Token Burn

Burn is not proof. Output is not victory. Accepted work that repeats is leverage.

today · in flight
913K active
14 sessions · spend 24.1M · Claude Code 913K
by project: las 913K · <home> 0
top session 4b15a997 (las, 913K, —)
2026-09-06
as of 2026-09-06T13:00:21Z
Stupid button · 28m ago · gemma4:e4b · 3656t in / 944t out · 121.5s

Trend vs last week

What worked

What wasted tokens

Why the big number is big

A subagent spawn forks its parent, so the child starts holding the parent's whole transcript and the same conversation is charged once per fork, on every turn. The fork_amplification entry shows root: "01a04b01-5edc", 6 threads, 3 forks, max\_depth 1, charged\_h "33.8M", replay\_pct 94.3, output\_h "124K", cost\_per\_output 273.5, and was associated with agents: "Chandrasekhar", "Faraday", "Harvey", "James", "Pascal".

What needs a verdict

One contrarian observation

The <home> directory has 1.5M waste\_active\_h, but its acceptance\_ratio is 0.0 because it is unmeasurable, as noted in its acceptance\_note.

---

Measurability caveat (generated, not model-authored): <home> (1.5M) ran outside any git repo. There is no diff to credit, so a 0.0 acceptance ratio there is the meter failing to see the work — not a judgement on it. Only ratios on measurable: true projects are verdicts.

Rollups — acceptance lane

Headline is accepted tokens (active lane, manual label wins over git auto-classification). Unsettled = committed-not-merged + open-loop. "Active" excludes cache replay; "spend" includes it.

Last 7 days
10.3M accepted
unsettled 2.0M · acceptance 75%
spend 357.5M · 167 sessions · Claude Code 12.7M / Codex 1.0M
Last 30 days
16.2M accepted
unsettled 71.8M · acceptance 18%
spend 2.40B · 745 sessions · Claude Code 64.1M / Codex 27.4M
Last 90 days
16.2M accepted
unsettled 160.4M · acceptance 9%
spend 5.29B · 910 sessions · Codex 97.0M / Claude Code 83.1M

Daily heatmap — last 60 days

Each cell = one UTC day, color intensity = log10(active tokens). Hover for date + counts.

LessMore· log-scale, exact lane
29.1M on 07-28

Top 10 days

DateActiveSpend SessionsTop projectOutcomes
2026-07-2829.1M1.09B22<home> (29.1M)?=22
2026-08-1822.3M554.5M38<home> (21.6M)?=43
2026-08-0616.5M592.1M51<home> (12.0M)?=58
2026-07-1615.7M522.1M23<home> (15.7M)?=23
2026-08-127.7M116.0M25dreamweaver (6.5M)?=27
2026-08-086.1M358.6M27dreamweaver (3.0M)?=37
2026-08-076.0M352.8M28<home> (4.1M)?=35
2026-08-315.9M144.8M26las (5.8M)no_repo=24 · open_loop=3 · accepted=2 · PRs #18
2026-06-155.9M132.6M9<home> (5.9M)?=9
2026-08-195.6M205.7M27pal (2.3M)?=36

By project — last 7 days

ProjectActive 
las10.9M
<home>1.5M
memory1.1M
pal144K
pal/templari61K
mainstay-av30K

By outcome — last 7 days

accepted 7 #17 (MERGED), #19 (MERGED), #18 (MERGED)
committed 1 #21 (OPEN)
open_loop 10
no_repo 165

Effectiveness — last 30 days

Where delegated work shipped (top) vs. where high burn never crossed an acceptance line (bottom). Bars: accepted · unsettled · waste. Normalized to the widest project in window.

Leverage — most accepted

ProjectAcceptedRatioMix
las15.0M79%
memory1.1M56%
pal00%
mainstay-av00%
pal/templari00%
templari00%

Waste candidates — lowest acceptance ratio (≥100K active)

ProjectAcceptedRatioMix
<home>00%
dreamweaver00%
pal00%
mainstay-av00%
pal/placeholders00%
pal/dashboards00%

Needs validation — last 14 days (20 unsettled)

Sessions where auto-classification couldn't tell shipped work from waste. Highest-burn first. Manually label these with `label: "used"|"partial"|"redo"|"review"|"not_useful"` in token_burn.jsonl to override the auto verdict.

DateProjectSessionActiveAutocwd
2026-08-24mainstay-av5af22ad71.8Mopen_loop.../mainstay-av/docs/entity
2026-08-29mainstay-av5af22ad71.6Mopen_loop.../mainstay-av/docs/entity
2026-09-06las4b15a997913Kcommitted · #21 (OPEN).../las/.worktrees/mvp
2026-08-28lasd33e7f08689Kopen_loop.../las
2026-09-05las4b15a997645Kopen_loop.../las/.worktrees/mvp
2026-09-03las4b15a997165Kopen_loop.../las/.worktrees/mvp
2026-08-24lasd33e7f08146Kopen_loop.../las
2026-08-31pal/templarid33e7f0861Kopen_loop.../pal/templari
2026-09-05pal01a0734757Kopen_loop.../pal
2026-09-04mainstay-avd33e7f0830Kopen_loop.../mainstay-av
2026-08-24mainstay-avd33e7f0828Kopen_loop.../mainstay-av/apps/worker
2026-08-24pald33e7f0827Kopen_loop.../pal
2026-08-24templarid33e7f0827Kopen_loop.../templari
2026-09-01pal5af22ad724Kopen_loop.../pal
2026-08-31pal5af22ad723Kopen_loop.../pal
2026-09-04pald33e7f0817Kopen_loop.../pal
2026-09-05pal01a0728d13Kopen_loop.../pal
2026-08-31pald33e7f0810Kopen_loop.../pal
2026-08-28pal/templarid33e7f084Kopen_loop.../pal/templari
2026-08-28pald33e7f082Kopen_loop.../pal

Recent events — last 30 rows

Latest enriched rows from the JSONL (sorted by date desc, active desc as tiebreaker). M = manually labeled.

DateToolProjectSessionActiveOutputOutcome
2026-09-06CClas4b15a997913K101Kcommitted #21
2026-09-06CC<home>055986cf00no_repo
2026-09-06CC<home>25fcf28d00no_repo
2026-09-06CC<home>436555db00no_repo
2026-09-06CC<home>4ef5621b00no_repo
2026-09-06CC<home>5b5e8c1300no_repo
2026-09-06CC<home>7c062c9300no_repo
2026-09-06CC<home>90e18de400no_repo
2026-09-06CC<home>b32d95d200no_repo
2026-09-06CC<home>c2facd4300no_repo
2026-09-06CC<home>ce66ef1000no_repo
2026-09-06CC<home>d154575600no_repo
2026-09-06CC<home>d190593b00no_repo
2026-09-05CClas4b15a997645K12Kopen_loop
2026-09-05Codex<home>01a07287139K12Kno_repo
2026-09-05Codex<home>01a07329128K6Kno_repo
2026-09-05Codex<home>01a0732b57K4Kno_repo
2026-09-05Codexpal01a0734757K1Kopen_loop
2026-09-05Codex<home>01a04b0147K88no_repo
2026-09-05Codexpal01a0728d13K7open_loop
2026-09-05CC<home>0a19a29900no_repo
2026-09-05CC<home>103ed8a500no_repo
2026-09-05CC<home>19f6220600no_repo
2026-09-05CC<home>2dec65eb00no_repo
2026-09-05CC<home>3e9a31f700no_repo
2026-09-05CC<home>45e0475100no_repo
2026-09-05CC<home>5379f2b800no_repo
2026-09-05CC<home>6067a19e00no_repo
2026-09-05CC<home>6290217c00no_repo
2026-09-05CC<home>7a0b538b00no_repo

Generated 2026-09-06 13:00 UTC from ~/ideas/pal/dashboards/token_burn.jsonl (1041 rows). Sources: ~/.codex/sessions/ (exact), ~/.claude/projects/-home-khoward/ (exact). No estimate lane currently active.