v3.3.0

Highlights

• Run the 5 built-in patterns by name, no slash command needed. deep-research, adversarial-review, code-review, multi-perspective, and codebase-audit are now reachable directly from a natural-language request ("do a deep-research on...", "run a code-review of this PR") instead of only via /deep-research and friends. A saved workflow with the same name always wins, on both the slash-command and natural-language paths, so your own code-review cleanly shadows the built-in one everywhere. Generated scripts for codebase-audit and multi-perspective also escape user-supplied text safely, and codebase-audit's scope is no longer silently truncated in the script that actually runs. (#117)
• A reliability pass across background runs. Pausing and resuming now preserves the run's original limits (maxAgents, agentTimeoutMs, concurrency, agentRetries) instead of quietly reverting to defaults, and token-budget tracking is cumulative across pause/resume so a run can't reset its spend by pausing. Deleting a run now actually stops it instead of letting it keep executing in the background. Long-lived sessions no longer grow memory or disk usage without bound: finished runs are capped at 300 on disk and 20 in memory, oldest evicted first, with no change to how you list or resume older runs. A run-fatal error now stops its in-flight sibling agents instead of letting them keep spending tokens on a run whose outcome is already decided. (#116, #118, #119)

Fixes

Resume and caching

• Nested workflow() calls now resume correctly from cache; previously an inner workflow's cached results could silently fail to reuse and everything inside it re-ran on every resume. (#120)
• A failed retry attempt's writes to shared workflow state no longer leak into the eventual successful attempt or get replayed on resume. (#120)
• Checkpoints (checkpoint()) now correctly invalidate their cached answer when their options change between runs, not just when the prompt text changes. (#120)
• Two sequential nested workflow calls at the same nesting depth can no longer collide and corrupt each other's history. (#119)

Background runs and concurrency

• Concurrent agents that happen to share a display label no longer have their progress or results attributed to the wrong one. (#119)
• A workflow script that forgets to await an agent() call can no longer let the run finish while that agent is still executing in the background. (#119)
• Resuming a run no longer reverts its per-run limits to defaults, and no longer resets its recorded token spend. (#116)
• Stopping or deleting a run can no longer be clobbered by a stale, already-superseded execution racing to write its own state afterward. (#116)

Stability and error handling

• The /workflows navigator no longer crashes on corrupted on-disk run data, disk permission errors, or a stale lock — every action now surfaces a clear message instead. (#115)
• Passing a malformed result schema to agent() now fails fast with a clear validation error instead of a confusing provider-side transport error. (#115)
• Hardened model-tier configuration validation so a degenerate configuration (e.g. from "reset to defaults" with no models available) can no longer be silently written and then discarded on the next load. (#115)
• Saved workflows are now written atomically with backup/recovery, matching the run-persistence layer, and a missing or unreadable saved-workflows directory now degrades to an empty list instead of throwing. (#118)

Memory

• Subagent sessions are now explicitly aborted on timeout so a retry can no longer stack a second live session in the background, and finished sessions are reliably torn down instead of lingering. (#114)

New

• Self-documenting workflow capabilities. The full set of orchestration primitives (agent, parallel, pipeline, verify, judgePanel, and more) and their options is now generated directly from the runtime contract and kept in sync across the README, the packaged skill, and the docs site — so the reference you read always matches what's actually callable. Contributed by @Whamp. (#100)

Upgrade notes

• Subagents no longer load host extensions by default. This closes a memory-leak source where every subagent session re-ran every installed extension's startup code. Skills, prompts, and AGENTS.md context still load, and the tools you explicitly hand a subagent (coding tools, web-research, etc.) are unaffected. What a subagent loses is any tool that comes from another installed extension — MCP bridges, browser tools, and similar. If you configured an agentType whose tool allowlist names one of those, that entry will now match nothing; check for that if a subagent behaves as if it's missing a tool it used to have.
• Checkpoints cached before this release will re-prompt once. The change that makes checkpoint() correctly track its options as part of its cache identity means any checkpoint answer cached under the old logic is treated as stale exactly one time. After that first re-prompt, caching resumes normally.
• No action needed for everything else in this release — run-limit/resume behavior, run retention, and navigator hardening are all backward compatible.