PR #85: feat(skills): add /cost-impact for weekly Claude spend reports

• Repository: idvorkin/chop-conventions
• Author: idvorkin-ai-tools
• State: closed | merged: True
• Link: https://github.com/idvorkin/chop-conventions/pull/85

Description (problem / solution / changelog)

Summary

New `/cost-impact` skill that computes a Claude Code cost-impact report across a time window (default 7 days), grouped by repo with per-session detail. Complements `/changelog` — the latter summarizes git history, this one summarizes Claude session history, answering "where did my Claude-time actually go this week."

What it does

Scans every `~/.claude/projects//.jsonl` main session and `//subagents/agent-*.jsonl` subagent transcript, filters turns by timestamp into the requested window, and bills each turn at its model's published list price. Key design points:

• Per-turn per-model pricing — Opus 4.6 $5/$25, Sonnet 4.6 $3/$15, Haiku 4.5 $1/$5, older Opus 4 / 4.1 tiers at their historic $15/$75. Table sourced from platform.claude.com/docs/en/about-claude/pricing.
• Subagent tokens roll into the parent session's total — so "what did this session cost" reflects all work done on its behalf, not just the main turns.
• Groups by (project, parent_session_uuid, day) — a session that spans midnight is split across days so daily totals add up cleanly.
• Flat pricing across the 1M window for 4.5+ — no 200k threshold, per Anthropic's long-context docs.

Report shape

1. Summary — total $, total duration, total turns (main/sub/total), sessions in window, cost breakdown (input/output/cache writes 1h+5m/cache reads), without-cache reference, cache savings %
2. Cost by model — per-model totals and share
3. Per day — sessions, main+sub turns, costs
4. Per repo summary table — sessions, actual $, share %
5. Sessions grouped by repo — collapsible `<details>` sections, one per repo. Summary line shows repo name, total $, session count, duration, main+sub turns. Expands to a per-session table with clickable PR links extracted from `gh pr create` calls in the scrollback.
6. Footnotes — methodology, fast-mode/peak-hours/TTL-bug caveats

Invocation

```bash python3 skills/cost-impact/_impl.py [days] ```

The skill prompts the user whether to post the report to a gist or save locally, then runs the right `gh gist create` or `cp` command.

Dogfood test results

Ran against my own 7-day window:

``` Wrote /tmp/cost-impact.md (16,937 bytes, 64 session-days) Actual: $1,248.28 | no-cache ref: $6,960.82 | savings: 82% Repos: 11, top 3: activation-energy-game, settings, blog4 ```

Also ran with `--days 1` to verify the short-window path ($331, 5 repos, 11 session-days).

Known caveats (surfaced in the report's own footnotes)

• Fast mode invisible: `usage` field doesn't flag `/fast` turns, so if the user hit fast mode those are under-reported by 6×. No current detection path.
• Peak-hours quota burn (weekday 5–11am PT, per Anthropic session limits update 2026-03-26) only affects quota burn rate, not $/token. The report prices at list; peak vs off-peak is not reflected in the dollar column.
• TTL bug (anthropics/claude-code#45381): if `DISABLE_TELEMETRY` or `CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC` is set, sessions silently fall from 1h to 5m cache tier. Footnote alerts the user.

Test plan

• Dogfood with 7-day window — produces coherent $1,248 report
• Dogfood with 1-day window — short-window edge case works
• Verify collapsible `<details>` sections render on GitHub (tested on a gist of the output)
• Verify PR link extraction finds `gh pr create` commands and shows titles with state
• All pre-commit hooks pass (ruff lint/format, prettier, fast tests)

Skill structure

• `skills/cost-impact/SKILL.md` — entry point with trigger phrases, run/output instructions, hard rules, known caveats
• `skills/cost-impact/_impl.py` — standalone Python (~370 lines), safe to run directly: `python3 skills/cost-impact/_impl.py [days]`

🤖 Generated with Claude Code

Changed files

• skills/cost-impact/SKILL.md (added, +166/-0)
• skills/cost-impact/_impl.py (added, +655/-0)
• skills/cost-impact/test_impl.py (added, +436/-0)

---

PR #90: docs(CLAUDE.md): don't hardcode factual claims in report output

• Repository: idvorkin/chop-conventions
• Author: idvorkin-ai-tools
• State: closed | merged: True
• Link: https://github.com/idvorkin/chop-conventions/pull/90

Description (problem / solution / changelog)

Summary

Adds a short rule adjacent to Diagnostics: Code Over Prose warning against hardcoded narrative footnotes in script output.

Why

LLM-written report code naturally reaches for claims that sound measured — "0 X observed", "no errors found" — but are actually string literals that rot silently and can flip from true to false without anyone noticing.

Concrete example from #85: the `skills/cost-impact/_impl.py` TTL-bug footnote unconditionally claimed `0 ephemeral_5m_input_tokens observed`. Real data over a 7-day window was 10,531,293 c5m tokens / $57.70 — the hardcoded "reassurance" was silently lying to the user about whether they were affected by anthropics/claude-code#45381. Copilot flagged it in review.

The existing Diagnostics: Code Over Prose rule covers skill prose going stale. This is related but distinct: it covers script output asserting facts the script never computed. Both are flavors of "derive, don't declare", but in different surfaces.

The rule

Factual claims in generated output must be computed from the actual data, not hardcoded as constants. [...] If a report asserts a number or category, derive it from the input; if you can't derive it, drop the assertion or phrase it as a question the reader must answer, not a fact.

Test plan

• Fits under the 5-line cap
• Concrete example cited (cost-impact TTL footnote)
• Durable, non-obvious, actionable — not already covered by existing rules
• Distinct from Diagnostics: Code Over Prose (which is about skill prose, not script output)

🤖 Generated with Claude Code

Changed files

• CLAUDE.md (modified, +4/-0)