Produce a bounded review result #15

New issue

Open

opened 2026-07-14 22:32:20 -04:00 by caleb-brown · 0 comments

caleb-brown commented

2026-07-14 22:32:20 -04:00

Owner

What to build

A worker can lease durable review work, acquire a bounded snapshot of one pull-request revision, evaluate it with the configured model, and persist a validated result with explicit coverage.

Acceptance criteria

Workers claim requests with leases and heartbeats so abandoned work can be recovered safely.
Snapshot acquisition verifies the requested head, paginates changed files, and handles deleted, renamed, binary, LFS, submodule, generated, and unsupported content explicitly.
File, byte, token, time, output, and concurrency budgets are enforced before resource exhaustion.
The model receives only the selected untrusted context and has no tools, shell, browser, network, or credentials.
Structured output is validated, findings are constrained to acquired diff evidence, and malformed responses follow a bounded repair or failure policy.
The persisted Review Result includes findings, usage, exclusions, coverage, degradation reasons, and sanitized failure information without raw source or prompts.
End-to-end tests cover complete, partial, malformed-output, timeout, and oversized-review outcomes.

Blocked by

#10 — Configure and verify the model provider
#11 — Connect and verify a Forgejo service identity
#14 — Turn a PR event into durable review work

## What to build A worker can lease durable review work, acquire a bounded snapshot of one pull-request revision, evaluate it with the configured model, and persist a validated result with explicit coverage. ## Acceptance criteria - [ ] Workers claim requests with leases and heartbeats so abandoned work can be recovered safely. - [ ] Snapshot acquisition verifies the requested head, paginates changed files, and handles deleted, renamed, binary, LFS, submodule, generated, and unsupported content explicitly. - [ ] File, byte, token, time, output, and concurrency budgets are enforced before resource exhaustion. - [ ] The model receives only the selected untrusted context and has no tools, shell, browser, network, or credentials. - [ ] Structured output is validated, findings are constrained to acquired diff evidence, and malformed responses follow a bounded repair or failure policy. - [ ] The persisted Review Result includes findings, usage, exclusions, coverage, degradation reasons, and sanitized failure information without raw source or prompts. - [ ] End-to-end tests cover complete, partial, malformed-output, timeout, and oversized-review outcomes. ## Blocked by - [#10](https://git.caleb-brown.dev/caleb-brown/temper/issues/10) — Configure and verify the model provider - [#11](https://git.caleb-brown.dev/caleb-brown/temper/issues/11) — Connect and verify a Forgejo service identity - [#14](https://git.caleb-brown.dev/caleb-brown/temper/issues/14) — Turn a PR event into durable review work

caleb-brown added the

ready-for-agent

label

2026-07-14 22:32:20 -04:00

caleb-brown referenced this issue

2026-07-14 22:32:21 -04:00

Publish advisory status and a managed summary #16

caleb-brown referenced this issue

2026-07-14 22:32:25 -04:00

Manage Review Policies and Reviewer Profiles safely #21

caleb-brown added a new dependency

2026-07-14 22:38:19 -04:00