Owner directive (2026-08-27)
The harness ships in two flavors. This is the product shape, not a fallback arrangement:
|
Workhorse (stable) |
Frontier (experimental) |
| Package |
v1.87.0 via opus-4.6 dist-tag |
current main / @latest |
| Driver |
claude-opus-4-6[1m], effort max (no xhigh on 4.6) |
Opus 5 |
| Reviewer |
GPT-5.5 at its ceiling (probe-confirmed it still resolves) |
GPT-5.6 Sol |
| Design authority |
— (escalation only) |
Fable 5 |
| Escalation valve |
Fable / GPT-5.6 for genuinely hard, complex problems only |
n/a (already frontier) |
Rationale, in the owner's words: the frontier models "get lost in harnesses"; the 4.6 + 5.5 pairing is what they run at work daily and is the most reliable with an SDLC like this. Two flavors is the count — Sonnet already exists in AI_SETUP_LANES.md as the budget lane; no third flavor now.
Seat provenance — read this before editing any doc
The stable flavor's seats are the owner's live work config, not what v1.87.0 shipped. Both facts, side by side:
- v1.87.0's own shipped docs name GPT-5.6 as the reviewer — verified from the tag.
- The owner runs GPT-5.5 at work and finds it reliable — their call, and it defines the flavor.
Do not retro-claim the package shipped 5.5. The resolution here is "owner owns the flavor," not "normalize the record."
Blocking unknown: the [1m] suffix
The bare claude-opus-4-6 pin measured a 200k window on 2026-08-26 (#629, #697). The owner's work form is claude-opus-4-6[1m], which may be the 1M launch path #689's matrix is missing. Owner-observed, unmeasured: "~250k autocompact, ~350k max." Test: scratch session → /model claude-opus-4-6[1m] → /context → paste verbatim here.
Acceptance criteria
Relations
#689 (safe line — this formalizes what it protects), #694 (frontier audit — its subject is the Frontier flavor), #697/#698–#702 (consumer defects both flavors inherit).
Owner directive (2026-08-27)
The harness ships in two flavors. This is the product shape, not a fallback arrangement:
opus-4.6dist-tag@latestclaude-opus-4-6[1m], effortmax(no xhigh on 4.6)Rationale, in the owner's words: the frontier models "get lost in harnesses"; the 4.6 + 5.5 pairing is what they run at work daily and is the most reliable with an SDLC like this. Two flavors is the count — Sonnet already exists in AI_SETUP_LANES.md as the budget lane; no third flavor now.
Seat provenance — read this before editing any doc
The stable flavor's seats are the owner's live work config, not what v1.87.0 shipped. Both facts, side by side:
Do not retro-claim the package shipped 5.5. The resolution here is "owner owns the flavor," not "normalize the record."
Blocking unknown: the
[1m]suffixThe bare
claude-opus-4-6pin measured a 200k window on 2026-08-26 (#629, #697). The owner's work form isclaude-opus-4-6[1m], which may be the 1M launch path #689's matrix is missing. Owner-observed, unmeasured: "~250k autocompact, ~350k max." Test: scratch session →/model claude-opus-4-6[1m]→/context→ paste verbatim here.Acceptance criteria
[1m]measurement recorded here verbatim[1m]resultRelations
#689 (safe line — this formalizes what it protects), #694 (frontier audit — its subject is the Frontier flavor), #697/#698–#702 (consumer defects both flavors inherit).