Codex Weekly Usage — August 3–9, 2026
An integration-heavy week paired Hyphenomenon publishing and design work with sustained multi-mailbox operations and Shopify delivery: 34 de-duplicated sessions across 82 active rollout logs, verified delivery traces for four repositories, and a Friday–Saturday model-context concentration that accounted for 63.5% of processed tokens.
Template-generated · Analyzer: codex-usage-report-v6 · Model: gpt-5.6-sol · high
Input plus output tokens in retained local Codex telemetry; includes cached input and is not a cost or productivity score.
Top-level rollout files active in the window; delegated rollouts are counted separately.
Subagent or spawned-thread rollout files with timestamped activity in the window.
Recorded turn contexts across top-level and delegated rollouts; not elapsed work time.
Token Usage
Cumulative runtime counters are differenced per rollout. Repeated snapshots are ignored; cached input stays visible instead of being mistaken for fresh context.
Token Usage
Retained local Codex telemetry only; not an account-wide usage or billing statement.
- Input
- 3,879,675,815
- Cached input
- 3,763,420,544
- Non-cached input
- 116,255,271
- Output
- 10,271,877
- Reasoning output
- 3,051,301
Cached input is 97% of input. Reasoning output is included within output tokens.
Tokens By Day
Processed input plus output assigned to the local day of each positive cumulative-counter increment.
Activity By Day And Clock Hour
The day row shows each date in the report window. The hour chart below combines the same clock hour across all seven days; it is not a single-day timeline.
- Mon Aug 313
- Tue Aug 42
- Wed Aug 51
- Thu Aug 64
- Fri Aug 77
- Sat Aug 84
- Sun Aug 93
Each count is a de-duplicated session start assigned to that local calendar day.
Total: 18,215 tool calls. “c” marks successful commit operations in that hour.
Where The Work Landed
Repository activity and changed lines show which codebases received attention during the report window.
Tool Calls By Repository
Working repository resolved from tool workdir or turn context.
Output Shape
Compare the rollouts carrying the largest patches with the mix of file categories touched.
File Touch Mix
Verified Delivery Ledger
Outcomes come from terminal return records. Intent-only command counts remain in the technical appendix.
Command Outcomes
Repositories In Successful Push Output
Agent Stack Used
Observed runtime evidence only; the available catalog is excluded. Status chips distinguish completed calls, skill loads, and explicit requests.
- Invoked means a completed runtime call was recorded.
- Loaded means a skill file was read or discovered.
- Requested marks an explicit skill request in the source.
- The logs do not preserve a reliable causal chain between every layer, so the diagram shows observed layers rather than inventing one.
Apps and Connectors
Completed connector calls recorded by the runtime.
Tool Providers
Completed MCP provider calls, grouped by server namespace.
Skills
Explicitly requested or loaded skills; application is not inferred.
Chronicle Context Across The Week
These panels are calculated directly from persisted Chronicle summaries. Lane counts overlap and are not duration.
Work Lane Mentions
Evidence Coverage
Technical Telemetry
Raw tool, event, workspace, and patch diagnostics used to audit the analyzer. These counts are intentionally secondary to outcomes.
Tool Surfaces
Function Calls
Custom Tool Calls
Specific custom calls and their share of all recorded tool events.