feat: measured output format changes and harness event classification (ADR-007, ADR-008) - #9
Merged
Merged
Conversation
The success path took the first non-empty line with no filtering, so a script's own "--- HEAD ---" was reported as the state of a working tree and a test runner's "RUN v4.0.18" as the result of a test run. Failures already skipped known-noise shapes; this gives success its own set (echoed section headings, progress lines, version banners) and strips ANSI escapes on both paths. Section headings are recognized only when the same rule closes the line, so a diff's "--- a/file.go" survives as an excerpt. The success-only shapes are deliberately not applied to failures: a banner-shaped line can be the error itself when the call failed. Measured on a 1,258-Bash-call session: 292 lines (23%) now carry a better excerpt, none lost one, and ANSI escapes drop from 95 lines to zero. A prose usage banner (gh's "Work seamlessly with GitHub from the command line.") has no shape to key on and is not covered; a test pins that limit. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8xGh8fwz3YjLhFTMYqvXB
#3) Two changes that pay for each other. The Bash summary printed only the description Claude wrote for itself, so 78% of a session's tool calls could not be checked against what actually ran. And "-> ok" annotated 1,550 of 1,604 calls with the state a reader already assumes. The verb is not a prefix of the command. Real commands open with `cd <absolute path> &&`, an env assignment, a shell keyword, or an echoed heading, so a 20-character prefix reads "cd /Users/maple/Desk" and names nothing; a 60-character one costs +9.8% tokens for the same reason. Stepping over those to the first real program name reaches one on 1,258 of 1,258 Bash calls in the measured session at +2.0%, which the -2.0% from the dropped marker cancels. Measured together over 5 sessions: -0.0%. Success now renders its excerpt or diff stat alone, still introduced by "->" so the call stays separable from its result. FAILED keeps its marker, and with it formatRetryCollapse's insertion point for "×N". Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8xGh8fwz3YjLhFTMYqvXB
"[08-06 03:06]" repeated a date identical on every message of a day and had no year at all, so a read output gave no way to place a session in time (context could fall back to its session header; read has none). The date moves to a "--- 2026-08-06 ---" marker emitted when it rolls over, and the space it freed buys second precision. Measured over 5 sessions this is -0.4% tokens while carrying strictly more than the old format: seconds and the year. Sessions really do span days in the sample (2 and 3), but sparsely: the marker fired twice across a 1,808-event session. An unparseable timestamp renders a placeholder and leaves the day state alone, so a malformed event cannot invent a date boundary. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8xGh8fwz3YjLhFTMYqvXB
The 20,000 limit existed to keep a page inside the harness's Bash stdout cap, but that cap was never measured. Binary search on 2026-08-29 put it at 30,000 characters: 30,000 came back inline, 31,000 was persisted to a file with a 2KB preview. 28,000 leaves headroom for the page marker and footer; the largest real page now renders at 28,073 bytes. A 871KB session drops from 44 pages to 32, so inheriting it costs 12 fewer Bash round-trips, each of which was a separate API request paying cache read on the whole prefix. The constant is exported so the boundary tests derive their cases from it instead of restating 20,000, which had already stopped describing the real boundary in three of them. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8xGh8fwz3YjLhFTMYqvXB
The set lands at +0.3% tokens rather than the predicted -0.4%. The gap is decision 1 doing its job: skipping a noise line reaches a longer, more informative one. Also records the two rules the implementation tightened (symmetric section headings only, indented code not skipped) and the shell control keywords the verb extraction had to step over. Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01M8xGh8fwz3YjLhFTMYqvXB
…-008) Reader now recognizes eight entry types the CLI added after the noise whitelist was written, classifies six harness-injected user messages in the parser layer and renders them under a `harness:` role label, detects teammate messages by a set of markers instead of a disclaimer string that no longer matches, and resolves skill-body injections through the sourceToolUseID link to the Skill tool_use so bundled skills without a base-directory line compact to `[skill: name]`. Turn counting moves out of the character-accounting loop into UserMessage.CountsAsTurn(), calibrated against transcripts: teammate messages, task notifications, and compaction summaries start a round and count; interrupts and agents-stopped notices do not. Docs: SKILL.md and README record the 28K page size and the harness role. Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TcvWErnKnuADJsomnZfwFr
Mapleeeeeeeeeee
force-pushed
the
feat/adr-007-format-changes
branch
from
September 1, 2026 20:05
fe92038 to
3e80212
Compare
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
為什麼
用過 cc-session 的 Claude 會把工具的成功摘錄當成執行結果讀、無法驗證 Bash 實際跑了什麼、分不出哪幾行是人打的。這兩份 ADR 記錄的改動都先量過再做。
ADR-007：五項格式改動（
docs/adr-007-format-changes-measured.md）| <程式名 子命令>，先剝掉cd、環境變數、shell 關鍵字再取動詞，1,258 個 Bash 呼叫全部抽到-> ok，只標FAILED[HH:MM:SS]，日期在換日時獨立一行--- YYYY-MM-DD ---，比原本多秒與年份ADR-008：harness 事件漂移（
docs/adr-008-harness-event-drift.md）noiseTypes補 8 個 CLI 後來加的 entry type（60 天內 597 KB 原本在 stats 裡隱形）harness:角色標籤（context格式為H:），讀的人分得出哪幾行不是人打的UserMessage.CountsAsTurn()把 turn 判定從字元統計的continue拆出來，政策以 transcript 校準（8 個 session 誤差 70 → 29），修正 teammate 密集 session 的 K 高估sourceToolUseID結構連結，沒有 base-directory 行的 bundled skill 不再全文渲染驗證
gofmt/go vet/go test ./...全綠，新增success_excerpt_test、command_verb_test、timestamp_test、turn_test、harness_classify_test、harness_role_test、skill_injection_link_testedf313157.6 → 4.3、544c7a3d8.6 → 5.5、其餘 ±1 以內b11858cf：46 則user:→ 36 則user:+ 10 則harness:；artifact-designskill 內文從全文渲染變成[skill: artifact-design]沒有包含
<agent-message>teammate 變體（80/1,112）與 teammate 偵測是否改成只認標籤，留待下次決定search子命令，等自己的 ADR🤖 Generated with Claude Code
https://claude.ai/code/session_01TcvWErnKnuADJsomnZfwFr