docs(agent-eval-adoption): dynamic driver + optimizePrompt; fix stale topology claims

[tangletools]

The agent-eval-adoption skill predated agent-runtime 0.33.0 and was actively
steering adopters wrong:

• Topologies section was stale/incorrect. It said "Refine + FanoutVote
  shipped; others deferred" and explicitly told people the Decompose
  topology was NOT shipped and to fork a custom Driver. But createDynamicDriver
  • createSandboxPlanner (agent-authored topology) ship in 0.33.0 and subsume it.
• createFanoutVoteDriver signature was wrong ({variants, scoreFn} → actual
  {n, selector}).
• No mention of optimizePrompt (identity-gated prompt-surface optimization).

Changes:

• Refresh Topologies: add createDynamicDriver / createSandboxPlanner,
  move Decompose out of "deferred", correct the fanout-vote signature. Council +
  Pipeline remain the only deferred topologies.
• New subsection "Prompt-surface optimization — optimizePrompt (identity-gated)":
  the reusable [SURFACE]-prompt optimization recipe (extract prompt → domain
  scenarios → judge dims → runWithPrompt → gepaDriver + held-out gate) and the
  footguns learned the hard way — gepaDriver is text/H2-only, scenarios must be
  domain-real, extend-don't-fork an existing runImprovementLoop, cost via ctx.cost.
• Frontmatter + plugin.json descriptions updated for discoverability (0.33.x).

Pairs with the self-contained agent-runtime-adoption skill now shipped inside
the agent-runtime repo (tangle-network/agent-runtime#77), so external consumers
of that package need nothing from this marketplace.

[drewstone]

✅ Auto-approved tangletools PR — 93f4cd24

This PR was opened by the trusted tangletools automation account.
The full PR reviewer audit still runs separately and will publish findings if it detects issues.

_{tangletools · auto-approval · reason: tangletools_author · 2026-05-31T01:22:32Z}

[tangletools]

✅ No Blockers — 93f4cd24

Readiness 82/100 · Confidence 65/100 · 4 findings (2 medium, 2 low)

	deepseek	kimi-code	aggregate
Readiness	82	82	82
Confidence	65	65	65
Correctness	82	82	82
Security	82	82	82
Testing	82	82	82
Architecture	82	82	82

Full multi-shot audit completed 1/1 planned shots over 2 changed files. Global verifier still owns final merge decision. | Full multi-shot audit completed 1/1 planned shots over 2 changed files. Global verifier still owns final merge decision.

🟠 MEDIUM Key docs section retains stale package version references inconsistent with updated frontmatter — plugins/agent-eval-adoption/skills/agent-eval-adoption/SKILL.md

The frontmatter description was updated to reference @tangle-network/agent-eval (0.50.x+) and @tangle-network/agent-runtime (0.33.x+), and the body documents createDynamicDriver / optimizePrompt as agent-runtime 0.33.0+ APIs. However, the Key docs section at lines 1118–1126 was left unchanged and still lists @tangle-network/agent-eval@0.36.x README and @tangle-network/agent-runtime@0.28.x/loops. An adopter following the Key docs index may look at the wrong package version documentation and fail to find the newly documented APIs. Fix: update Key docs to @tangle-network/agent-eval@0.50.x+ README and `@tangle-netw

🟠 MEDIUM Stale version references in Key docs section after description bump — plugins/agent-eval-adoption/skills/agent-eval-adoption/SKILL.md

L1118 references @tangle-network/agent-eval@0.36.x but the description frontmatter (L3) was bumped to 0.50.x+. L1122 references @tangle-network/agent-runtime@0.28.x/loops but the description now says 0.33.x+ and new features (createDynamicDriver L145-146, optimizePrompt L180) require 0.33.0+. The agent-runtime mismatch is clearly wrong and will misdirect readers to stale API docs. Fix: update L1118 to 0.50.x and L1122 to 0.33.x to match the frontmatter version target.

🟡 LOW Comment references wrong variable name in optimizePrompt example — plugins/agent-eval-adoption/skills/agent-eval-adoption/SKILL.md

L194 comment says assign result.prompt unconditionally but the destructuring at L187 binds prompt directly (not result.prompt). The comment should say prompt to match the destructured variable. Minor reader confusion risk.

🟡 LOW optimizePrompt example destructures prompt but prose refers to result.prompt — plugins/agent-eval-adoption/skills/agent-eval-adoption/SKILL.md

Lines 187–194 destructure the return value as const { prompt, improved, decision, delta } = await optimizePrompt(...), yet the comment on line 194 says // assign result.prompt unconditionally, and the prose on line 199 says `result.prompt` is the baseline UNLESS `decision === 'ship'`. A reader copying the d

---

_{tangletools · 2026-05-31T01:29:30Z · trace}