Human Interaction Points
Human interaction points
The pipeline pauses and returns control to the user at the following points. Points marked blocking require a response before the pipeline can continue; points marked informational present output with no further input needed.
Input Validation
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 1 | mcp__forge-state__validate_input returns an error (empty, too short, malformed URL) | Error messages; pipeline stops | Yes — pipeline aborts |
| 2 | LLM judges input as gibberish or unrelated to software development | Rejection message with specific reason and valid-input examples; pipeline stops | Yes — pipeline aborts |
| 3 | Jira URL provided but mcp__atlassian__getJiraIssue tool unavailable | Error with plugin install instructions; pipeline stops | Yes — pipeline aborts |
Workspace Setup
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 4 | Current git branch is not main/master | Branch name shown; choice to use the current branch or create a new one | Yes — waits for choice |
| 5 | Effort level selection (always required) | User selects effort level (S / M / L) and sees which phases will execute for that choice | Yes — waits for selection |
| 6 | full template and --auto flag used together | Warning that full mandates manual checkpoints; asked to continue without auto-approve or abort | Yes — waits for choice |
Checkpoint A — Design Review
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 7 | Auto-approve conditions met (--auto + AI verdict APPROVE or APPROVE_WITH_NOTES, no CRITICAL findings) | One-line notice: "Auto-approving Checkpoint A (AI verdict: …)" | No — informational |
| 8 | Human approval required (AI returned REVISE, or no --auto, or full template) | Design summary: approach, key changes, risk level, AI verdict, any MINOR findings, workspace path. Asked to approve or give feedback. Sound notification plays. After each revision cycle the updated design is re-presented and the pipeline stops again | Yes — STOP AND WAIT |
Checkpoint B — Tasks Review
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 9 | Auto-approve conditions met | One-line notice: "Auto-approving Checkpoint B (AI verdict: …)" | No — informational |
| 10 | Human approval required | Task overview: task count, risk level, AI verdict, any MINOR findings, workspace path. Asked to approve or give feedback. Sound notification plays. After each revision cycle the updated task list is re-presented and the pipeline stops again | Yes — STOP AND WAIT |
Implementation (Phase 5–6 loop)
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 11 | A task's impl-reviewer returns FAIL and the per-task retry limit (2) is exhausted | Failure report for that task; asked how to proceed | Yes — waits for instruction |
| 12 | A subagent returns empty or incoherent output and the single retry also fails | Failure reported; phase-fail recorded in state | Yes — pipeline stalls until user intervenes |
| 13 | Test suite fails after implementation completes | Failure output presented; phase-fail recorded in state | Yes — pipeline stalls |
Final Verification
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 14 | Verifier finds failures it cannot fix | Failure report presented to user | Yes — pipeline stalls |
Pipeline End
| # | Trigger | What the user sees | Blocking |
|---|---|---|---|
| 15 | summary.md written successfully | Full contents of summary.md displayed (request, branch, PR, task table, improvement report, execution stats). Sound notification plays. | No — informational |
Skipped checkpoints: Checkpoint B is skipped for effort S and M (only effort L runs Checkpoint B). Phase 4b (task reviewer) is also skipped for effort S and M. Use
--autoto allow the AI reviewer verdict to auto-approve Checkpoint A (not available withfulltemplate).