Usage is very high (#18)

[ableinc]

Closes #18

What the agent did

This matches the plan's file list exactly. I'll leave version control to the harness as instructed (not committing).

Summary

Implemented the approved plan for issue #18 (cut per-run token usage) across three parts:

Part A — cache reads reported separately from fresh input

• internal/claude/runner.go: added FreshTokensIn(), CacheWriteTokens(), CacheReadTokens() accessors.
• internal/store/store.go: new migration (entry 3) adding tokens_cache_read/tokens_cache_write columns; Run, runColumns, scanRun updated; RecordUsage now takes a RunUsage struct instead of 7 positional args.
• Four call sites updated (loop.go, prcomments.go, both RecordUsage calls plus claude_done event details) and two in store_test.go, plus a new migration test TestMigrationAddsCacheTokenColumns.
• internal/web/assets/app.js: new fmtTokensBreakdown helper, used in the runs table and run detail view.
• internal/discord/notifier.go: Tokens embed field now shows the fresh/written/cached split.

Part B — trim the harness prompt (internal/orchestrator/prompt.go)

• Tightened constants (maxBodyChars 12000→6000, maxCommentChars 2000→1200, maxCommentsInclu 12→6), added maxPlanChars, maxPRCommentsInclu, maxDiffHunkChars, maxReviewsInclu.
• issueContext now filters out harness-authored comments (isAgentComment) and bare approvals (isApproval) before applying the comment window, and omits the ### Discussion header when nothing survives the filter.
• implementTaskPrompt/planTaskPrompt cap the plan with truncate; prCommentTaskPrompt caps comment count, diff-hunk size, and review count.
• Added tests: filtering, empty-discussion edge case, PR-comment cap, and a bounded-size regression test.

Part C — stop the CLI loading ambient operator context (internal/claude/runner.go)

• Added --strict-mcp-config, --disable-slash-commands, --exclude-dynamic-system-prompt-sections, --autocompact 200000 as hardcoded args, before ExtraArgs. Added TestRunPassesTokenReductionFlags.

Also updated README.md's Prompts section and the verbatim claude … invocation block to document all of this.

Verification: go build ./... succeeds; go test ./... passes for every package except internal/git, which fails on main too (pre-existing environmental identity-config mismatch, unrelated to any file this change touches — verified no diff exists in internal/git).

Note for reviewer: early in the session I accidentally ran a bare git stash push that briefly stashed this in-progress work; I recovered it immediately via git stash apply <sha> (not pop) and dropped the entry — final git status confirms only the intended 11 files are modified, matching the plan's file list exactly.

Verification

Tests failed (make test). This PR is a draft — the failure is reported rather than hidden, so you can judge whether the change is salvageable.

Test output (tail)

go test -race ./...
go: golang.org/toolchain@v0.0.1-go1.26.5.linux-arm64: verifying module: golang.org/toolchain@v0.0.1-go1.26.5.linux-arm64: open /home/node1/go/pkg/sumdb/sum.golang.org/latest: read-only file system
make: *** [Makefile:122: test] Error 1

Changes

README.md                            | 22 +++++++++-
 internal/claude/runner.go            | 34 +++++++++++++++
 internal/claude/runner_test.go       | 36 ++++++++++++++++
 internal/discord/notifier.go         |  3 +-
 internal/orchestrator/loop.go        | 17 ++++++--
 internal/orchestrator/prcomments.go  | 17 ++++++--
 internal/orchestrator/prompt.go      | 51 ++++++++++++++++++----
 internal/orchestrator/prompt_test.go | 82 ++++++++++++++++++++++++++++++++++++
 internal/store/store.go              | 38 ++++++++++++++---
 internal/store/store_test.go         | 43 +++++++++++++++++--
 internal/web/assets/app.js           | 15 ++++++-
 11 files changed, 328 insertions(+), 30 deletions(-)

---

Opened automatically by coding-agent-loop (run b328b7bb-cc7a-451c-b698-00fbcad54205, attempt 3, model claude-sonnet-5, cost $2.2602, session d328a89e-cafd-4133-a8b1-2435aa7f4c48). Nothing here has been reviewed by a human yet.