agenthropic

Phase-0 Corpus Probe — the empirical CD-1 verdict

The evidence behind CD-1. The best-path memo named one do-this-now action: run the full-corpus read-only probe and write the CD-1 verdict with real numbers. This is that probe, executed against the real ~/.claude/projects/ corpus on this Mac Mini on 2026-07-04. It was run as an 8-agent read-only workflow (census → reconstruct → reconcile → independent re-verification → judge; Opus / high effort; 0 errors; ~326k agent tokens), and its two load-bearing claims were then re-verified by hand.

CD-1 = CONDITIONAL-GO → BUILD. Confidence 85/100. JSONL is a trustworthy, outage-surviving single source of truth for the persisted subagent DAG and the per-agent token ledger — but only for a parser that keys on the Agent/Workflow spawn tools (not Task), walks both on-disk layouts, indexes subagents as parents, and sums tokens from child transcripts (parent rollups are effectively absent). Those four are the CD-1 acceptance gate, not nice-to-haves.

This probe corrects three things the earlier docs asserted — see §4. Every integer below was computed from disk, not inferred.

The census below is a 2026-07-04 snapshot and is no longer the census of record (note added 2026-08-15). The corpus has grown by roughly six weeks of daily use since, and the numbers in §1 have moved with it — parser-spec.md §4.2 holds the current figures and is what to quote. Nothing here is retracted: the probe’s verdict rests on structural facts (which spawn tools carry the join key, which layouts exist, where the tokens live), and those have held. It is the magnitudes that expired, so treat §1 as dated evidence rather than as a description of the corpus today. One counting subtlety survives the move and is worth carrying across: a subagent-transcript count is not a session count, and the two have been confused before — §4.2 states both explicitly.

---

1. The corpus (census)

Metric	Value
Project slugs	17
Top-level session transcripts	117
Sessions with a subagents/ dir	33 (23 flat-only · 7 nested-only · 3 both)
Sessions with zero subagents	84 / 117 (72%) — the DAG is sparse; optimize for the empty case
Subagent transcripts (agent-*.jsonl)	~976–997 (live-growing) — 148 flat + ~849 nested
Nested share	85.2% (verified by hand)
*.meta.json sidecars	1:1 with agent files (0 missing)
Spawn-depth histogram	depth-1 = 968 · depth-2 = 6 · depth-3+ = 0
agentType split	workflow-subagent 828 · general-purpose 121 · Explore 27
CC versions (agent lines / all transcripts)	7 / 13 (2.1.70 … 2.1.200; 2.1.197 dominant = 593)

The corpus grew while being probed (this very session spawns workflow-subagents), so meta-denominator percentages are moving snapshots, not fixed constants — noted wherever it matters.

2. The four CD-1 questions, answered with numbers

Q1 — Is the depth-1 parent→child edge a HARD key? → YES, 0% orphan.

The probe first measured 0.893% orphans (10/1120); the independent verifier could not reproduce it and got 0.000% (0 / 146). Trusting the verifier (safe direction, and stronger): all 146 flat toolUseIds resolve to a real spawn tool_use block — 140 in the owning session, 6 inside a parent-agent transcript. Two exact structural keys: meta.toolUseId == parent Agent tool_use.id and filename agent-<hex> == toolUseResult.agentId. JSONL is the authoritative substrate; the depth-1 edge is a data fact.

Q2 — Are depth-2 / nested edges recoverable? → YES, 100% for the proven populations.

Two distinct populations were conflated in earlier notes:

• Genuine depth-2 = only 6 flat Explore grandchildren in a single session (b24be30c), recovered 6/6 via the same toolUseId key — but the key lives inside a depth-1 agent transcript, so the parent index must be self-referential (a session→child-only index silently drops them).
• The ~828 “nested” workflow-subagents are all spawnDepth-1, not depth-2. They carry no toolUseId (0/828) and parentUuid=null; they attach to their owning session 844/844 (100%) via sessionId == session-uuid + wf_<id>/ directory containment + the parent’s Workflow tool_use.
• Residual (open gap): intra-workflow “who spawned whom” inside a wf_ dir is not toolUseId-keyed and was not proven to 100% — it needs a journal.jsonl (started/result rows) + promptId reconstruction. This is the one edge family CD-1 leaves unproven.

Q3 — Version / layout drift? → SIGNIFICANT — one parser, two join code paths.

7 CC versions across agent lines, 13 across session transcripts; both layouts present. The key correction: layout is not a version boundary — flat and nested coexist within the same versions. It is driven by the spawn mechanism: the Workflow tool writes a nested workflows/wf_*/ subtree; the Agent tool writes flat agent-*.jsonl. A parser must branch on directory shape, not on version. The message.usage token schema is identical across both, but the edge join needs two paths (toolUseId for flat; workflow-journal + directory for nested).

Q4 — Token attribution & reconciliation? → 100% attributable; sum from children; parent rollup ≈ 0%.

Every one of the ~18,213 assistant message.usage lines carries an agentId → 100% attributable (filename hex == internal agentId; 0 unattributable — the 251 zero-token files are <synthetic> stubs, nothing to attribute). The parent side records nothing to reconcile against: async Agent spawn tool_results are async_launched with no token rollup (0/142); only the 24 synchronously-completed results carry one (14.5% overall, 0% for the async majority). Σ over child transcripts is the sole exact method — verdict NEEDS_SUMMATION. Summed totals (internally exact — the model histogram sums to every usage line):

Tier	Tokens
input (non-cache)	≈ 7.75 M
output	≈ 5.53 M
cache_creation	≈ 83.3 M
cache_read	≈ 584 M
grand total incl. cache	≈ 678 M

Models seen: claude-opus-4-8, claude-fable-5, claude-sonnet-5, claude-haiku-4-5, <synthetic>.

3. Ingest invariants surfaced (not CD-1 blockers, but correctness-critical)

Invariant	Evidence	Consequence for the ingester
Compaction is a context reset	63 / 117 sessions carry a compaction/PreCompact/summary boundary	Treat isCompactSummary/compactMetadata lines as resets, not normal turns.
Session wall-clock is NOT an isolation boundary	92 / 117 sessions overlap a same-slug sibling; some span up to ~190 h resumed	Key strictly on session-uuid + timestamps; never assume one live session per project-slug.
The DAG is sparse	84 / 117 sessions never spawn a subagent	Optimize the reader for the empty case; the graph is the exception, not the rule.
Crashes are ~nonexistent historically	1 crash-proxy hit — and it is this live probe session itself	The durable-outbox correctness argument has ~no empirical support (see §4c).

4. Three corrections to the earlier docs

a) The spawn tool is Agent / Workflow, not Task (highest-impact — verified by hand: Task blocks = 0, Agent = 142, Workflow = 29). The design digest (DESIGN.md) and prior analysis say “Task tool”. A parser hard-keyed to name=='Task' reconstructs an empty DAG. This must be encoded in DESIGN.md, the reconstructor spec, and a regression test.

b) Layout drift is spawn-mechanism-driven, not version-driven. The best-path memo said the layout “churns on patch bumps (2.1.198 → 2.1.199).” Wrong — both layouts coexist within the same versions; the driver is Agent-tool (flat) vs Workflow-tool (nested). The parser branches on directory shape. (Version detection is still worth keeping for provenance/telemetry.)

c) The durable outbox is YAGNI-leaning, not load-bearing. The best-path memo insisted “do not demote WP-IN11 (durable outbox).” The evidence downgrades it. JSONL is an already-durable, complete, self-reconciling source that Claude writes independently, so state can always be re-derived by re-reading the tree (backfill). Historical crash rate ≈ 0. The outbox buys latency, not correctness. Trigger to add it: a sub-second live-freshness requirement, or hooks becoming a data source not also present in JSONL. Until then, a periodic / notify-driven JSONL re-scan suffices and is simpler. The memo’s other hedges survive intact — see §5.

5. The load-bearing split (what the evidence proves you must build vs may defer)

Hedge	Verdict	Why
Dual-layout parsing	🔒 LOAD-BEARING (hard)	85% of agents are nested workflow-subagents with no toolUseId; a single toolUseId-keyed reader recovers only ~15% and drops the entire Workflow family.
Child-transcript token summation (CD-3 backfill)	🔒 LOAD-BEARING (hard)	Parent rollup is 0% for the async majority; tokens live in exactly one place — message.usage inside each child transcript. The summation is the ledger.
Self-referential parent index	🔒 LOAD-BEARING	The 6 depth-2 edges resolve only against a depth-1 agent transcript; a session→child-only index drops them.
Durable outbox (WP-IN11)	🟡 DEFERRABLE (was “keep”)	JSONL self-reconciles by backfill; ~0 historical crashes; buys latency not correctness. Add on the trigger in §4c.

6. Parser requirements — the CD-1 acceptance gate

The reconstructor is GREEN only when it satisfies all of these (each is forced by a measured number above):

1. Match spawn tools on {Agent, Workflow}, never Task (input keys description / prompt / subagent_type). Zero Task blocks exist.
2. Dual-layout directory walker — flat subagents/agent-<hex>.jsonl and nested subagents/workflows/wf_<id>/agent-<hex>.jsonl, each with a .meta.json. Nested = 85%.
3. Two join schemas — (a) flat: meta.toolUseId == parent Agent tool_use.id (and toolUseResult.agentId == filename hex); (b) nested: parent Workflow result’s transcript-dir path → wf_ members, anchored by agent.sessionId == enclosing session uuid.
4. Self-referential parent index — index subagent transcripts as candidate parents (depth-2).
5. Structural block-id equality only — join on tool_use.id equality, never substring / text grep of agent-<hex> / toolu_ ids (transcripts quote each other’s ids → false links).
6. Token summation from child transcripts — group message.usage by agentId; sum all four tiers. Do not rely on parent rollups.
7. Legacy fallback for CC 2.1.70 bare {agentType:'Explore'} metas (no toolUseId/ spawnDepth) — join via raw agentId in parent progress lines. Tiny (2 files) but present.
8. Compaction handling — treat compaction lines as context resets (63/117 sessions).
9. Concurrency-safe ingest — key on session-uuid + timestamps; never one-session-per-slug (92/117 overlap).
10. CC-version detection from the per-line version field for provenance — but branch parsing on directory shape, not version.
11. Intra-workflow edge reconstruction via journal.jsonl (started/result by agentId) + promptId — the one edge family not covered by toolUseId; build and validate it, or ship intra-workflow ordering as best-effort (see §7).

7. Open gaps (what this probe did NOT prove)

• Intra-workflow parent→child ordering inside a wf_ dir is unproven to 100% — only agent→session (844/844) and flat depth-2 (6/6) are proven. Trigger to downgrade the build: if this can’t be reconstructed at a high rate on a real workflow corpus, ship the DAG at agent → session + Agent-tool granularity and mark intra-workflow ordering best-effort, not a data fact.
• Depth-2 sample is tiny and non-diverse — 6 edges in one session; no depth-3 anywhere. Generalization to deeper/wider trees is unmeasured.
• Outage survival is unstressed — ~0 historical crashes means the outbox-vs-backfill call is a design inference, not a measured recovery test. A deliberate kill-mid-run test would harden the YAGNI claim.
• The 0.893% → 0.000% orphan discrepancy was resolved in the safe direction but suggests the first probe used a restricted denominator — worth confirming the join harness on a frozen corpus snapshot (the live corpus polluted the run).

8. What this means for the plan

• CD-1 is answered: CONDITIONAL-GO → build, with the §6 list as the literal acceptance gate for Track S / the reconstructor. No production code before the gate is codified as tests.
• Amend DESIGN.md + the reconstructor spec for the Agent/Workflow-not-Task correction and the directory-shape branch (§4a/§4b).
• Keep CD-2/CD-3 (immutable substrate + token backfill) and multi-layout parsing — proven load-bearing. Defer the durable outbox to a triggered add (§4c) — this frees WP-IN11 from the v1 critical path, reinforcing the memo’s “alerts + non-moat off the critical path” move.

---

Read-only probe; no files under ~/.claude modified. 8 agents, 0 errors, ~326k tokens, ~24 min. Full agent returns: the run’s journal.jsonl. The two load-bearing facts (Agent≠Task; 85% nested) independently re-verified by hand. Companion strategic memo: best-path-decision.md. Decisions refined: concept-analysis-v2.md CD-1…CD-3. Build items: development-plan.md Track S · Open work: ../../TODO.md.

This site is open source. Improve this page.