{ "title": "Pocket Ledger: normal Codex versus the LeanTask launcher", "date": "2026-09-21", "publisher": "Thinkelution", "design": "A deliberately created sample app-building task, one measured run per approach; not a customer production deployment or a statistical savings estimate.", "model": "gpt-6-astra", "reasoning": "high", "codex_version": "codex-cli 0.153.4", "leantask_version": "0.3.0b6", "leantask_source_commit": "7cb9ca66de09a79b33b89a5920e920cccbd88410", "platform": "macOS", "python_version": "3.13.13", "node_version": "25.2.1", "run_order": [ "leantask", "baseline" ], "order_random_seed": 21, "timeout_seconds_per_run": 900, "sandbox": "workspace-write", "fresh_conversations": true, "identical_starting_fingerprints": true, "configuration": "Both CLI runs used --ignore-user-config and explicit model/reasoning, retaining existing policy rules. LeanTask added its source launcher guidance and five task-scoped MCP tools; no global companion-plugin installation.", "original_prompt_sha256": "2acda5264fc8a962acbecf5c2ba6287f386f073578e1e8f8275cbc79bc630e5d", "original_prompt_characters": 2227, "fixed_acceptance_tests": 23, "profiles": { "baseline": { "usage": { "available": true, "failed_event": false, "turns": 1, "input_tokens": 435704, "cached_input_tokens": 396800, "output_tokens": 26973, "total_tokens": 462677, "uncached_input_tokens": 38904 }, "elapsed_seconds": 857.314, "execution_status": "completed", "mcp_calls": {}, "native_command_count": 13, "raw_turn_usage": [ { "input_tokens": 435704, "cached_input_tokens": 396800, "cache_write_input_tokens": 0, "output_tokens": 26973, "reasoning_output_tokens": 1686 } ], "patch_verification_failures": 1, "verification": { "fixed_acceptance_tests_passed": 23, "fixed_inputs_unchanged": true, "javascript_syntax_passed": true, "independent_browser_checks": { "empty_state": "pass: initial $0.00 and no entries", "add_and_totals": "pass: $12.50 Food + $5.00 Travel = $17.50, count 2", "date_order": "pass: Sep 3 Bus before Sep 2 Tea", "search": "pass: trimmed case-insensitive bus search gives $5.00/count 1", "combined_filters": "pass: Food and September gives Tea/$12.50; nonmatching combination gives empty state", "delete": "pass: delete Tea reduces stored count and survives reload", "persistence": "pass: added entries survive reload", "inline_error": "pass: amount 0 rejected with helpful visible alert; total unchanged", "safe_text": "pass: HTML-like description renders as literal text in browser", "responsive": "pass: 375px viewport reports page width 375px; mobile and 1280px layouts visually inspected", "csv": "Export CSV button exercised with Travel filter; downloaded file not independently verified. CSV function checked by fixed acceptance tests.", "console": "no warnings or errors reported by browser during checks", "malformed_storage": "source reviewed; not independently exercised in the real browser" } } }, "leantask": { "usage": { "available": true, "failed_event": false, "turns": 1, "input_tokens": 136659, "cached_input_tokens": 111488, "output_tokens": 11163, "total_tokens": 147822, "uncached_input_tokens": 25171 }, "elapsed_seconds": 356.952, "execution_status": "completed", "mcp_calls": { "tasklean_read": 3 }, "native_command_count": 6, "raw_turn_usage": [ { "input_tokens": 136659, "cached_input_tokens": 111488, "cache_write_input_tokens": 0, "output_tokens": 11163, "reasoning_output_tokens": 427 } ], "patch_verification_failures": 0, "verification": { "fixed_acceptance_tests_passed": 23, "fixed_inputs_unchanged": true, "javascript_syntax_passed": true, "independent_browser_checks": { "empty_state": "pass: initial $0.00 and no entries", "add_and_totals": "pass: $12.50 Food + $5.00 Travel = $17.50, count 2", "date_order": "pass: Sep 3 Bus before Sep 2 Tea", "search": "pass: trimmed case-insensitive bus search gives $5.00/count 1", "combined_filters": "pass: Food and September gives Tea/$12.50; nonmatching combination gives empty state", "delete": "pass: delete Tea reduces stored count and survives reload", "persistence": "pass: added entries survive reload", "inline_error": "pass: amount 0 rejected with helpful visible alert; total unchanged", "safe_text": "pass: HTML-like description renders as literal text in browser", "responsive": "pass: 375px viewport reports page width 375px; mobile and 1280px layouts visually inspected", "csv": "app reports Exported 1 expenses as CSV; browser download event timed out, so downloaded file not verified. Fixed CSV function tests pass.", "console": "no warnings or errors reported by browser during checks", "malformed_storage": "source reviewed; not independently exercised in the real browser" } } } }, "observed_tokens_fewer_with_leantask": 314855, "observed_token_reduction_percent": 68.05, "separate_optimizer_inference_tokens": 0, "launcher_added_prompt_characters": 764, "leantask_tool_character_diagnostics": [ { "operation": "command", "calls": 1, "source_chars": 1170, "returned_chars": 1730 }, { "operation": "read", "calls": 3, "source_chars": 3336, "returned_chars": 4059 } ], "interpretation": "Observed difference only; not a causal attribution to compression or a general savings rate. Small file reads and the captured log grew with metadata.", "confounders": [ "Normal Codex retried one invalid patch.", "Normal Codex attempted additional browser checks blocked by the subprocess sandbox; both runs later reported passing simulated-DOM checks.", "Single pair, no repetitions or statistical uncertainty estimate; model variability, cache behavior and independent execution choices were not isolated." ], "accounting": "Input plus output. Cached input is a subset of input; reasoning output is not added again. Token counts are not dollar costs or subscription allowance.", "excluded_scope": "Coordinating conversation, fixture creation, independent verification and report preparation.", "rejected_setup_attempts": { "count": 2, "model": "gpt-5.4", "reason": "Account rejected the configured model before code generation.", "usage": null, "treatment": "Unknown, not zero. Excluded from the successful-model pair; total account spend is not established." }, "evidence_scope": "Curated usage and verification record. Full prompts are supplied separately; full execution transcripts, local paths and account identifiers are not published.", "fixture_sha256": { "package.json": "fbe3cf224544565e9f93d1da093aa85a3ded4b5d5c6f06e6adb544dbda142bb7", "README.md": "f82d4679dbdb98075b899f0e8780ac641a7e9cb749855d80f45398153400264c", "test/ledger.test.js": "79148667a17e1b6d74224b43a44464c6a0b57d31b8b95988bce4d8de770bf1c6" }, "sandbox_retry_adjustment": { "kind": "Accounting-only subtraction from a measured run; not a sandbox-free rerun or a prediction.", "source": "Detailed normal-run session token_count records, using last_token_usage per model request.", "all_request_records_count": 13, "all_request_usage_reconciles_with_completed_run": true, "reconciled_fields": [ "input_tokens", "cached_input_tokens", "output_tokens", "total_tokens" ], "excluded_requests": [ { "request_index": 9, "usage_recorded_at_utc": "2026-09-21T12:24:42.407Z", "usage": { "input_tokens": 41350, "cached_input_tokens": 37888, "cache_write_input_tokens": 0, "output_tokens": 236, "reasoning_output_tokens": 51, "total_tokens": 41586 }, "description": "Retry browser verification through local-file access after the temporary HTTP server was denied." }, { "request_index": 10, "usage_recorded_at_utc": "2026-09-21T12:24:52.762Z", "usage": { "input_tokens": 41678, "cached_input_tokens": 41216, "cache_write_input_tokens": 0, "output_tokens": 192, "reasoning_output_tokens": 38, "total_tokens": 41870 }, "description": "Adjust browser-process cleanup and retry local-file verification after the previous attempt failed." } ], "excluded_usage": { "input_tokens": 83028, "cached_input_tokens": 79104, "output_tokens": 428, "total_tokens": 83456 }, "adjusted_normal_codex_usage": { "input_tokens": 352676, "cached_input_tokens": 317696, "output_tokens": 26545, "total_tokens": 379221, "uncached_input_tokens": 34980 }, "leantask_measured_total_tokens": 147822, "numerical_gap_tokens": 231399, "numerical_gap_percent": 61.02, "retained": [ "Initial browser verification attempt", "Successful fallback simulated-DOM checks", "Patch retry", "Final checks and answer", "All remaining requests at their original measured context and cache usage, including any failure-related text" ], "limitations": [ "A successful browser check would return different output and could change subsequent actions.", "This does not remove all downstream effects of the failures or model the alternate successful path.", "The adjusted comparison is not a measured savings rate, a causal estimate of LeanTask benefit, or an upper/lower bound on a sandbox-free run.", "No adjusted elapsed time, dollar cost or subscription allowance is inferred." ] }, "recurring_work_hypothesis": { "claim": "Repeated work on the same task can offer greater opportunities for token savings when focused reads, saved notes and retrievable logs avoid redundant context.", "evidence_status": "Expected use case; not measured in this single app-building pair.", "qualification": "Ordinary Codex also supports continuation and efficient reads. Actual savings depend on reuse and context avoided; growing history and tool overhead can offset benefits." } }