Platform observability¶
Platform observability is a development-only diagnostic over CodeClone's own runtime telemetry. It exists to help maintainers build and profile CodeClone itself.
Not a repository quality signal
Observability measures CodeClone's runtime, not your code. High database query counts, large MCP payloads, or a hot semantic reindex say nothing about the quality or safety of the repository being analyzed. It never affects reports, gates, baselines, memory facts, or edit authorization.
Disabled by default¶
Observability is off unless you explicitly enable it through environment
variables. When disabled (or when no telemetry store exists), the diagnostic
returns an inert status=disabled / status=no_store envelope rather than an
error.
| Environment variable | Purpose |
|---|---|
CODECLONE_OBSERVABILITY_ENABLED |
Master switch — turns telemetry capture on |
CODECLONE_OBSERVABILITY_PERSIST |
Persist captured spans to the local store |
CODECLONE_OBSERVABILITY_PROFILE |
Enable per-phase profiling detail |
CODECLONE_OBSERVABILITY_CAPTURE_PAYLOAD_SIZES |
Record payload sizes (numeric only) |
CODECLONE_OBSERVABILITY_CORRELATION_ID / CODECLONE_OBSERVABILITY_PARENT_OPERATION_ID |
Correlate spans across a chain |
How to read it¶
The maintainer-facing surface is the MCP tool query_platform_observability, a
read-only sectioned slicer. It emits numeric metrics only — never raw SQL or
payloads. Start at the summary section and follow the recommended next
sections:
summary · slow_operations · memory_pipeline_cost · db_cost ·
agent_context · mcp_tool_matrix · correlated_chains · costly_noops ·
pipeline · analysis_phase_cost
detail_level accepts compact, normal, or full (aggregate sections
downgrade full to normal); limit clamps to [1, 100]. The
branded HTML cockpit remains the human-facing everything-view.
For the maintainer-only skill, see codeclone-platform-observability; in an MCP
client, help(topic="observability") gives the compact contract.