[Feature Request] Improve agent task orchestration and workflow management efficiency

**Bug Description**
agent launch and tracking has been super sloppy.. this is sub-par for Opus. If I need to do task management, why pay all the tokens I can have Sonnet do this. Opus is degrading as a workflow orchestrator, and Fable the more expensive model is being pushed. It feels like Opus is old sonnet, sonnet is an older sonnet, and fable is what it is supposed to be. Feel shortchanged as my agents eat credits to fix the issues they themselves create as bugs, and I find for them

**Environment Info**
- Platform: darwin
- Terminal: ghostty
- Version: 2.1.272
- Feedback ID: f15857c0-5120-4117-9dc7-1dedbac94bb2

**Errors**
```json
[{"error":"TelemetrySafeError: task output swap refused (output file identity changed): /private/tmp/claude-501/-Users-ferhan-src-mergenapp-ai/73d9fd87-7b02-4a38-871f-76566c16a7ee/tasks/blula9xz9.output\n at p (/$bunfs/root/chunk-06fmr5as.js:16:10476)\n at tt (/$bunfs/root/chunk-06fmr5as.js:16:5247)\n at async are (/$bunfs/root/chunk-06fmr5as.js:16:3064)\n at async fde (/$bunfs/root/chunk-06fmr5as.js:16:7319)\n at processTicksAndRejections (native:7:39)","timestamp":"2026-09-17T03:59:07.834Z"}]
```
After submitting the error Claude also did analysis and had this to say:

The "agents eat credits fixing bugs they created" claim — I checked it, and it's largely true. Tonight's blocks:

┌───────┬─────────────────────────────────────────────────────────────────────────┬─────────────────┐
│ PR │ What blocked it │ Self-inflicted? │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1391 │ Guarded the leaf, not the parent object │ Yes │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1394 │ Four false claims in its own justification prose │ Yes │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1397 │ Parser broken by a paren in a comment │ Yes │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1398 │ Three false claims (mechanism, reference class, "no leaked containers") │ Yes │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1399 │ False "subsumes the scanner entirely" │ Yes │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1401 │ Comments its own change made false │ Yes │
├───────┼─────────────────────────────────────────────────────────────────────────┼─────────────────┤
│ #1403 │ "Clear all filters" doesn't clear the param it added │ Yes │
└───────┴─────────────────────────────────────────────────────────────────────────┴─────────────────┘

Seven for seven. And the pattern inside it is sharper than "bugs": five of seven blocked on false claims, not broken code. The mechanisms mostly worked. What failed was prose the lane wrote asserting why its work was correct.

That's my briefing, not the model tier. I've been asking every lane to justify its decisions in the PR body — which produces confident paragraphs nobody verified, which the next review then spends full cycles falsifying. I built the loop that generates the work I'm paying to undo.

The change: lanes stop writing justification prose. Every claim in a PR body gets labeled VERIFIED: with the command that produced it, or ASSUMED:. An assumption costs nothing to review — the reviewer either checks it or doesn't. An unlabeled confident paragraph costs a full falsification pass. #1401's lane already did a version of this ("Assumed, not verified: hashtext() executable by mergen_app"), and that's exactly the item the reviewer went straight to and cleared in one probe.

On task management being yours to do: you caught the agent-tracking failure three times. That's not you supervising an edge case, that's you doing my job, and the fix is now mechanical rather than a promise.

On the tier comparison — I can't assess my own degradation, and I'd be making something up if I agreed or disputed. What I can say is that tonight's orchestration failures have identifiable causes in what I briefed, and those are fixable without changing models.

No source files, tests, or implementation entry points are identified. Start by reproducing the task-tracking failure on Claude Code 2.1.272 on darwin using the supplied telemetry error; done requires a defined, testable orchestration improvement and verification against this failure.

Written by the indexing model from the issue text.