Agent repair walkthrough

Agent (or you) runs the CLI. Compiler returns structured JSON — no file paste.

What the agent reads

{
  "schemaVersion": "point.core.check.v1",
  "ok": false,
  "diagnostics": [
    {
      "code": "unknown-field",
      "message": "Unknown field unknownField on LaunchSignals",
      "path": "fn.launchReadinessScore.if.condition",
      "ref": "point://semantic/Math/rule.launch readiness",
      "severity": "error",
      "span": {
        "start": {
          "line": 12,
          "column": 1,
          "offset": 207
        },
        "end": {
          "line": 12,
          "column": 36,
          "offset": 242
        }
      },
      "expected": [
        "has bundle id",
        "submitted for review",
        "has passing tests"
      ],
      "actual": "unknownField",
      "repair": "Use one of: has bundle id, submitted for review, has passing tests.",
      "relatedRefs": [
        "point://semantic/Math/record.Launch Signals.field.has bundle id",
        "point://semantic/Math/record.Launch Signals.field.submitted for review",
        "point://semantic/Math/record.Launch Signals.field.has passing tests"
      ]
    }
  ]
}

Measured from real CLI output on this fixture (1,051 chars). CI asserts context stays under 1,200 chars.

Agent picks a field from expected, replaces line 12. No repo search — ref and repair tell it where.

One-line change

-   add 30 when signals.unknown field
+   add 30 when signals.has bundle id

CI applies the golden line from the fixed fixture — no LLM — and point check passes.

Same gate as CI. If check passes, the repair loop is done.

Terminal

$ point check tests/fixtures/agent-repair/unknown-field-broken.point
Point core check passed: tests/fixtures/agent-repair/unknown-field-broken.point

Compiler returns a line number in emitted JS/TS. No list of valid field names.

What the agent reads first

error TS2339: Property 'unknownField' does not exist on type 'LaunchSignals'.
  at launchReadinessScore (lib/math.ts:18:15)

Real error shape; token count from this message only.

Typical Cursor/Codex workflow: paste component, lib, tests — agent hunts for the typo.

Representative paste (truncated)

// ReadinessPanel.tsx — excerpt (~320 lines total in real repos)
import { useMemo } from "react";
import type { LaunchSignals } from "../../types";
import { launchReadinessScore, scoreStatusLabel } from "../../lib/math";

export function ReadinessPanel({ signals }: { signals: LaunchSignals }) {
  const score = useMemo(() => launchReadinessScore(signals), [signals]);
  const label = scoreStatusLabel(score);
  return (<section><h2>Launch readiness</h2><p>Score: {score} — {label}</p></section>);
}

// lib/math.ts — agent often pastes this too when tsc fails
export function launchReadinessScore(signals: LaunchSignals): number {
  let score = 0;
  if (signals.unknownField) score += 30;
  if (signals.submittedForReview) score += 40;
  if (signals.hasPassingTests) score += 30;
  return score;
}

Illustrative ~12,000-char paste heuristic. Not from a logged agent session.

If the model emits camelCase or patches the wrong file, the next turn reloads the same context.

Second turn (common)

Same ~3,000 tokens pasted again + new error output

Not CI-tested. Shown because retry loops are common without structured repair hints.