fix(plugin): insights shows savings with the right sign

[amiddavid]

Summary

• /context-guru:insights now shows the right sign for every saving. Keep-alive pings show +$60 where they showed -$180.
• The lever rows now add up to the "Saved" headline.
• Fixes fix(plugin): insights shows the keep-alive saving with the wrong sign #582. This unblocks docs(readme): replace the plugin section with "What's cool" #579.

What changed

In /api/stats, a saving is a negative delta_usd and a cost is positive (dash/overview.go, waterfall()). levers() read each key like this:

Key	Sign in Go	levers() before	Now
compaction	saving, negative	negated (right)	same
cachesplit_saved	saving, negative	as is (wrong)	negated
keepalive_saved	saving, negative	as is (wrong)	negated
keepalive_ping	cost, positive	abs()	as is
decl_filter_saved	saving, negative	as is (wrong)	negated
cg_llm	cost, positive	-abs()	-value
self_removed	saving, negative, outside total_saved	as is (wrong)	row removed

• No blind abs(): a keep-alive that costs more than it saves still shows as a loss.
• self_removed is the account's own doing. The waterfall keeps it out of total_saved (it is in total_reduced). With it in the rows, they could not add up to the headline. Open question: do you want it back as a separate line, labelled as not ours?
• Same change in codex-marketplace/plugins/context-guru/scripts/cg.py. The two files stay byte-identical.
• Test fixtures (stats_*.json, test_tv.py) had positive savings. They now use the real sign. Goldens did not change.

How it was tested

• Two new tests in cg_test.go:
  • TestInsightsLeversAddUpToTotalSaved: every key non-zero, rows sum to total_saved.
  • TestInsightsKeepAliveLossStaysNegative: saved 10, spent 40 shows -30.
• Mutations, each caught:
  • keepalive_saved sign reverted (the old code): 4 tests fail, both new tests among them.
  • cachesplit_saved sign reverted: the sum test fails.
  • decl_filter_saved sign reverted: the sum test and the goldens fail.
  • abs() on the keep-alive: both new tests fail.
  • self_removed row restored: the sum test fails.
• Full go test ./context-guru-plugin passes on box 1 (318 s).
• Read-only check against the local proxy, 7 days: rows sum to $147.15 = total_saved. Keep-alive shows +$62.34 = keepalive_net_usd.

Fixes #582

🤖 Generated with Claude Code

[amiddavid]

Decision on the open question (from the repo owner): the self_removed row stays removed. What Claude Code removes by itself is not context-guru's saving, so insights does not show it.

[amiddavid]

Review, round 1

Verdict: approve.

What I checked

• Signs. I read waterfall() in dash/overview.go. Savings are -X (compaction, cachesplit_saved, keepalive_saved, decl_filter_saved, self_removed). Costs are +X (cg_llm, keepalive_ping). The new levers() matches every key.
• Net loss. No abs() is left. A keep-alive that costs more than it saves, or a cachesplit with a negative value, still shows as negative.
• Fixtures. The changed fixtures now match the real shape. The live /api/stats of the local proxy (read-only) has keepalive_saved -127.69, keepalive_ping +65.46, compaction -90.03, cg_llm +6.24.
• Sum. The live rows give 90.03 + (127.69 - 65.46) - 6.24 = 146.03. This equals total_saved.
• Other readers. I grepped every key in the plugin and in Codex. cg.levers() is the only reader of the waterfall. tv.py:185 calls cg.levers(), so it is fixed too. statusline.py, insights.py and the skills do not read the waterfall. No other reader has the defect.
• Copies. The two cg.py files are byte-identical.

Tests

• go test ./context-guru-plugin -run 'Insights|Levers|CG' on box 1: pass.
• Full go test ./context-guru-plugin on box 1: pass (313 s).
• test_tv.py run direct (pytest is not installed here): 36 tests, OK.
• My own mutations, one at a time, each caught:
  • cachesplit_saved sign flipped: sum test fails.
  • decl_filter_saved sign flipped: sum test and goldens fail.
  • keep-alive ping sign flipped: both new tests and goldens fail.
  • cg_llm sign flipped: sum test and goldens fail.
  • keepalive_saved sign flipped: both new tests and goldens fail.

self_removed

• The old row showed the waterfall value as is. That value is -SelfRemovedUSD, so it showed negative. The row was wrong, and it broke the sum.
• cg.py shows no other place for this number. total_reduced is not shown either.
• Removing the row is correct for the "Saved" table. I suggest a follow-up, not a change here: a separate line below the table, such as "You removed this yourself, not counted above: $X". It must stay outside the sum.

Non-blocking

• test_tv.py still has self_removed at 0, so no test checks that levers() ignores a non-zero self_removed. TestInsightsLeversAddUpToTotalSaved covers it on the Go side.

[amiddavid]

Review, round 1 (update)

• Withdrawn: my follow-up idea of a separate self_removed line. The row stays removed, as decided. Nothing blocks on it.
• Checked: no other insights output adds self_removed to a context-guru saving. cg.levers() and tv.py do not read the key. In insights.py it is a separate finding ("self-removed-credit", severity ok). It is used for ranking only and is never added to a saved total.
• Verdict: approve.