Agent coding loop

Summary

If you use Cursor Agent, Codex, or Claude Code, you already run a repair loop: compiler error → paste context → model patch → verify. On a TypeScript repo, that loop is expensive and fragile. Point replaces paste-and-pray with a compiler that returns JSON shaped for agents: ref, expected, repair, and span.

Live walkthrough with real code: Agent repair tests.

The bug (same application logic, two stacks)

A typo on launch readiness scoring — wrong field name on a record:

TypeScript repo — logic hides inside a function:

// src/lib/math.ts
import type { LaunchSignals } from "../types";

export function launchReadinessScore(signals: LaunchSignals): number {
  let score = 0;
  if (signals.unknownField) score += 30;  // typo
  if (signals.submittedForReview) score += 40;
  if (signals.hasPassingTests) score += 30;
  return score;
}

Point source — logic is a named rule (a typo on one field line triggers unknown-field in check-json):

module Math

record Launch Signals
  has bundle id: Bool
  submitted for review: Bool
  has passing tests: Bool

rule launch readiness
  input signals: Launch Signals
  output score: Int
  score starts at 0
  add 30 when signals.has bundle id
  add 40 when signals.submitted for review
  add 30 when signals.has passing tests
  return score

TypeScript + agent loop (what you do today)

1. tsc fails — text error, no valid field list:

error TS2339: Property 'unknownField' does not exist on type 'LaunchSignals'.
  at launchReadinessScore (math.ts:5:15)

2. You or the agent paste surrounding code — math.ts, types.ts, often the React file that imported the helper (~3,000+ tokens in a real app).

3. Model guesses — without an expected list, it may invent hasBundleId, patch the wrong file, or hit a line number that moved after format-on-save.

4. Retry — wrong fix means another turn: re-paste files + chat history. Token cost compounds.

Problems for agents: patch target is a line in emitted JS; no repair hint; context is implementation glue, not the product rule name.

Point agent loop

1. Structured diagnostic only — no React tree in context:

point check-json math.point

{
  "ok": false,
  "diagnostics": [{
    "code": "unknown-field",
    "ref": "point://semantic/Math/rule.launch readiness",
    "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.",
    "span": { "start": { "line": 12 } }
  }]
}

2. Patch one line at the ref — pick from expected (Point field syntax, not camelCase):

module Math

record Launch Signals
  has bundle id: Bool
  submitted for review: Bool
  has passing tests: Bool

rule launch readiness
  input signals: Launch Signals
  output score: Int
  score starts at 0
  add 30 when signals.has bundle id
  add 40 when signals.submitted for review
  add 30 when signals.has passing tests
  return score

3. Verify with the same gate CI uses:

point check math.point

Why it wins: stable semantic ref; repair sentence + field list; ~130–260 tokens of context vs thousands for a TS paste; authoritative source is .point — emit stays derived.

Side-by-side

Token estimates use ~4 chars/token. Exact counts vary by model and chat history; the order-of-magnitude gap holds on real repos.

Quality, not just tokens

Smaller context helps, but the main win is shape: agents parse JSON fields reliably; they guess from prose and line numbers unreliably. CI in the Point repo proves check-json alone is sufficient to reach a passing check — apply the golden line at the diagnostic span and point check passes (see tests/fixtures/agent-repair/).

Optional: run bun run benchmark:agent-repair-models with provider API keys to compare GPT, Claude, and Gemini repair success under both workflows.

Commands (copy-paste loop)

point check-json path/to/file.point
# read ref, expected, repair, span — patch .point at ref
point check path/to/file.point

For multiple errors: point repair-plan path/to/file.point. Optional context: point explain path/to/file.point <ref>.

See also

• Agent workflow — recommended CLI loop
• check-json — diagnostic schema
• Repair loops
• Proof of concept — live demos and CI fixtures