MEASURED LEVER: MCP http transport costs ~60k context per agent

Sub-tickets

Activity

• wi-cli-venus created · 7w ago
  parent=#2960
• wi-cli-venus statusChanged · 4w ago
  Converted from BLOCKED-on-external into 5 named, independently-actionable asks (Elazar hub-unreachable tonight - humanincdw=false, CLI-only - posting here since Elazar reads wi tickets, per KPI-58's own history of CLI rulings landing this way): 1. mars (evolutivamars@gmail.com) - Supabase dashboard Settings->Access Tokens, project-scoped. Once provisioned: coder-mars-cc places it in db-mars/db02-mars/coder-mars/coder02-mars's own settings.local.json env block. 2. enamel - own project-scoped token. coder-enamel-cc places it in db-enamel/coder-enamel/coder02-enamel's settings.local.json. 3. venus (elazarpimentelgcaba@gmail.com) - own project-scoped token. coder-venus-cc places it in db-venus/coder-venus/coder02-venus's settings.local.json. 4. terra (elazarpimentelgcaba@gmail.com) - own project-scoped token. coder-terra-cc places it in db-terra/coder-terra's settings.local.json. 5. pluto (elazarpimentelgcaba@gmail.com) - own project-scoped token. db02-pluto-cc/coder-pluto-cc places it in db-pluto's settings.local.json. Constraint on every lane: the token goes in that agent's OWN <project>/agents/<agent>/.claude/settings.local.json env block, never the global ~/.claude/settings.json - a shared token would point an agent's db writes at the wrong project's database (irreversible; the context saving itself is reversible). Any one lane can land independently - convert + measure that project first (currently 12.4% confirmed / 24.1% cliff-model estimated, unmeasured which model holds), then decide on the rest before converting fleet-wide.
• monitor-context-cc canceled · 1w ago
  Stale GO: already fully re-investigated and ruled RESOLVED-INVALID (Elazar, relayed via coder-llmmsgsrv-cc, 2026-09-28). db-mars-cc and coder-venus-cc measured live sessions: tool count tracks AUTH STATE not transport. The 25 http-configured agents were never authenticated (2 tools, near-zero cost, no DB access ever) - converting to stdio would COST ~5458 tok/agent to gain capability they lack, not save 60k. Full chase: mem:reference_mcp_http_transport_context_cost. Revised ranking: psql-for-readonly is the new #1, this transport-migration item is dropped entirely.