Session audit — 2026-09-04

04 Sept 2026

← Back to the curated DOs and DON’Ts

• projectsDir: /Users/foyzul/.claude/projects
• sessions audited: 854 (9 active excluded)
• date window: 2026-07-12 → 2026-09-04

Headline

854 sessions · 28,237 turns · cacheRead 3989.1M · cacheCreation 47.9M · hit ratio 0.988

findings: 916 · headline waste 25,869K tokens ≈ $56.85

NO_SUBAGENT carries a separate counterfactual of 44,256K ≈ $73.04 — it is not summed into the headline (different counterfactual: delegate-the-phase rather than read-less).

Unpriced-model share: 17.2% (8 local/cheap models — MiniMax-M2.7, MiniMax-M3, Ornith-1.0-9B-4bit, Ornith-1.0-9B-oQ4-fp16, Ornith-1.5-9B, Ornith-1.5-9B-MLX-4bit, Qwen3.8-27B-OptiQ-4bit, gpt-oss-20b-MXFP4-Q8). 82.8% of waste carries a price. Dollar and token totals are a floor: thinking tokens and compaction events are not measured by any rule.

Findings by rule (headline)

NO_SUBAGENT       ██████████████████████████████████████████████████  41 findings / 41 sess / $73.04  (separate counterfactual)
CACHE_TTL_EXPIRY  █████████████████████████                           54 findings / 36 sess / $35.26
DUP_TOOL_CALL     ████████████                                       652 findings / 156 sess / $14.01
BIG_TOOL_OUTPUT   ██████                                             71 findings / 66 sess /  $7.58

(bar widths scaled to NO_SUBAGENT, the largest rule — CACHE_TTL_EXPIRY’s bar shows its share of the headline ratio, not the absolute.)

Cache-hit ratio distribution

>=0.95         ██████████████████████████████████████████████████  155 sess  (healthy)
0.50-0.90      ██████████████████████████████                       96 sess  (warning)
0.90-0.95      ██████████████████                                  55 sess
<0.50          ████                                                16 sess

532 sessions have no cache and are excluded from the histogram (local/free models that bypass caching). Six of the costliest low-ratio sessions are NOT flagged by CACHE_MISS_RATE because they exceed the entry threshold; they are listed in the worst-sessions table below.

Projects by waste (dollars)

personal-claude-lens              ██████████████████████████████████████████████████  212 sess / $16.17
personal-fine-tuning              ██████████████████████████                         28 sess /  $8.20
personal-aswe-lms                 ████████████████████████                          130 sess /  $7.76
personal-agentic-vod              ██████████████                                    91 sess /  $4.46
al-claude-lens--worktrees-36      █████████                                          4 sess /  $2.92
al-claude-lens--worktrees-41      ████████                                           6 sess /  $2.73
al-claude-lens--worktrees-37      ████████                                           6 sess /  $2.48
al-claude-lens--worktrees-35      ████                                               5 sess /  $1.36

personal-fine-tuning ranks 3rd by tokens (1,868K) but 2nd by dollars ($8.20) — a 3-5× per-token premium from Opus. The dollar order is the order to act on; the worktree projects share a single codebase but are ranked separately because each worktree path is its own session namespace.

Worst sessions by waste

429adc51 (personal-fine-tuning) is the worst by dollar-per-token ratio at $5.00 per K tokens — about 5× the directory average — driven by 8 user_idle TTL gaps in a single 146-turn session where the user paused to read and ask follow-ups.

Idle-gap cost curve (raw cache-creation tokens)

Total raw excess 12,567K = 14,453K cost-equivalent (compared with CACHE_TTL_EXPIRY waste of 11,409K — the gap curve and the rule corroborate each other). 52 of 54 TTL findings carry gapKind: user_idle (96%); only 2 are tool_runtime (priced at zero by design).

Trend by date

2026-07-12→08-08  ██████████████████████████████████████████████████  231 sess / 19,118K waste / 0.70% rate
2026-08-09→09-04  ██████████████████                                94 sess /  6,751K waste / 0.55% rate

The window is split at the median session date. Waste rate dropped from 0.70% to 0.55% — a ~21% relative improvement, but most of that drop is the smaller read volume in the later window (1,238M vs 2,750M). Token-wise the per-turn rate is roughly steady.

Ranked fixes

#1 · TTL expiry after user_idle gaps — $35.26 / 11,409K (62% of headline)

• Evidence: 52 of 54 TTL findings are user_idle; idle >30m shows a 94.1× amplification in cache-creation per turn. Worst single session: 449adc51-bff4-4b1f-9092-753a7dd1c328 (personal-fine-tuning, $6.28) has 8 user_idle gaps — three of them >1 hour — in one 146-turn fine-tuning adapter session. User said “I actually have almost zero idea about how to create a fine tune adapter, let alone the other keywords. if possible, prepare your own recommendations options and use them” — they paused to read between each prompt.
• Inspect: node {base_directory}/bin/audit.js fetch 449adc51-bff4-4b1f-9092-753a7dd1c328 --kind user_text --limit 5 --max-bytes 500
• Cost: $35.26 / 11,409K tokens (62% of headline waste, 100% of which is user_idle).
• Attribution: habit — running Claude for long stretches across multiple thinking pauses without /clear.
• Target: by the next audit, CACHE_TTL_EXPIRY should drop under $15 / 5,000K (the gap rule fires on >5min idle, but the curve shows the >30m bucket alone is 6,835K raw).

#2 · DUP_TOOL_CALL on Read — $11.66 / 7,660K (21% of headline)

• Evidence: Read alone accounts for 295 of 652 DUP findings and 7,660K of 9,178K wasted tokens. Worst single session: 8e689d85-7c4e-4a8e-ab83-b7888a02355f re-reads the same 483K bytes 11 times in one session (592K waste). 4a5aac31-3f1c-4762-8896-a8be5693bd1c (personal-claude-lens, $5.70) reads the same 100K file 4× and the same 73K file 7× — total 1,290K of waste from re-reads in one session, while the user asked it to “invoke dev-pipeline:review skill” and then “resume all subagents”.
• Inspect: node {base_directory}/bin/audit.js fetch 4a5aac31-3f1c-4762-8896-a8be5693bd1c --kind user_text --limit 3 --max-bytes 500
• Cost: $11.66 / 7,660K (21% of headline).
• Attribution: habit — context_growth compounds and the agent re-Reads files it already saw rather than trusting its context.
• Target: by the next audit, DUP waste should fall under $7 / 4,500K; the 8e689d85 outlier (11× same 483K read) should not recur.

#3 · NO_SUBAGENT in marathon sessions — 44,256K / $73.04 (separate counterfactual, 1.3× headline)

• Evidence: 41 sessions over 73 with NO_SUBAGENT findings. Top: 7302d6bc-c494-4023-be03-a443fb6acf77 did 93 scan calls / 468K bytes inline (6,074K waste). Sessions >100 turns with no subagent: 55 of 81 (68%). Top NO_SUBAGENT sessions overlap the top CONTEXT_GROWTH peaks (7302d6bc, a52bf19a, 3ae6d031, 4a5aac31) — the agent absorbs all the scan work into its main loop, which compounds with TTL and DUP waste.
• Inspect: node {base_directory}/bin/audit.js fetch 7302d6bc-c494-4023-be03-a443fb6acf77 --kind user_text --limit 3 --max-bytes 500
• Cost: 44,256K / $73.04 — equal in magnitude to the entire rest of the headline combined. This is NOT summed into the headline per spec.
• Attribution: skill_file — dev-pipeline:plan-architecture should mandate subagent delegation in Phase 1 when scope exceeds a threshold.
• Target: by the next audit, the share of sessions >100 turns with NO_SUBAGENT should fall under 50% (currently 68%).

#4 · BIG_TOOL_OUTPUT on Read — $7.58 / 5,279K (13% of headline)

• Evidence: 69 of 71 BIG findings are Read; sample d4a968ab Read 38,629 bytes for a 8,629-byte excess over the 30KB threshold. Read is the largest tool by bytes-into-context (18.47MB across 3,467 calls).
• Cost: $7.58 / 5,279K (13% of headline).
• Attribution: skill_file — dev-pipeline:implement should recommend offset/limit Reads for files >500 lines.
• Target: by the next audit, BIG waste should fall under $3 / 2,000K; the threshold should move from 30KB to 20KB once this is achieved.

#5 · Opus per-token premium on personal-fine-tuning — $8.20 from 28 sessions

• Evidence: personal-fine-tuning ranks 3rd by tokens (1,868K) but 2nd by dollars ($8.20) — Opus prices each token at 3-5× directory average. Worst session 449adc51 (above) is at $5.00 per K tokens.
• Cost: $8.20 — 14% of headline by dollars, only 7% by tokens.
• Attribution: config — model default for fine-tuning and routine implementation tasks should be Sonnet; switch up only for the architecture parts.
• Target: by the next audit, the per-K price for personal-fine-tuning should drop to ~$2.50 (Sonnet-equivalent), reducing project waste to ~$4.

#6 · personal-claude-lens concentration — 7,183K / $16.17 (28% of headline)

• Evidence: 212 sessions / 6,060 turns / 7,183K waste; hosts the longest sessions in the directory (294 turns in 4a5aac31, 698 turns in 7302d6bc). These overlap with NO_SUBAGENT (#3) and TTL (#1) — fixing #1 and #3 cuts this project’s waste substantially.
• Cost: $16.17 — 28% of headline.
• Attribution: habit — long single-session runs that should have used dev-pipeline:start-task to spawn a feature branch with its own conversation.
• Target: by the next audit, personal-claude-lens should fall under $10 ($16.17 → $10).

Do this first

#3 NO_SUBAGENT is the highest-leverage single fix by sheer magnitude ($73.04 — equal to all the rest combined) and is the only one reachable via a skill_file edit that keeps paying. The TTL/DUP fixes (#1, #2) require indefinite discipline; the Opus fix (#5) is a config change but its dollar savings depend on whether the model choice is genuinely negotiable.

Pick: edit dev-pipeline:plan-architecture/SKILL.md to add a Phase-1 rule: “if the spec mentions ≥5 files or ≥8 distinct tasks, spawn a research subagent before drafting the plan; cite its findings in the plan.” That single rule converts the worst NO_SUBAGENT pattern (93 inline scan calls, 468K bytes, 6,074K waste) into a delegated scan that holds its own context. Cost: one file edit. Permanence: indefinite.

Then #1 TTL (CLAUDE.md already documents /clear-after-break; the audit shows it isn’t being applied — add a one-line reminder at the top of global CLAUDE.md).

Trend

First audit — no baseline yet (~/.claude/audit-reports/ was empty before this run). The within-run date split (Jul 12–Aug 8 vs Aug 9–Sep 4) shows waste rate dropping from 0.70% → 0.55%, but the later window is half the volume of the earlier one, so the rate drop is largely a denominator effect. Next audit’s headline number is what will tell us if the trend is real.

Audit self-cost

• fetch_log.jsonl (lower bound — escalation fetches only; excludes Phase 1 landscape and own reasoning turns): 6 fetches, ~2,100 bytes
• 5 user_text fetches at --max-bytes 400 (well under the 2,000-byte per-fetch cap) + 1 narrow aggregate query.

Stats this report wanted and could not get

• Skills breakdown by waste contribution: views prints skills by use count, not by waste. A skill that is invoked often on a small session is reported alongside one invoked rarely on a marathon session — they look equal. Would need a per-session join from overview.json to compute, and I did not do that to keep the report transcript cheap (I8).
• Per-session subagent savings attribution: NO_SUBAGENT’s $73.04 is a capacity counterfactual, not an observed spend. There is no per-session view of how much a delegated scan would have cost; the $73.04 is computed from the bytes the main loop did pull, not from a comparable delegation that ran. Treat as an upper bound on what is recoverable from #3.

• ai
• agentic
• workflow
• efficiency