valksor / kvelmo

84%

push

github

Make claude-mcp work as the conductor's agent through full phases

With pipe working, the conductor (plan/implement/simplify/optimize/submit)
still stalled at "planning": claude-mcp launched but the phase never completed.
Three causes, all fixed here, so the full workflow e2e passes with claude-mcp
as the default agent:

1. The orchestration tools were never loaded. claude defers MCP and built-in
   tools behind tool-search and only surfaces ones it judges relevant to the
   literal prompt, so kvelmo_save_artifact / kvelmo_signal_complete were absent
   from the model's context. ENABLE_TOOL_SEARCH=false in the child env loads all
   tools upfront; with --strict-mcp-config (only the kvelmo server) the cost is
   small.

2. MCP tool calls blocked on an approval prompt. acceptEdits auto-approves file
   edits but still raises an interactive prompt for Bash and MCP tools, which
   hangs forever with no human present. The default is now bypassPermissions,
   the correct mode for an autonomous headless orchestrator.

3. The rendezvous socket path overflowed the ~104-byte Unix limit. adapter.sock
   lived under WorkRoot (KVELMO_HOME/work/session-.../claudemcp-.../), which is
   ~139 bytes when KVELMO_HOME is a long temp path — so the bind failed, the
   spawn errored, and the phase stalled with no agent. It now routes through
   paths.ShortenSocketPath (exported for this), the same short-path fallback the
   global and worktree sockets already use.

Idle-detection is also extended to conductor sessions (a longer grace than the
one-shot pipe path): the interactive TUI does the phase work, writes its
deliverable, then idles without exiting, so the same PTY-quiescence signal that
ends a pipe turn ends a phase. The conductor then detects the written spec and
advances, exactly as it does when the binary claude adapter exits.

Verified: TestCLIFullCycle (plan through submit, PR merge, finish) passes with
KVELMO_E2E_AGENT=claude-mcp.

2818 of 3769 branches covered (74.77%)

Branch coverage included in aggregate %.

55 of 60 new or added lines in 3 files covered. (91.67%)

48698 of 57765 relevant lines covered (84.3%)

2.16 hits per line