chore(deps): take the cohort to agent-eval 0.140.1 so the runtime's own journal parses

[drewstone]

Why

src/durable/spawn-journal.ts writes every supervision event wrapped in an envelope:

// src/durable/spawn-journal.ts:273, type at :423-424
await this.appendRecord({ kind: 'event', root, event: ev })

Through agent-eval 0.140.0, parseSupervisorTree read only the flat dialect — it took row.kind at face value, so it saw the literal string event, matched no tree event kind, and folded nothing into the tree. Every read of the runtime's own journal reported zero spawns.

0.140.1 adds readJournalRow, which unwraps {kind:'event', root, event}, ignores the {kind:'begin', root, at} header, and tags the source journalDialect: 'runtime-envelope'.

Runtime's peer range was >=0.139.2 <0.140.0, which excluded the fix for its own output format.

Measured, not assumed

Both versions installed side by side, fed the exact envelope spawn-journal.ts emits (2 spawned + 1 settled under one begin header):

agent-eval	spawns	closes	journalDialect
0.139.3	0	0	(field absent)
0.140.1	2	1	runtime-envelope

dist/integrity-*.js is byte-identical between 0.139.3 and 0.140.0 (md5 7b10cb0fee10265d4245db65ee5c364f), and readJournalRow appears in neither — 0.140.1 is the first version that can read this repo's journal.

Places updated

The version is load-bearing in five source-of-truth files that must agree, plus two generated ones:

File	Line	Before	After
pnpm-workspace.yaml	21	'@tangle-network/agent-eval': 0.139.2	0.140.1
pnpm-workspace.yaml	23	'@tangle-network/agent-knowledge': 7.0.3	7.0.4
package.json	171	">=0.139.2 <0.140.0"	">=0.140.1 <0.141.0"
.github/workflows/ci.yml	80	36de070…513f # v0.139.2	55d2faacc6ba2c92ad21b42b9b607ab16ae0fc7c # v0.140.1
.github/workflows/ci.yml	87	d932983…3635 # v7.0.3	b67218f308ff0c8df6fe0cc8d8e47fa6351d0ecb # v7.0.4
.github/workflows/publish.yml	146	36de070…513f # v0.139.2	55d2faacc6ba2c92ad21b42b9b607ab16ae0fc7c # v0.140.1
.github/workflows/publish.yml	154	d932983…3635 # v7.0.3	b67218f308ff0c8df6fe0cc8d8e47fa6351d0ecb # v7.0.4
docs/canonical-api.md	9	>=0.139.2 <0.140.0	>=0.140.1 <0.141.0
pnpm-lock.yaml	—	regenerated (pnpm install)	0 remaining 0.139.2 / 7.0.3 rows
docs/api/primitive-catalog.md	10	generated header	regenerated (pnpm run docs:api)

Both cohort SHAs were resolved from the tags, not typed by hand:
gh api repos/tangle-network/agent-eval/commits/v0.140.1 → 55d2faa…, …/agent-knowledge/commits/v7.0.4 → b67218f….
The agent-interface pin is untouched: eval 0.140.1 and knowledge 7.0.4 both depend on 0.40.0 exactly.

Why agent-knowledge moves too

This is not scope creep — the cohort is a four-package lockstep. scripts/verify-packed-cohort.mjs:530 runs assertExactDependency(agentKnowledge, agentEval), requiring knowledge's dependencies['@tangle-network/agent-eval'] to exactly equal the packed agent-eval version:

• agent-knowledge@7.0.3 → "@tangle-network/agent-eval": "0.139.2"
• agent-knowledge@7.0.4 → "@tangle-network/agent-eval": "0.140.1"

Bumping eval alone would have failed cohort verification with agent-knowledge requires @tangle-network/agent-eval@0.139.2, packed 0.140.1 — the same class of multi-place drift that sank the v0.110.0 tag.

controlAdapter is not a breaking change here

Checked before assuming. controlAdapter is on DspyRlmTraceEngineOptions and is optional (controlAdapter?: 'chat' | 'two-step' | 'tolerant', default tolerant) in both 0.140.0 and 0.140.1 dist/skill-usage-*.d.ts — not required. Independently, agent-runtime has zero call sites: grep -rni 'dspy|controlAdapter' over the repo returns 3 prose mentions in docs/ and 3 false positives on a local endSpy variable in tests/runtime/worktree-cli-executor.test.ts. Blast radius is zero, and typecheck agrees.

Evidence

All run locally on this branch; the baseline was measured on origin/main @ 56dd6eb before the change.

Gate	origin/main @ 56dd6eb	This branch	Verdict
pnpm run typecheck	0 errors	0 errors	clean
pnpm run test — files	1 failed | 196 passed | 2 skipped (199)	1 failed | 196 passed | 2 skipped (199)	identical
pnpm run test — tests	1 failed | 2260 passed | 6 skipped (2267)	1 failed | 2260 passed | 6 skipped (2267)	identical
pnpm run check:publish-workflow	—	pass	Publish workflow keeps package creation separate from npm authority.
pnpm run lint	—	pass	542 files, no fixes applied
pnpm run docs:freshness	—	pass	version pin + 3 substrate peers accurate

Zero new failures. The single failure is identical on both sides — same test, same 15000ms timeout, same tests/knowledge-improvement-job.test.ts:296 — a known flake under parallel load. Re-run isolated on this branch (i.e. against agent-knowledge 7.0.4): 3/3 green, 3 passed (3) in 13.1s / 19.3s / 10.8s.

docs/api/primitive-catalog.md regenerated to a one-line diff (the version in the generated header), confirming the API surface agent-runtime consumes is unchanged across the bump: 2351 catalogued symbols, 0/0 undocumented own callables.

Version number and tagging are deliberately untouched — this branch only moves the dependency cohort.