Agent UX: pm close recovery bundle emits a suggested_retry that cannot parse and drops the flags the caller supplied #953

Summary

When pm close rejects an invocation for missing closure fields, the suggested_retry line in the recovery bundle is not a runnable command. It renders each missing flag as a quoted positional value with the flag name duplicated in parentheses, and it silently drops the --reason flag that the operator did supply, leaving the reason text as a bare positional.

An agent that copies suggested_retry verbatim — which is exactly what the field invites — produces a second, differently-broken invocation.

Reproduction

Tracker configured with a strict close policy (requires --resolution, --expected-result, --actual-result).

pm close pmc-s2bk --reason "Aligned changelog:check to the replace-mode generator" --author claude

Actual recovery bundle

provided_fields: --reason, --author
missing: --resolution (--resolution), --expected-result (--expected-result), --actual-result (--actual-result)
suggested_retry: pm close pmc-s2bk "Aligned changelog:check to the replace-mode generator" "--resolution (--resolution)" "<value>" "--expected-result (--expected-result)" "<value>" "--actual-result (--actual-result)" "<value>"

Three defects in that one line:

1. --reason is dropped. The supplied reason becomes a bare positional argument, so the retry loses the flag that carried it.
2. Flag names are emitted as quoted values. "--resolution (--resolution)" is passed as an argument rather than as the flag --resolution. The parenthesised duplicate appears to be the missing: field's display label (--flag (--flag)) being interpolated directly into the command string instead of the flag token.
3. --author claude is dropped, so a retry that did parse would fail authorship validation next.

Expected

suggested_retry should be a command that runs as-is once the <value> placeholders are filled:

pm close pmc-s2bk --reason "Aligned changelog:check to the replace-mode generator" --author claude --resolution "<value>" --expected-result "<value>" --actual-result "<value>"

That is: preserve every flag the caller already provided, and append each missing flag as --flag "<value>".

Why this matters for agents

The recovery bundle is the single most machine-readable part of the error surface, and suggested_retry reads as the authoritative "do this next" instruction. When it is malformed, an agent burns a full retry cycle on a command that cannot parse, then has to reconstruct the invocation from the missing: list anyway — which makes suggested_retry net-negative compared to omitting it.

The same generator appears to back other commands' recovery bundles, so the fix is likely one shared code path.

Related

• Agent UX: pm close recovery bundle suggests --validate-close "<value>" but the flag is an enum (off|warn|strict), causing a guaranteed invalid retry #519 — pm close recovery bundle suggests --validate-close "<value>" when the flag is an enum. Same field, but that issue is about an invalid value placeholder; this one is about the command structure being unparseable and dropping caller-supplied flags.
• Agent UX: pm show rejects --id while create/close/notes/activity accept it, and the recovery points at the wrong command #950 — recovery text pointing at the wrong command.