Skip to content

feat(ce-run,plan): wrapped-phase recognition hook + Erin phase-isolation plan#9

Merged
jcasimir merged 2 commits intomainfrom
feat/erin-phase-isolation
May 8, 2026
Merged

feat(ce-run,plan): wrapped-phase recognition hook + Erin phase-isolation plan#9
jcasimir merged 2 commits intomainfrom
feat/erin-phase-isolation

Conversation

@jcasimir
Copy link
Copy Markdown
Member

@jcasimir jcasimir commented May 8, 2026

Summary

  • ce-run/SKILL.md gains a small Step 5 branch: when a phase has wrapped: true in its frontmatter, ce-run yields control to the orchestrator's wrapped-phase behavior section instead of invoking the skill in-thread. ce-run does not define a generic primitive — it only recognizes the flag.
  • Bundles the brainstorm, plan, and the Unit 1 platform spike (depth-2 Agent dispatch + workaround findings).
  • Plan revisions bake in the spike's Workaround A: the wrapped /ce:work invocation is pinned to Inline execution strategy and forbidden from invoking /ce:review or other Agent-dispatching skills, since subagents in Claude Code lack the Agent tool.

Pairs with JumpstartLab/ce-reviewers-jsl#7. Both must merge before /ce:refresh + plugin reinstall picks up the new behavior end-to-end.

Test plan

  • Merge both PRs
  • Reinstall/update compound-engineering plugin from the marketplace
  • /ce:refresh to sync erin.md
  • Dogfood: run /ce:run erin <real feature> end-to-end (Unit 4)
  • Capture observations in docs/solutions/2026-05-XX-erin-wrapping-dogfood.md

🤖 Generated with Claude Code

jcasimir and others added 2 commits May 7, 2026 23:27
Three-pass refinement cycle for the Erin phase-isolation feature:
brainstorm → plan → plan-review → Unit 1 spike. The spike caught
the load-bearing platform constraint (subagents lack the Agent
tool, so direct depth-2 dispatch fails) before any code shipped.

Two workarounds verified in the findings doc:
  A. Constrain wrapped /ce:work to inline-only execution + no
     Agent-dispatching sub-skills. Sufficient for v1.
  B. Subprocess via `claude -p` to bypass the constraint when
     nested dispatch is genuinely needed. ~16s overhead, streaming
     fidelity regression, but unlocks arbitrary nesting.

Plan v1 status: active, proceeding with Workaround A. Units 2–4
(ce-run hook, erin.md update, dogfood) still ahead.

Compound artifact: tests/spikes/depth-2-dispatch.md is a durable
regression harness — re-run on any future Claude Code update; if
it ever produces 5/5 successful dispatches the platform constraint
has lifted and the original architecture becomes available.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
ce-run/SKILL.md gains a small Step 5 branch: when a phase has
`wrapped: true`, yield to the orchestrator persona's wrapped-phase
behavior instead of invoking the skill in-thread. ce-run does not
define a generic primitive — it just recognizes the flag.

Plan revisions bake in the Unit 1 spike's Workaround A: the
wrapped /ce:work invocation is pinned to Inline execution strategy
and forbidden from invoking /ce:review or other Agent-dispatching
skills, since subagents lack the Agent tool.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
@jcasimir jcasimir merged commit 1405a64 into main May 8, 2026
2 checks passed
@jcasimir jcasimir deleted the feat/erin-phase-isolation branch May 8, 2026 17:22
@github-actions github-actions Bot mentioned this pull request May 8, 2026
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant