[vicgupta]

Summary

• New Executor.FormatError check before dispatch: a call is a format error when its arguments are invalid JSON / not an object, the tool name isn't in the run's catalog, a schema required field is missing (or null), or a field has the wrong JSON type. Applies to MCP tool schemas too.
• Malformed calls are not executed (and no hooks run); the model gets format error: <reason>. Call one of: <tools> with arguments matching its schema. (capped at 500 chars). Well-formed calls in the same response run as before.
• A step whose calls are all malformed increments a consecutive counter; any well-formed call resets it. At N (default 3) the run ends with new status RepeatedFormatError, empty submission, non-zero exit; limits are still checked first.
• Config max_format_errors and flag --max-format-errors (flag overrides config). Must be an integer >= 1.
• Visibility: tool.result gets "format_error": true; result event gets format_errors; summary line shows N format errors (+ last reason on RepeatedFormatError); trajectory gets format_errors, consecutive_format_errors, format_error_messages (message indexes), last_format_error. Original arguments are kept in the trajectory.
• Resume: counter persists and is restored; checkpoint validation now accepts malformed calls only when their result is a marked format-error correction; RepeatedFormatError checkpoints are resumable like limit statuses.
• README: flag table, config example, Exit status, new "Malformed tool calls" section.

Decisions on open questions

1. No-tool-call responses stay Completed (unchanged).
2. 0 is rejected (exit 2 for the flag); no "no cap" mode.
3. Truncation (finish_reason=length) detection not added.
4. Type validation is JSON types only (no enum/min/max). Integers still accept numeric strings, as the tools always did.
5. No schema_version bump; fields are additive/omitempty and ToolSignature is unchanged, so existing checkpoints still resume.
6. Names as proposed: RepeatedFormatError, max_format_errors / --max-format-errors, format_error.

Other choices:

• In --read-only, calling a non-read-only built-in keeps the existing read-only denial (not a format error); a truly unknown tool is a format error (one test expectation updated accordingly).
• Invalid config values (0, -1, abc, 1.5) fail before any provider request with a message naming max_format_errors, but exit 1 like every other config error (AC8 asked for 2; flags do exit 2). 1.5 is explicitly rejected (yaml would otherwise truncate it to 1).
• Help shows integer defaults unquoted ((default 3)); usage golden updated for the wider flag column.

Tests

internal/agent/format_test.go covers AC1–AC7, AC9, AC11 (trajectory markers, resume with consecutive=2 → stops on next malformed step, unmarked malformed call rejected by validation), plus wrong-type and mixed-step reset. cmd/tinyagent/format_test.go covers AC5/AC7 end-to-end via a fake OpenAI server (default 3, flag 1, config 5, flag overrides config), AC8 (flag/config validation, zero provider requests), AC10 (--json stays pure JSONL, format_error and result.status/format_errors). Existing suite unchanged otherwise (AC12). gofmt/vet/build/test pass locally.

Closes #66

[vicgupta]

Rebased onto main by the hourly merge job (head now 051ebe1). Two notes on the rebase:

1. Beyond conflict resolution: removed the duplicate captureStdout helper from cmd/tinyagent/format_test.go — it is identical to the one in cmd/tinyagent/schema_test.go (landed on main via Add --output-schema to end runs with a validated, schema-conforming JSON result for scripting and CI #72), and keeping both failed go vet with a redeclaration error. Main's copy is kept.

2. The internal/agent/loop.go conflict (main refactored the tool-call loop into toolStep/recordToolResult helpers) was resolved by keeping those helpers: well-formed calls still go through toolStep, while format-error calls are handled in the loop body, still skip hooks, and still emit tool.result with format_error: true. No behavior change intended.