Repository navigation
Two-lane model strategy + plugin parity (port from claude-sdlc-wizard v2.0.0) #183
Description
Activity
- changed the title
[-]Two-lane model strategy: Reliable + Frontier (port from claude-sdlc-wizard)[/-][+]Two-lane model strategy + plugin parity (port from claude-sdlc-wizard v2.0.0)[/+]on Sep 27, 2026 Lessons to port from claude-sdlc-wizard (session 2026-09-27)
1. Model IDs have no point releases — use exact IDs
Wrong Correct claude-fable-5-1claude-fable-5claude-opus-5-5claude-opus-5No
.1or.5variants exist for any current Claude model. This issue's body carries the old wrong IDs — fix before porting.2. Per-lane reviewer split (the portable pattern)
The reviewer model differs by lane. The claude wizard now does:
- Reliable: GPT-5.5 xhigh (field-proven)
- Frontier: GPT-5.6 Sol xhigh (newer)
For codex-sdlc-wizard (families inverted):
- Reliable: Opus 4.6 or Fable 5 via
claude -p(field-proven Anthropic reviewer) - Frontier: whatever newest Anthropic model fits
The point: the reviewer is not one-size-fits-all across lanes.
3. Exhaust the driver before escalating (THE key principle)
95% confidence is the escalation threshold, but sub-95% does NOT mean escalate immediately. The driver must actively try to reach 95% first:
- Web search, read more files, grep the codebase, check git history
- Each rung is expensive and more effective when it knows what was tried
- When escalating: pass forward what you researched and what specific question remains
- "I checked X, Y, Z and I'm stuck on this" >> "I'm not sure, please review"
This must be explicit in the shipped skill, not just implied by the 95% threshold.
4. Regression tests must prove they catch regressions (mutation testing)
Every shipped surface (skills, hooks, plugins, docs) needs a test. But a green test suite is worthless unless it goes RED on the wrong version. Proven today:
- Introduced wrong model ID → test caught it (FAIL)
- Swapped lane reviewer → test caught it (FAIL)
- Broke cowork/skill parity → test caught it (FAIL)
- Found one direction NOT guarded → added the test
The codex wizard should have the same: for every regression test, a mutation that proves it fires.
5. Local test runner runs EVERYTHING
No skip gates, no env vars, no API-key checks. One command, all suites. The maintainer has the subs — the runner assumes it.
./tests/run-local.shwent from 6 suites to 84 (2,602 checks, 0 failures).6. GPT-6 models are live in Codex CLI
Codex 0.157.1 shows GPT-6-Astra (default), GPT-6-Sol, GPT-6-Luna. GPT-5.5 is "Legacy." GPT-5.6-Sol is "Older." The codex wizard's lane tables need updating — model mappings TBD by maintainer with Codex directly.
What claude-sdlc-wizard v2.0.0 just shipped
Two lanes with npm dist-tags
Plugin install (same result, no npm):
The portable pattern (builder/reviewer/advisor + 95% escalation ladder)
Escalation ladder:
Each rung passes forward what it learned. No model redoes work.
claude-sdlc-wizard model table
@latest)@frontier)max(Anthropic)xhigh(Anthropic)xhigh(OpenAI)xhigh(OpenAI)high(Anthropic)high(Anthropic)codex-sdlc-wizard — THE INVERSE
Builder is OpenAI. Reviewer is Anthropic (cross-family diversity). Families flip:
@latest)@frontier)xhigh(OpenAI)claude -pWhat needs porting
Two-lane dist-tag infrastructure —
release.ymlwithscripts/derive-dist-tag.sh(already proven in claude-sdlc-wizard #713). Prerelease versions get--tag frontier, stable gets--tag latest.Model table in all shipped docs — exact model IDs, exact effort levels, no aliases. "GPT-5.6 Sol xhigh" not "Sol" or "Codex high".
Brain escalation ladder — documented in wizard doc, skills, tested in
tests/test-escalation-ladder.sh(live E2E, isolated sessions). 95% confidence threshold documented in 3+ shipped files.Plugin parity (#725 on claude-sdlc-wizard) —
codex --pluginshould give the same experience as npm install.Model pin tests —
tests/test-model-pins.sh(6 assertions: frontier model present, reviewer present and defaulted, advisor present, no stale refs).Reliable = what you actually use
Same principle: Reliable is what the maintainer dogfoods. Frontier tracks newest models for experimenting. Owner will figure out the Codex-side model pins interactively.
Reference