Commit Graph

2 Commits

Author SHA1 Message Date
sim
8a0c1bce2e test(#3149): correct stale tdd_mode assertion and drop a marker-token collision
Two failures from the remote runner on 654b2cc10, both introduced here.

1. tests/mcp-catalog-parity.install.test.cjs greps emitted workflow files
   for the bare substring 'gsd:section' and treats its presence in a
   composed file as an un-stripped marker. debug.md's new Step 0 prose
   documented the field by writing that token literally, so the emitted
   file tripped the gate even though the parser correctly ignored it as
   prose. Reworded to 'applicability-section markers'. Same class as
   DEFECT.PROMPT-INJECTION-SCAN-COLLISION.

2. tests/debug-session-management.test.cjs asserted debug.md contains the
   literal 'config-get workflow.tdd_mode'. That call is gone. The
   invariant it protected -- tdd_mode comes from the workflow.tdd_mode
   key, never a bare top-level one -- is unchanged, and now has a
   stronger behavioral home: init-debug.test.cjs row A9 asserts a bare
   key is ignored and the canonical key is honored, whatever the read
   mechanism.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-07 10:09:18 -04:00
sim
2bead6ca1d feat(#3149): add dedicated init.debug entry point for /gsd:debug
/gsd:debug was one of the last workflows with no cmdInit* of its own: its
Step 0 made three separate round-trips (state.load, resolve-model
gsd-debugger, config-get workflow.tdd_mode) to assemble one context. Because
no debug-scoped fact was computed at any entry point, ADR-1671 admission gate
(2) could never be satisfied for debug — an applicability atom naming such a
fact would evaluate FALSE forever and silently exclude its section.

Adds cmdInitDebug (init.debug), registers it in the init router and the
command-alias table, and collapses debug.md Step 0 to one call. Every field
resolves through the same primitive the call it replaces used: loadConfig for
commit_docs, withProjectRoot for response_language (#2402), planningPaths for
debug_dir, resolveModelInternal for debugger_model, and the existing
Boolean(workflow.tdd_mode) idiom for tdd_mode.

PlanningPaths gains a debug field so state.load and init.debug share ONE
debug-directory expression rather than two kept in sync by hand. state.load
keeps emitting debug_dir: it is a shipped query surface with its own test
anchor, so narrowing it would break unseen consumers for no gain.

No WHEN_VOCABULARY atom and no gsd:section marker: gate (1), a consuming
section of at least 400 bytes, belongs to the change that adds the section.

Closes #3149

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-07 09:30:27 -04:00