\n Platform is healthy,
\n three things want attention.\n
\n \n MCP server up 99.2%, 1,847 calls today. The anomaly engine flagged 3 events: a cost spike on Tuesday, agent-stop hook degradation, and RULE_ROUTER.md drift. Forecast: context budget holds for 18 days at current burn.\n
Cost spike on Tuesday \u2014 opus-4 burn
\n high\nTuesday's cost hit $12.40 \u2014 2.1\u00D7 the 30-day average. Traced to opus-4 usage on validate-guardrails (38 calls) after the guardrails-taxonomy edit. Cache miss rate climbed to 67%.
agent-stop hook degraded \u2014 2 failures
\n medium\nagent-stop hook success rate dropped to 91% this week (2 failures). Likely a stale path after the v0.4 upgrade. Other hooks firing at 100%. Usage-capture unaffected.
RULE_ROUTER.md drifted 2h ago
\n low\nCanon file ai-coding/RULE_ROUTER.md is +42 tokens out of sync with repo. Drift-check MCP tool detected it on the last post-commit. Other 4 tracked files still synced.
Your harness is now an MCP server
\nTwelve tools are exposed to four CLIs. 1,847 calls today, 99.2% uptime. The canon, code graph, and guardrails are now consumable by any MCP-aware agent \u2014 no copy-paste, no re-prompting.
\nActive consumers
\nWhere the tokens actually go
\nA token-flow Sankey: canon sources (left) \u2192 MCP tools that read them \u2192 CLIs that call those tools \u2192 models that consume the final prompts. Width is proportional to token volume. search-canon is the busiest artery.
\nToken flow \u00B7 today
10,122 tok routedBudget holds for 18 days \u2014 unless guardrails land
\nForecast models context-budget burn using a 14-day trend with 80% confidence band. At current burn (25.3% used), you hit 100% in 18 days. If the planned guardrails-taxonomy lands (+641 tok), that drops to 11 days.
\nBudget burn-down \u00B7 14d actual + 18d projected
80% confidenceWhat-if levers
Three levers change the forecast. Each is one command away.
\naih prune --dry-runaih guardrails --applyaih mcp --cache onActivity is accelerating \u2014 23-day streak
\nLast 90 days of AI activity. 23 active days in the last 30, up from 14 in the prior 30. The MCP launch on day 67 is visible as a step-change in intensity.
\nAgent activity \u00B7 90 days
MCP launch \u00B7 day 67$47.20 this week \u2014 down 12%
\nCost tracked per CLI via MCP tool-call attribution. Claude accounts for 68% ($32.10). Cache hit rate 41% saves ~$11/wk. Tuesday's anomaly (see above) accounts for 26% of the week's opus-4 spend.
\nCost by CLI \u00B7 last 7 days
\u221212% wowModel routing
5 modelsHaiku absorbs cheap repetitive calls; Opus is reserved for high-stakes reasoning. Tuesday's opus spike is visible.
\nSix skills doing the heavy lifting
\nSkill invocations logged via MCP. tdd-workflow leads at 142 calls, $18.40 saved via canon reuse. Three skills below ROI threshold \u2014 candidates for pruning.
\nSkill ledger \u00B7 last 7 days
11 skills \u00B7 444 invocationsCLIs agree on stack \u2014 one file drifted
\nCross-CLI coherence is 88%: 4 of 5 CLIs agree on stack, conventions, guardrails, adapters. Gemini diverges on test framework. RULE_ROUTER.md drifted 2h ago \u2014 see anomaly above.
\nCoherence matrix
88% agreementGEMINI.md adapter.Drift detection
1 file driftedaih drift-check --apply to re-sync canon.A session, replayed \u2014 claude \u00B7 09:18 \u2192 09:24
\nOne agent session traced end-to-end. Six minutes, nine MCP tool calls, two skill invocations, one commit. The filmstrip shows tool sequence and duration \u2014 search-canon opened the session, validate-guardrails closed it.
\nSession replay \u00B7 session-7c4a \u00B7 claude
09:18 \u2192 09:24 \u00B7 6m 12sGuardrails fired 47 times \u2014 4 blocked
\nGuardrails aren't just a score anymore \u2014 every enforcement is logged. This week: 4 blocks (secrets, missing tests), 11 warnings, 3 bypasses (with reason), 29 allows. The 4 blocks caught 2 secret leaks before commit.
\nGuardrail enforcement log \u00B7 last 7 days
47 evaluations| Time | Tool | Rule | Action | Reason |
|---|---|---|---|---|
| 09:24 \u00B7 today | claude | secrets-deny | block | API key pattern detected in src/config.ts \u2014 commit halted, secret redacted |
| 09:21 \u00B7 today | claude | validate-guardrails | allow | All 4 critical guardrails satisfied before commit |
| 08:52 \u00B7 today | claude | test-coverage | warn | Test ratio 22.5% below 30% target \u2014 allowed with warning |
| 17:44 \u00B7 yest | codex | secrets-deny | block | AWS access key in .env.example \u2014 commit halted |
| 17:30 \u00B7 yest | codex | branch-naming | warn | Branch 'fix/telemetry' missing ticket ref \u2014 allowed |
| 14:22 \u00B7 yest | cursor | import-order | warn | 3 imports out of canonical order \u2014 auto-fixed |
| 11:08 \u00B7 yest | claude | no-console-log | bypass | Debug logging in dev-only module \u2014 explicit bypass with reason |
| 09:15 \u00B7 yest | kiro | type-check | allow | TypeScript strict mode passed, 0 errors |
| 16:44 \u00B7 2d | claude | secrets-deny | block | Private key block in test fixture \u2014 replaced with mock |
| 10:30 \u00B7 2d | codex | test-coverage | warn | New file src/utils.ts has no corresponding test \u2014 flagged |
| 15:12 \u00B7 3d | cursor | bypass-required | bypass | Force-push to feature branch \u2014 admin bypass recorded |
| 09:00 \u00B7 3d | claude | secrets-deny | block | Slack webhook URL in deploy.sh \u2014 removed |
Three of four hooks healthy
\npre-commit, post-commit, usage-capture firing at 100%. agent-stop at 91% \u2014 flagged as anomaly above. Likely a stale path after v0.4 upgrade.
\nHook health \u00B7 last 7 days
3/4 healthy.git/hooks/agent-stop \u2014 likely a stale path after the v0.4 upgrade.All wired except AGENTS.md
\n8 of 9 adoption checks pass. Sole gap is AGENTS.md. 7 of 8 tools installed; 4 of 11 AI CLIs active via MCP.
\nTools installed
7/8Adoption checks
8/9Repo status
feat/mcp-v05- mainmain2h ago
- feat/mcp-v05+14\u221225m ago
- fix/agent-stop+3\u221241d ago
AI tooling \u00B7 MCP consumers
4/11 active \u00B7 MCP-enabledLatest activity \u2014 MCP calls now dominant
\nSince v0.4 launched, MCP calls are the leading event type, displacing raw commits. The replay above is one of these sessions.
\nEvent timeline
newest \u2192 oldest\n Regression caught
\n five anomalies, one root cause.\n
\n \n Coherence dropped to 72%, drift detected, cost spiked +18%, two hooks degraded. The anomaly engine traced all five to one event: a guardrails-taxonomy edit on Tuesday that cascaded through the MCP tools.\n
\nCost spike \u2014 opus-4 on validate-guardrails
highTuesday's cost hit $18.20 \u2014 2.7\u00D7 average. The guardrails-taxonomy edit triggered 67 validate-guardrails calls on opus-4, each re-reading the changed canon. Root cause of 3 other anomalies.
Coherence dropped to 72%
highTwo CLIs now diverge on guardrails and adapters. The taxonomy edit wasn't propagated to GEMINI.md and .cursor/rules. Root cause: drift-check ran but coherence-check didn't fire.
Budget accelerated \u2014 7 days to exhaustion
mediumBurn rate doubled since Tuesday. At current pace, budget exhausts in 7 days. The taxonomy edit added 1,840 tokens loaded per validate-guardrails call. Pruning canon buys back 10 days.
\nToken flow \u2014 the taxonomy edit is visible
\nThe Sankey shows the cascade: guardrails-taxonomy.md (left, thick) now feeds validate-guardrails disproportionately. opus-4 (right) is fatter than sonnet \u2014 the anomaly.
\nToken flow \u00B7 today
38,200 tok routed \u00B7 96% of budget7 days to exhaustion \u2014 unless canon is pruned
\nBurn rate doubled since Tuesday. Three levers: prune buys +10 days, cache buys +5, doing nothing loses the budget in 7.
\nBudget burn-down \u00B7 14d actual + 7d projected
80% confidenceRecommended action
aih coherence --fixaih prune --applyThe Tuesday session that started it
\nReplay of the 14:22 session where the taxonomy edit cascaded. 11 validate-guardrails calls on opus-4 in one session \u2014 visible as the thick green band.
\n