Savings

savings — Session Savings Report

Report token savings from rtk and jcodemunch usage during this session.

Procedure

1. Run rtk gain --format json to get current savings data. If rtk is not available, skip rtk reporting.
2. Gather the session cache files (aggregate, do not pick one):
   • Use Bash(ls /tmp/rig-session-*.json) to list candidates (or the Glob tool with pattern /tmp/rig-session-*.json).
   • For each candidate path, use the Read tool (not cat or grep) to load the JSON and check the cwd field. Collect ALL files whose cwd matches the current project directory ($CLAUDE_PROJECT_DIR or pwd) — hooks running from subagent contexts or under different session ids write separate cache files keyed by (cwd, session id), and each holds part of this session's tool-call counters.
   • Anchor staleness to session start: session-start recaptures the metrics baseline every session, so the most recent matching file with a non-null metricsBaseline (highest updatedAt among those) provides the session-start anchor via its metricsBaseline.capturedAt — subagent-context fragments never run session-start and carry a null baseline, so they cannot anchor. Discard matching files whose updatedAt predates the anchor, and as an outer bound discard any file older than 24 hours — files from earlier sessions today (same cwd, different session ids) hold a previous session's work and would inflate "this session" counts.
   • Sum metricCounters (rtkCalls, jmCalls, efficientCalls, graphifyCalls) across the remaining matching files. These summed values are the session call counts used in the report.
   • Take metricsBaseline and environment (rtkAvailable, jcodemunchAvailable) from the most recent matching file (highest updatedAt) — baselines and environment snapshots must not be summed.
   • If no file has a matching cwd, fall back to the single most recent file by modification time.
   • rig init auto-allows Bash(ls /tmp/rig-session-*) and Read(/tmp/rig-session-*.json), so this flow should not prompt for permission. If it does, .claude/settings.json is out of date — re-run rig init --force.
3. Compute the rtk session delta: current total_saved - baseline totalSaved.
4. For jcodemunch: call mcp__jcodemunch__get_session_stats and read session_tokens_saved, session_calls, total_tokens_saved, and tool_breakdown directly. These are reliable per-session counters maintained by the MCP server process. If jcodemunch MCP is not available, skip jcodemunch reporting.
5. For graphify: check the most recent matching cache file for graphifyStats in the metricsBaseline field. This is a per-directory record mapping absolute paths to GraphifyProjectStats objects. If present, include the graphify section in the report (see Output Format below). Alternatively, if graphify MCP is available, call mcp__graphify__graph_stats to get live stats. If graphify is not available, skip graphify reporting.
6. For headroom (context-compression proxy): check the most recent matching cache file's environment.headroomInitialized field. If true, run headroom perf --format json --hours 24 and read tokens_saved, savings_pct, total_requests, and cache_hit_pct. Include the headroom line only when total_requests > 0. These are context-layer savings — they overlap tool-layer (rtk/jcodemunch) savings at the boundary, so report them on their own line and NEVER add them to the rtk/jcodemunch numbers. If headroomInitialized is false/absent or the command fails, skip headroom reporting.
7. Format and print the report (see Output Format below). Do NOT write any explanatory text before or after the report — output ONLY the report lines.

Output Format

With session data (baseline + delta available), single project:

[rig] Session Savings
  rtk: X.XM saved (N calls, +XK this session)
  jcodemunch: XK saved (N queries, 150M total all-time)
  headroom: XK saved (context layer — N requests, X% compression, X% cache hits; not summed with tool-layer savings)
  graphify: N nodes, M edges, K communities (X% EXTRACTED, X% INFERRED, X% AMBIGUOUS)

The headroom line appears only when the proxy is initialized for this project and had requests in the window — omit it otherwise.

With session data, multi-project (scouted external repos):

[rig] Session Savings
  rtk: X.XM saved (N calls, +XK this session)
  jcodemunch: XK saved (N queries, 150M total all-time)
  graphify:
    project-a: N nodes, M edges, K communities (X% EXTRACTED, X% INFERRED, X% AMBIGUOUS)
    project-b: N nodes, M edges, K communities (X% EXTRACTED, X% INFERRED, X% AMBIGUOUS)

Without session data (no cache file or baseline is null):

[rig] Session Savings (all-time)
  rtk: X.XM saved (N commands, XX.X% avg savings)
  jcodemunch: XM saved all-time (N queries)
  graphify: N nodes, M edges, K communities (X% EXTRACTED, X% INFERRED, X% AMBIGUOUS)

When baseline is missing, still call mcp__jcodemunch__get_session_stats for all-time totals and mcp__graphify__graph_stats for live graph stats. Only skip a line if the tool is genuinely unavailable.

If rtk is not installed:

[rig] Session Savings
  rtk: not installed

Formatting Rules

• Format token counts: >=1M as X.XM, >=1K as XK, else raw number.
• Round percentages to 1 decimal.
• jcodemunch session_tokens_saved and session_calls come directly from get_session_stats. total_tokens_saved is the all-time cumulative.
• graphify graphifyStats is a Record<string, GraphifyProjectStats> — keys are absolute directory paths, values are stats objects with nodes, edges, communities, extractedPct, inferredPct, ambiguousPct. Single-entry records render as one line; multi-entry records render as indented per-project lines with directory basename as label.
• Output the report and nothing else.