feat(subagent-dev): accumulate discoveries across tasks

[mvanhorn]

What problem are you trying to solve?

When using subagent-driven-development, each subagent starts completely fresh. Knowledge discovered during one task is invisible to the next. Real example from #601: Task 9 discovered that Drizzle ORM wraps PostgresError inside DrizzleQueryError (check error.cause). Task 10 hit the same issue independently, debugged it from scratch, and fixed it again. Wasted tokens and time.

Two independent users confirmed this in #601. @Koroqe built a file-based workaround in their own project to fix it.

What does this PR change?

Adds a Project Discoveries field to the implementer report format and controller logic to accumulate and inject those discoveries into subsequent subagent prompts. 16 lines across 2 files.

Is this change appropriate for the core library?

Yes. This affects the SDD controller-subagent coordination pattern, which is core infrastructure. It follows the existing "controller curates context for subagents" architecture. No new concepts, no new files, no domain-specific behavior.

What alternatives did you consider?

1. File-based persistence (like @Koroqe's workaround): Have subagents write discoveries to a file and later subagents read it. Rejected because it adds filesystem state and breaks the controller-curates-context pattern. The controller already constructs prompts - this extends that.
2. Full project briefing system (test commands, patterns, imports): More comprehensive but larger scope. This can be a follow-up if discoveries prove useful.

Does this PR contain multiple unrelated changes?

No. Both changes serve the same purpose: one adds the report field, the other adds the accumulation and injection logic.

Existing PRs

• I have reviewed all open AND closed PRs for duplicates or prior art
• Related PRs: feat(subagent-driven-development): add fresh-eyes review step for cross-task issues #738 (fresh-eyes review - different approach to cross-task learning), Add CLAUDE.md reading instructions to subagent prompts #804 (CLAUDE.md reading - different kind of context injection)
• No PR implements the specific discovery-accumulation approach from feat(subagent-driven-development): accumulate learnings across tasks and feed to subsequent subagents #601.

Environment tested

Harness	Harness version	Model	Model version/ID
Claude Code	2.1+	Claude	claude-opus-4-6

Evaluation

The initial prompt was reviewing #601 and proposing a minimal implementation. This is a Markdown instruction change, so evaluation is structural: verified frontmatter parses correctly, template fences are balanced, report format includes the new field, and the new SKILL.md section follows existing voice and structure. The change is small enough that its correctness is verifiable by reading the 16-line diff.

Rigor

• If this is a skills change: I used superpowers:writing-skills and completed adversarial pressure testing (paste results below)
• This change was tested adversarially, not just on the happy path
• I did not modify carefully-tuned content (Red Flags table, rationalizations, "human partner" language) without extensive evals showing the change is an improvement

No carefully-tuned content was modified. The change adds a new section and a new report field. Adversarial check: the Project Discoveries field is marked optional so subagents that have nothing to report skip it without friction.

Verification

Frontmatter valid: name=subagent-driven-development
Template well-formed: 2 backtick fences (balanced)
Report format includes all fields including new Project Discoveries
New section "Accumulated Discoveries" found at line 14 and 21 of SKILL.md

Human review

• A human has reviewed the COMPLETE proposed diff before submission

Addresses #601.

This contribution was developed with AI assistance (Claude Code).