Intent assistant validates proposals only with the parser - extend the repair loop through a dry-run generation

[delchev]

Problem

The intent assistant validates a proposal with the parser and nothing else: IntentAgentService.validationIssues runs IntentParser.parse and returns. Everything the parser accepts reaches the developer unexamined — including models that generate nothing for an authored requirement, or generate references the generation pass itself would refuse. The repair loop (ProposalRepairLoop) can only ever correct parse errors, so a proposal that "parsed fine, generated wrong" is handed over looking finished.

The generation layer already collects exactly the right issues: IntentGenerationContext.addIssue / getIssues is where reportDroppedGlue, the cross-model resolution failures and the lifecycle warnings land. What is missing is a way to run generation without writing.

Fix

1. A dryRun mode on the generation pass. A flag on IntentGenerationContext guarding every write/claim surface — writeModelFile, writeModelFileIfAbsent, keepExistingModelFile — plus the stale-output scrub, which lives in IntentGenerationService (the flag must be consulted there too, not only on the context). Reads stay live: cross-model resolution needs the repository.
2. A dry-run entry point on IntentGenerationService returning the collected issues. Note the current generate(...) signature requires projectRoot / projectName / workspaceName — and AgentRequest is {yaml, message, history}, carrying no project coordinates — so the dry-run entry point needs a synthetic-context shape, not just a flag.
3. validationIssues runs parse, then the dry run, and concatenates — the repair loop then sees the generation-layer issues too.
4. Wire the Builder shell's gate. The Builder's client-side re-validation of an agent proposal calls POST /services/ide/intent/parse — parse-only, the same blind spot. It should consume the extended check (a dry-run parameter on the parse endpoint, or a dedicated endpoint), so its auto-apply gate stops accepting generation-invalid models.
5. Raise MAX_REPAIR_ROUNDS from 2 to 4 — one constant, worthless alone, sensible once the validator can see more than a parse error. Land it after the upstream call is streamed (Intent AI client: stream the upstream call, send adaptive thinking, raise the max-tokens default #6955): each repair round is a full re-emission of the document, and more rounds before streaming only multiplies exposure to the 120-second wall.

The load-bearing risk — review this hardest

The agent has no repository context for the proposed YAML (see the AgentRequest shape above). Cross-model references (uses: aliases) resolve against the owner's .model in the repository; with no project context they fall back to conventions and would report false "unresolvable reference" issues. A false positive is worse than no check: it burns one of the few repair rounds correcting something that is not wrong, and teaches the model to "fix" correct YAML. Cross-model resolution issues must therefore be filtered out of the dry-run result for the agent path (the Builder path, which knows the project, can keep them). That filter is the part to get right.

Acceptance

• A proposal that parses but whose generation would drop glue / fail a cross-model-independent check comes back to the repair loop with those issues, and the repaired proposal reaches the developer.
• A dry run writes nothing and scrubs nothing (assert the workspace is byte-identical before/after).
• A valid multi-model intent proposed with no project context produces zero cross-model false positives.
• IntentAgentServiceTest gains the parse-ok-generation-fails path against a scripted upstream.

Complements #6953 (which moves the map-key class of defect all the way forward to parse time); this issue catches the band the parser legitimately cannot see.