fix(ai): re-read agent.maxSteps live so a raised limit takes effect mid-run

[ndemianc]

The bug

In autopilot, changing Maximum tool-use steps (levelcode.ai.agent.maxSteps) from 25 to 1000 had no effect — the agent still paused at "step limit" after 25.

agentFlow read the cap once, as a plain number, when a goal started:

maxSteps: Math.max(1, cfg.get('agent.maxSteps', 25)),   // cfg = a snapshot from goal start

The loop in agent.js (while (step++ < ctx.maxSteps)) then compared against that frozen 25 for the whole run, so raising the setting mid-run never took effect.

The fix

Hand runAgent a live getter that re-reads a fresh config on every access — mirroring the autopilot getter three lines below it, which exists for the same "must be live mid-run" reason:

get maxSteps() { return Math.max(1, aiConfig().get('agent.maxSteps', 25)); },

aiConfig() returns a fresh getConfiguration('levelcode.ai') (VS Code config objects are snapshots, so re-fetching is what makes it live). The loop reads ctx.maxSteps every iteration and never copies it, so a changed limit now applies on the next step:

• Raise 25 → 1000 while it's running → it continues past 25, live.
• Lower it mid-run → it stops sooner. Dynamic both ways.
• The pause card's "paused after N steps" now reports the actual limit in effect, not a stale 25.

Tests

New test/agentMaxSteps.test.js — a source-invariant guard in this suite's existing idiom (webviewCss, providerMode), because a live runAgent needs a workspace + provider the pure-unit suite deliberately doesn't stand up. It locks down both ends and fails loudly if either regresses to a snapshot:

• extension.js passes maxSteps as a getter over a fresh aiConfig() (never a static capture),
• agent.js gates the loop on ctx.maxSteps directly (never hoisted/destructured).

Full extension suite: 23 passing, 0 failing.

🤖 Generated with Claude Code

[Copilot]

Pull request overview

Fixes autopilot’s step-limit behavior so levelcode.ai.agent.maxSteps is re-read live during an in-progress agent run (instead of being snapshotted at goal start), allowing users to raise/lower the cap mid-run and have it take effect on the next step.

Changes:

• Pass maxSteps into runAgent as a live getter that re-reads a fresh aiConfig() each access.
• Add a source-invariant unit test guarding (a) the contributed setting shape and (b) that both the extension call-site and agent loop continue to treat maxSteps as live.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
extensions/levelcode-ai/extension.js	Switches maxSteps passed to runAgent from a snapshotted number to a live getter re-reading aiConfig().
extensions/levelcode-ai/test/agentMaxSteps.test.js	Adds a regression test ensuring agent.maxSteps remains live (and the setting remains unbounded above).

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 2 out of 2 changed files in this pull request and generated no new comments.