADRs are filed for decisions that need additional context recorded ## Implementation notes - The agent creates a bot-owned branch for its work — branch protection prevents writes to protected branches - The agent's GitHub token is the broadest of the three roles: contents (read/write), pull requests (read/write), issues (read/write) per ADR 0009 - The 6-hour GitHub Actions job timeout (ADR 0007) constrains how long the implement-test loop can run - The agent discovers repo conventions from the repo itself and CLAUDE.md if present — no additional configuration required - Whether the agent applies labels itself or produces structured output for the entry point to apply is a design decision — MVP can have the agent apply labels directly - Can be developed and tested independently against a test repo before the dispatch entry point (Story 2) is complete ## Related PRs - [#108 — ADR 0009: GitHub App for agent identity and ephemeral credentials (undecided)](https://github.com/fullsend-ai/fullsend/pull/108) - [#106 — ADR 0007: GitHub Actions as initial execution platform](https://github.com/fullsend-ai/fullsend/pull/106) - [#107 — ADR 0008: Reusable workflows for credential isolation (undecided)](https://github.com/fullsend-ai/fullsend/pull/107) - [#104 — ADR 0005: Unidirectional control flow](https://github.com/fullsend-ai/fullsend/pull/104)