Problem
Agent recommendations can still be too generic. A miner should not just hear “pick one narrow change”; Gittensory should explain which repo, lane, work type, or cleanup action is most useful and why.
Desired behavior
Generate repo-specific next actions using contributor history, lane fit, queue pressure, duplicate risk, labels, language/path fit, maintainer-lane context, and prior outcome patterns.
Implementation notes
- Start in
src/services/agent-orchestrator.ts, src/services/decision-pack.ts, and the reward/risk signal builders.
- Use existing stored contributor evidence and repo signal snapshots before adding new storage.
- Keep sorting deterministic.
- Separate maintainer-owned repos from normal contributor/miner evidence.
- Make recommendations concrete: repo, lane, action type, reason, blocker, and rerun condition.
Public/private boundaries
Private MCP/API output can explain scoreability, risk, and blocker impact. Public GitHub output must stay sanitized and avoid score estimates or private reviewability details.
Acceptance criteria
- Agent plan names specific repos and work types to pursue or avoid.
- Maintainer-owned repo history does not inflate normal contributor-lane evidence.
- Cleanup-first actions outrank new work when open PR pressure blocks scoreability.
- Direct-PR, issue-discovery, split, and maintainer lanes produce different reasoning.
- Recommendations are deterministic for the same input data.
Tests
- Direct-PR repo fixture.
- Issue-discovery-only repo fixture.
- Maintainer-lane fixture.
- High open-PR pressure fixture.
- Duplicate-risk fixture.
- Deterministic sorting regression.
Problem
Agent recommendations can still be too generic. A miner should not just hear “pick one narrow change”; Gittensory should explain which repo, lane, work type, or cleanup action is most useful and why.
Desired behavior
Generate repo-specific next actions using contributor history, lane fit, queue pressure, duplicate risk, labels, language/path fit, maintainer-lane context, and prior outcome patterns.
Implementation notes
src/services/agent-orchestrator.ts,src/services/decision-pack.ts, and the reward/risk signal builders.Public/private boundaries
Private MCP/API output can explain scoreability, risk, and blocker impact. Public GitHub output must stay sanitized and avoid score estimates or private reviewability details.
Acceptance criteria
Tests