Package budgets and scoped correction rounds
Version 5.2.0 made each verifier faster: 7.6–11 minutes instead of 12, and 32–38 turns instead of 49. The pilot run still hit its one-hour cap. It ran three full reviews, and two correction cycles went mostly to minor findings. The second cycle, which fixed minor findings only, introduced a new major one. This release bounds each agent and keeps correction rounds for findings that matter.
Changes
- Budget ceilings. Each agent type has a ceiling for turns and tool calls: analyst 45 / 50, implementer 60 / 70. A package may set a lower budget. A plan whose budget exceeds its agent's ceiling fails with
BUDGET_EXCEEDS. The plan check prints each package's effective budget. - Runtime limit. The Claude Code adapter writes
maxTurnsinto the analyst and implementer agent files, so the runtime stops the agent at the ceiling. - Stop at the budget. An agent that reaches its budget stops and returns
BLOCKEDwith what it checked and what it did not. The lead escalates it rather than retrying, because a retry meets the same limit. - Usage reporting. The ledger records what each package spent, as the runtime reports it. The gate reports totals and the packages that went over budget. This is information only; it does not change the verdict.
- Correction rounds. A rework amendment whose findings are all minor fails with
REWORK_MINOR_ONLY. Minor findings stay open for the pull request. - Scoped re-review. The re-review of a correction covers only the change since the last reviewed head, plus the findings it answers.
Verification
- Skill validation passed, and all 155 deliberate validator mutations were caught.
- 69 check codes are covered by 228 adversarial cases, 5 of them new.
- Six orchestration scenarios were re-recorded and met all 26 expectations.
- CI passed on the merged branch.
- The 5.2.0 pilot result behind these numbers is in
.pilot/results/claude-multi-audit-5.2.0.*.
Limits
- No pilot run has measured this release yet. The 5.2.0 evidence comes from a single run, so how much the timing varies between runs is unknown.
- No runtime caps tool calls, and Codex caps neither turns nor tool calls. On Codex both limits are instructions only; on Claude Code the tool-call limit is.
- The scoped re-review is a rule the agents are told to follow. No check verifies it.
- Existing installs keep their current agent files. Re-running the installer writes the new versions, which include
maxTurnsand the new instructions, next to them with a.l-gevitysuffix. Merge them by hand. A projectalchemy-workspace.jsonthat overridesenvelopeskeeps its own instructions and budgets.
Includes #46.