Test coverage for memory + scratchpad subsystem

[mlund01]

Summary

Follow-up to #102. The post-ship audit flagged that the memory + scratchpad rework had solid coverage at the config-parse and store-build layers (~90%) but the tool surface, the agent auto-attach branch, the prompt formatter, and the wsbridge file browser were untested (~20-30%). This PR closes those gaps.

50 new tests, ~960 lines, no production-code changes.

What's covered

aitools/memory_tools_test.go (27 tests)

Direct unit tests for all six file_* tools against a fake MemoryStore that backs slots with real temp dirs (so syscalls exercise actual on-disk behavior):

• Every tool's schema advertises slot (not legacy folder)
• Legacy {\"folder\": ...} JSON payloads are rejected at unmarshal
• file_list: flat, recursive, pagination, empty, unknown slot
• file_read: max_lines, max_bytes, missing path, directory rejected, ../ escape
• file_create: create-new, fail-if-exists, overwrite, append (+ missing target), parent-dir creation
• file_delete: happy path, directory rejected
• file_search: regex match, no match, bad regex
• file_grep: flat, recursive, no-recursion-by-default

agent/internal/prompts/prompts_test.go (7 tests)

FormatMemoryContext behavior:

• nil store and empty store both return empty string
• memory slot gets the persistent-mission label
• scratchpad slot gets the ephemeral label
• Shared slots get no special label
• Output names the slot param and all six file_* tools
• Iteration order is preserved (sort lives in mission/memory_store.go, not prompts)

wsbridge/memory_browser_test.go (13 tests)

File-browser RPC layer that was previously fully untested:

• collectMemoryInfos: shared + per-mission memory both emitted with correct paths under SquadronHome, Editable: true for all
• Missions without memory { } are skipped; scratchpads aren't exposed (preserved behavior, now pinned by test)
• resolveMemoryPath: shared memory, mission memory, unknown name, collision (shared wins — pinned so a future priority swap is loud), mission without memory{} not resolvable
• resolveSafePath: rejects ../, mid-path ../, absolute paths; accepts valid relative and empty (root)
• End-to-end read of a real file via the resolver

agent/memory_tools_attach_test.go (3 tests)

The auto-attach contract — that file_* tools appear iff a MemoryStore is wired:

• All six tools appear when MemoryStore != nil
• None appear when MemoryStore == nil
• Attached tools are wired to the exact store passed in (not nil, not a different instance)

What's still uncovered

Low ROI, deliberately skipped:

• cmd/engage.go runScratchpadCleanupLoop — three-line shim around mission.SweepExpiredScratchpads
• End-to-end mission run with memory + scratchpad — covered by the manual smoke test in Rename folders → memory in HCL (folders kept as alias) #102; a true integration test would need an LLM call

Test plan

• go build ./...
• go test ./... — all 19 packages pass
• New tests run in ~2s combined, no flakes across multiple runs

🤖 Generated with Claude Code