Degradations
What still works in the AI factory when a capability is absent — the degradation matrix rendered from its single owner, .ai-factory/tier-home.md §3.
Experimental. The factory's pieces do not all have to be present for work to continue — but when a capability is absent, something specific degrades, and the degradation is named rather than papered over. A degradation row answers one question: with X missing, what still works, and what gets worse. The matrix below is copied verbatim — row text unchanged — from the one file that owns it; this page is a pointer, never a second source. Read the rows as evidenced claims, not probe-verified facts: each names the probe class that will validate it later, and that validation has not run yet.
Columns: | Capability absent | Tier-system degradation | Current evidence source | C3-validation status |
Reading note, ours not the owner's: two evidence cells cite CLAUDE.md:108 and
CLAUDE.md:130 — line numbers from before that doc was pointer-ized. The quoted text
now lives in the owner file's own §2 (.ai-factory/tier-home.md §2 on your install,
packages/core/templates/shared/tier-home.md §2 in the framework); the «§2 above»
inside those cells refers to that section, not to this page.
| Capability absent | Tier-system degradation | Current evidence source | C3-validation status |
|---|---|---|---|
no aif runtime (no runtime-bridge / no aif-handoff dispatcher) | Tier 1 (bridge-profile marker mechanic) and Tier 2 dispatch defaults both become unreachable — the packages/runtime-bridge header-region parse and the aif runtime profile config do not exist on the consumer's machine. Tier routing collapses: every non-Tier-0 task degrades to in-session SDD (superpowers:subagent-driven-development invoked directly, no kickoff dispatch, no pipeline). The senior interactive session does the Edit-then-review loop itself rather than dispatching through the aif pipeline. | CLAUDE.md:130 (pre-pointer-ization — verbatim-lifted to §2 above) names packages/runtime-bridge as the marker mechanic home + the aif runtime profile config as the Tier 2 defaults home. Spec §4 A3 (docs/superpowers/specs/2026-07-23-beta-program-design.md:261) names «no aif → in-session SDD». | TO-BE-VALIDATED by C3 — C3 probe class: aif-runtime-absence behavioral enumerator over setup.d/companions.manifest + PROFILE dispatch (verifies the absence propagates correctly through the install surface so the consumer never sees a half-wired dispatch). Built in umbrella C, post-A3. |
| no GLM subscription (no executor-tier model available on the harness) | The executor tier is absent from the available model set. Per night-mode «window slides», tiers slide to the next-cheaper available model: if only the top-tier model is available, Tier 1 work (executor) routes UP to the top tier (cost increase — the cheaper tier was the point of Tier 1); the SDD advisor collapses per the night-mode single-tier-harness rule. Tier routing CRITERIA are unchanged; only the INSTANTIATION slides (the criteria doc still classifies the same way; the harness fills the executor slot with whatever it has). | CLAUDE.md:108 (pre-pointer-ization — verbatim-lifted to §2 above): «Tiers are RELATIVE capability tiers, not hard-coded models … the window slides to whatever the active harness offers». .claude/skills/night-mode/SKILL.md:17: «The window slides to the available set». Spec §4 A3 (docs/superpowers/specs/2026-07-23-beta-program-design.md:262): «no GLM subscription → tiers slide (night-mode posture SSOT)». | TO-BE-VALIDATED by C3 — C3 probe class: model-tier availability enumerator over runtime-bridge/runtime-profiles resolution + AifHandoffBackend.ts two-step resolver behaviour when the named executor profile is absent. Built in umbrella C, post-A3. |
| no Fable (top-tier advisor seat absent — Claude-stack-specific instance of «the top-tier model is unavailable») | The advisor seat (night-mode «top tier → advisor + SDD's final whole-branch review», currently instantiated as Fable on this operator's Claude stack) slides down to the next-available tier. Per night-mode: «Claude with no Fable → [advisor Opus, executor Sonnet, cheap Haiku]» — Opus takes the advisor seat; the tier-routing criteria and the executor tier are unaffected. This is an instantiation-level degradation, not a criteria-level one. | .claude/skills/night-mode/SKILL.md:17: «Claude with no Fable → [advisor Opus, executor Sonnet, cheap Haiku]». Spec §4 A3 (docs/superpowers/specs/2026-07-23-beta-program-design.md:262): «no Fable → Opus tops». | TO-BE-VALIDATED by C3 — C3 probe class: top-tier advisor seat enumerator over runtime-bridge/runtime-profiles + night-mode/SKILL.md advisor-consult mechanism (delta item 7). Built in umbrella C, post-A3. |
| non-CC harness (Claude Code primitives absent — zcode/GLM/other) | Per night-mode/SKILL.md:19 portability table: the loop runs on any harness with sequential subagent dispatch; CC-named primitives degrade gracefully — Workflow (context-economy) → manual summaries-only discipline; ScheduleWakeup (quota-backoff) → harness sleep/resume or manual; isolation:"worktree" (parallel executors) → sequential; subagent hooks (digest/report injection) absent on zcode event set → carried in the dispatch prompt instead. Per-artifact: 16 zcode plugin twins (S1 inventory §1.7) carry the dispatch surface on the plugin channel. Net: only parallel-executor speedup and hook-based context-hygiene degrade. | .claude/skills/night-mode/SKILL.md:19: full portability table (verified 2026-07-04: ~/.zcode/cli/agents/ holds real subagent sessions). S1 inventory §1.7 (docs/meta-factory/research-patches/2026-07-25-beta-a-s1-inventory.md:86-94): 16 zcode twins catalogued. Spec §4 A3 (docs/superpowers/specs/2026-07-23-beta-program-design.md:262-263): «non-CC harness → per-artifact degradations (validated by C3)». | TO-BE-VALIDATED by C3 — C3 probe class: harness-portability degradation enumerator over plugin/hooks/ zcode twins + .claude/rules/zcode-parity-doctrine.md §2 census. C3 verifies the per-artifact degradation claims by exercising the non-CC harness path end-to-end (night-mode §5 declares portability «designed-not-proven» as of S3 stage time). Built in umbrella C, post-A3. |
The authoritative rows live in your install at .ai-factory/tier-home.md §3. That file
installs at env and factory depth only — on a core install it is absent by
design, because tier routing presupposes the multi-model contour that core does not
ship.
Naming the absent machinery precisely, at framework pin b069c593: the runtime-bridge
package's consumer-facing CLI surface is the vendored subset under
.claude/vendor/runtime-bridge/src/cli/ (dispatch, harvest, answer, questions, claim,
ensure-parallel — packages/runtime-bridge/vendor/README.md:28-63), so «no runtime-bridge on
the consumer's machine» means those entrypoints are gone, not merely unconfigured. The
backend they talk to is chosen by RUNTIME_BRIDGE_MODE (manual / aif-handoff / auto;
packages/runtime-bridge/src/resolver.ts:6-9), with ManualBackend as the guaranteed tail
(resolver.ts:11 — "ManualBackend is always the tail — never excluded"). The default
aif-handoff backend dispatches over REST on port 3009 (packages/runtime-bridge/src/AifHandoffBackend.ts:4-5
— "DISPATCH = REST (:3009)"), which is the channel the rows above call the dispatcher.
The «window slides» posture the rows cite is the night-mode skill's — shipped on the
env+ tier arm (setup.d/lib.sh:62), triggered «Use when running a task FULLY AUTONOMOUSLY
(overnight / unattended) as an orchestrator» (.claude/skills/night-mode/SKILL.md:3). Its
model posture assigns each role a relative capability tier, never a hard-coded model name,
and fills the seat with whatever the active harness offers — the window slides to the
available set (.claude/skills/night-mode/SKILL.md:17).
Next
Where the tiers come from: Overview — multi-model pipeline. The install that puts the file on disk: First Steps — env.
Daily cycle — factory
The factory day in four beats — /arch an idea, pick a launch preset, ask /pipeline what to start, dispatch and harvest. The loop from kickoff to finished branch.
FAQ
Direct answers: what an executable AGENTS.md is, whether getff needs an LLM, its license, and how it compares to Packmind, agnix, and CodeRabbit/Qodo.