Repository navigation
Cursor setup produces a Lab that can't run intake; no supported way to bind a planning agent #388
Description
Problem
A new user who picks Cursor during setup gets a Lab that can never start an Issue.
Reproduced on v0.4.2 against DanielNetzer/forgecell-trial:
forgecell init --harness cursorin a fresh Forgecell home. Doctor reports Cursor ready and signed in, and the bootstrap proposal is saved with onlyharness: {binding: cursor}.forgecell init --approve <assay>: approved, with no warning.forgecell run 3: the Molecule blocks at intake with
analysis failed: Cursor ticket analysis is unavailable: action-capable MCP/plugin isolation has not been verified; coding binding is preserved.
The Cursor adapter can't do read-only ticket analysis, by design. The documented fix is a per-role binding (harnesses.analysis: {binding: claude-code, ...}, docs/ticket-readiness.md). But no supported path creates one. lab/README.md says "Init never creates an override", the UI has no control for it, and the post-approval init --harness cursor refuses and says to declare harnesses.analysis first. The only way is editing the Formula YAML by hand.
Meanwhile docs/install.md (First run, step 3.2) tells users to "pick claude-code, codex or cursor", and the setup UI offers Cursor as a ready choice.
Change
- Bootstrap: when the chosen coding harness can't do analysis, propose a Formula that declares
harnesses.analysison a ready analysis-capable harness (claude-code or codex). If none is ready, refuse with the adapter's reason and the next step. Never save a Formula that can't run intake. - Setup UI: on the Coding agent step, say Cursor codes only and needs a second agent for planning, and let the user pick that agent. The Formula review step shows both bindings.
- CLI: a supported way to add or change a role override, such as
init --harness ID --role analysis|coding|meta, using the same reviewed-proposal flow as--model --role. - Docs: in install.md step 3.2 and the README, say Cursor needs a planning agent.
Done when
- A fresh Lab set up with Cursor, from the CLI or the UI, runs intake on Claude Code or Codex and coding on Cursor, end to end to a draft PR.
- No supported setup path produces an approved Formula whose intake can't run.
- Tests cover the bootstrap refusal and proposal, plus the new role flag.
Activity
DanielNetzer commentedon Nov 6, 2026
Two more things from the v0.4.3 Cursor retest, where planning was bound to Claude Code by hand with
harnesses.analysis:init --labapproved a Formula withharnesses.analysis: claude-code, the summary still said "GitHub Issue workflow: unsupported". It judges only the default harness, not the per-role bindings, so it should report the workflow as supported when the analysis role is bound to a harness that can plan.lab/internal/cli/cli.go:1381-1383). The label should name the bound coding harness, or say nothing for harnesses that have no allowlist.The Cursor coding failure from the same run is filed separately as #391.
DanielNetzer commentedon Nov 8, 2026
Split into three slices, in order: 413 (CLI role override), 414 (bootstrap pairing + per-role doctor/plan wording + docs), 415 (setup UI). This Issue closes when all three merge.
Verified Nov 8 on v0.4.7, with hand-added overrides: Codex plans + Cursor codes (forgecell-trial#8), Codex plans + Claude Code codes (forgecell-trial#7), Claude Code plans + Cursor codes (forgecell-trial#5).
DanielNetzer commentedon Nov 8, 2026
Done across #413 (#416), #414 (#417) and #415 (#419): a Cursor setup from the CLI or the UI now declares a planning harness (Claude Code, then Codex, or the one you pick), refuses when none is ready, and
init --harness ID --role ROLEchanges a role later. Ships in v0.4.8.