瀏覽代碼

docs: explore sufficiency -- the metric, its rules, what it caught (CG-8)

Records what each bucket means and which fix it points at, the four rules that
keep the classification honest (same-message calls, bookkeeping tools, subagent
threads, earlier-explore files), and the three real transcripts it was
hand-checked against -- including the excalidraw canvasNonce run, where it
independently found the data-flow frontier CLAUDE.md already documents: 0%
sufficient, without being told what to look for.

Also states what it does NOT say: sufficient is not correct, one Read is a vote
rather than a proof, and bucket 1 is ambiguous by construction.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Colby McHenry 1 月之前
父節點
當前提交
945e52f4ee
共有 2 個文件被更改,包括 148 次插入3 次删除
  1. 144 0
      docs/benchmarks/explore-sufficiency.md
  2. 4 3
      docs/benchmarks/residual-context-occupancy.md

+ 144 - 0
docs/benchmarks/explore-sufficiency.md

@@ -0,0 +1,144 @@
+# Explore sufficiency
+
+**What it measures:** whether a `codegraph_explore` response was *enough* — read
+off what the agent did next, which the harness was throwing away.
+
+The agent tells us on every call. It reads a file, or it explores again, or it
+answers. That next action is free ground truth, and it splits into buckets that
+each point at a different fix:
+
+| Next action | Bucket | What it means |
+|---|---|---|
+| another codegraph call | `explore again` | insufficient — the response did not answer |
+| `Read` of a file we **returned** | `Read a file we returned` | **allocation**: right file, wrong bytes |
+| `Read` of a file we did **not** return | `Read a file we did not return` | **recall**: the file never surfaced |
+| `Grep` / `Glob` | `Grep/Glob` | recall, weaker signal — still hunting |
+| `Edit`, a build, the final answer | `moved on / answered` | sufficient |
+
+Harness-only. Nothing is emitted from the product and nothing leaves the machine
+(decided 2026-08-03); it is parsed out of transcripts we already write.
+
+---
+
+## Running it
+
+Every run prints it — `run-all.sh` and anything else that calls `parse-run.mjs`:
+
+```bash
+scripts/agent-eval/run-all.sh /tmp/codegraph-corpus/express \
+  "How does res.send decide the Content-Type and ETag?"
+
+# Or over a log you already have:
+node scripts/agent-eval/parse-run.mjs /tmp/agent-eval/run-headless-with.jsonl
+```
+
+```
+  Explore sufficiency — what the agent did NEXT (2 answered calls):
+      1  50%  explore again                  insufficient: did not answer
+      1  50%  Read a file we returned        allocation: right file, wrong bytes
+      0   0%  Read a file we did not return  recall: file never surfaced
+      0   0%  Grep/Glob                      recall (weak): still hunting for the file
+      0   0%  moved on / answered            sufficient
+    1. "res.send Content-Type ETag generation" [3 files] → codegraph_explore
+    2. "response.js res.send function body" [3 files] → Read response.js
+```
+
+Interactive runs get the same block from `parse-session.mjs <project-dir>`.
+
+The per-call lines matter as much as the counts: they name the query that fell
+short and the file the agent went and read instead, which is usually enough to
+reproduce the miss with `probe-explore.mjs`.
+
+---
+
+## The rules that keep it honest
+
+**Only a later message counts as a reaction.** A `Read` fired in the *same*
+assistant message as the explore was issued before the response existed, so it
+is not a verdict on it. Those are stepped over and counted separately as
+`concurrent`.
+
+**Bookkeeping is stepped over.** `ToolSearch` (pulling a deferred tool schema)
+and `TodoWrite` say nothing about the response; the call behind them is the
+verdict.
+
+**Subagents are a separate thread.** Claude Code interleaves a subagent's tool
+calls into the same stream-json output, tagged `parent_tool_use_id` — verified
+on a real excalidraw run where a delegated search's greps landed between the
+parent's own calls. Reactions are matched within one thread, or a subagent's
+first grep gets scored as the parent's verdict on an explore it never saw. In
+interactive sessions the subagent lives in its own file instead; each
+`agent-*.meta.json` carries the `toolUseId` of the Task that spawned it, and
+`parse-session.mjs` stitches the threads back together with it.
+
+**A delegation is judged by what the subagent did first.** `Agent`/`Task` →
+first substantive call, shown as `Agent → Bash search`. Scoring the delegation
+itself as "moved on" is the one error a tuning metric must not make: on the
+excalidraw run below it reported 33% sufficient while the subagent was off
+grepping for the file. A delegation that never runs a tool stays "moved on".
+
+**Shell file access counts.** Both arms have Bash, and agents reach for
+`sed -n '100,200p' lib/x.js` as readily as for `Read`; `grep`/`rg`/`find`/`ls`
+count as search. Counting only the `Read` and `Grep` tools would score those
+explores as sufficient. A heredoc or a redirect is writing, not reading.
+
+**Returned means we shipped the source.** A file is "returned" if the response
+carried its per-file section (the ``**` `` marker). A file the response only
+*named* — a flow step, a blast-radius entry — is still a recall miss, flagged
+`(named, not returned)`, because pointing without delivering is its own failure.
+
+**A file an earlier explore shipped still counts as returned.** Re-reading it is
+an allocation miss wherever we shipped it; filing that as recall would aim the
+fix at the wrong end of the pipeline. The line says which
+(`Read utils.js (returned by an earlier explore)`).
+
+**Errored calls are counted, not bucketed.** An explore that came back
+`isError` or never returned has no response to judge; it is reported separately.
+
+---
+
+## Validation
+
+Hand-checked against real transcripts, then swept over all 76 A/B logs on this
+machine (0 crashes, 176 calls bucketed).
+
+**`cg22/ab-express/run-baseline-1` — the allocation bucket, by hand.** Sequence:
+explore *"res.send Content-Type ETag generation"* → explore *"response.js
+res.send function body"* → `Read /…/t-base/lib/response.js`. The second explore
+returned `lib/response.js`, and the agent read it anyway → `explore again`, then
+`Read a file we returned`. That is the #1500 allocation bug (the 583-byte stub)
+showing up as a bucket instead of as a hunch. The new-build arm of the same
+question: one explore, `moved on / answered`, 100% sufficient.
+
+**`cg15/ab-express/run-new-2` — the same verdict on a longer run.** Four
+explores; the fourth returned `lib/utils.js` and the agent then read
+`/…/t-new/lib/utils.js` at `offset: 195`. Right file, wrong window.
+
+**excalidraw `canvasNonce` — the recall bucket, end to end.** A fresh
+`run-all.sh` arm on the documented data-flow frontier: three explores, the last
+one delegating a subagent that immediately grepped for `sceneNonce` → 67%
+`explore again`, 33% `Grep/Glob`, **0% sufficient**. That matches what
+`CLAUDE.md` already records about this question (the residual reads and greps
+are all nonce data-flow, deliberately uncovered) — the metric found it without
+being told.
+
+`node scripts/agent-eval/parse-run.mjs --selftest` covers the classifier over
+synthetic transcripts with known answers: every bucket, the same-message rule,
+the thread rule, delegation, shell reads and searches, and errored calls.
+
+---
+
+## What it does not say
+
+- **Sufficient is not correct.** The agent moving on means the response was
+  enough to stop it, not that the answer was right. Answer quality is not
+  measured here.
+- **A Read is a vote, not a proof.** An agent sometimes re-reads a file it
+  already has. The bucket is still the right signal — it read *because
+  something was missing* — but a single call is noisy; read the counts over a
+  pass, not one run.
+- **Bucket 1 is ambiguous by construction.** "Explored again" means the response
+  did not answer; it does not say whether that was allocation or recall. The
+  follow-up explore's query usually does.
+- **Small-n.** Runs make 1–5 explore calls, so one run's percentages are coarse.
+  Compare arms over a pass (`RUNS>=2`, and the 7-repo campaign), never n=1.

+ 4 - 3
docs/benchmarks/residual-context-occupancy.md

@@ -194,6 +194,7 @@ is a baseline across the 7 README repos to compare future changes against.
   only its summary returns to the parent. Runs that delegate are measured on the
   parent's window alone.
 - **Occupancy is not sufficiency.** A small residual is only good if the answer
-  was still right. This metric says nothing about answer quality — that is
-  CG-8's job (sufficiency) and CG-9's (how much of the returned bytes the answer
-  actually used).
+  was still right. This metric says nothing about whether the response was
+  *enough* — that is [explore sufficiency](explore-sufficiency.md), which every
+  run now prints alongside this block — nor about how much of the returned bytes
+  the answer actually used (CG-9).