Installed Execution Compositions in DocsFab
Look up the installed DocsFab model and serving-route records. To select one, follow Select an Installed Composition. To change its defaults, follow Override Model Settings.
Installed Compositions
An execution composition binds a model, serving route, authentication route, transport, harness, and exact adapter versions. List the package-installed records before selecting one:
The executable compositions use Pi as a common harness. The following table is the current installed registry, not a release catalog.
The supported first-use path configures a model and endpoint that you choose through Install DocsFab. DocsFab has no default model. No composition in this table is eligible for release, and each is an explicit execution path with its own credential contract.
nvidia-inference-gpt-6-sol-pi is the composition that DocsFab’s own
live-model CI jobs bind. Its recorded evidence is diagnostic only: a bounded
probe on September 25, 2026 returned the harness-compatible verdict, and
diagnostic matrix runs on an operator route, not this sealed composition,
scored 24 of 24. No acceptance, cost, or release qualification has run for
this sealed composition. Select it with
docsfab setup --profile nvidia-inference-gpt-6-sol-pi.
Public installations omit private Inference Hub endpoint defaults. Set the
provider’s actual endpoint through MODEL_BASE_URL or
host.yaml active.writing.settings.baseUrl before selecting one of these
compositions. Keep an environment-supplied endpoint exported for later commands;
setup --profile does not save it. Refer to
Select an Installed Composition
for the persistent configuration procedure. No profile YAML copy is required.
nvidia-inference-claude-sonnet-5-pi remains installed as a
secondary option on the same route and credential. It has known workflow gaps,
tracked in issue #1428: provider stops during first-pass decomposition,
evaluation steps that do not finish within the acceptance time window, and
output truncation at its default 32,768-token output limit.
DocsFab sends thinking: {"type": "disabled"} on every Claude Sonnet 5
request whose reasoning is off, through the
nvidia-inference-claude-sonnet-5-pi composition’s own route or any other
endpoint. It sends reasoning_effort only when you set reasoningEffort,
which turns reasoning on and removes the thinking field. With reasoning
set to true and no effort, it sends neither field, so the model’s default
thinking applies; refer to
Change a Preset’s Settings. On this route,
Claude Sonnet 5 thinks by default when
a request has no thinking field. That thinking counts against the
per-request output limit but is not streamed, so one turn can spend its whole
limit and return no text. The composition’s model profile declares
reasoning: false, and the field makes each request match that declaration.
A probe of the route on September 28, 2026 measured the following behavior:
- Without a
thinkingfield, a request for a bare integer used all 16,384 output tokens it allowed and returned no visible text. - With
thinkingdisabled, the same request used 371 output tokens and returned its answer. reasoning_effort: "none"was accepted and had no effect.reasoning_effort: "low"and a request for 64,000 output tokens were refused at once with HTTP 408.
No request that allowed more than 32,768 output tokens reached its limit, so
the composition’s default output limit stays at 32,768 tokens. A
maxOutputTokens setting replaces it. Inference Hub Metadata reports
a 128,000-token output limit for this model, but the route measurement takes
precedence. The
nvidia-inference-claude-opus-5-pi composition does not send the
field, because its route behavior has not been measured.
Inference Hub Token Limits
The NVIDIA Inference Hub records seal the context, input, and output limits from each model’s Inference Hub Metadata record, read on September 28, 2026. A route measurement that contradicts Metadata takes precedence. The following table lists the sealed limits.
The package ceiling is the largest value across all records: 1,050,000 context
and input tokens and 384,000 output tokens. These limits are each composition’s
defaults. A contextWindow, maxInputTokens, or maxOutputTokens setting
replaces the composition’s value, above or below it, and DocsFab records and
warns about the difference.
nvidia-inference-deepseek-v4-pro-pi binds the model ID
nvidia/deepseek-ai/deepseek-v4-pro and carries no installed request policy.
On September 28, 2026, a probe sent that model ID 2 prompts with a 128-token
output limit and no tool prompt; both returned content without a reasoning
split. The model ID has not been checked with a tool prompt or a longer
answer. The former model ID, azure/deepseek-ai/deepseek-v4-pro, split its
reasoning without a control on August 26, 2026 and returned HTTP 403 on
September 28, 2026.
The NVIDIA API Catalog records do not take their limits from Inference Hub Metadata, so their sealed limits are unchanged.
Select one explicitly on run or do with --composition. Without it, the
run takes the model-profile path:
A retired composition stays listed so that completed run records keep
rebinding to it, but selecting it for a new run stops with
composition-retired and names its successor. A retirement cites one of two
reasons. provider-end-of-life quotes the provider’s notice.
superseded-by-operator-decision names the person, date, statement, and
record of the decision; DocsFab uses it when the provider still serves the
model. The direct OpenAI GPT-5.6 Sol composition is retired for the second
reason: the provider notice that retired GPT-5.6 in NVIDIA’s Codex
environment does not cover the OpenAI API. The composition-retired refusal
names only the reason code and the effective date. The rest of the retirement
record stays in the installed registry as internal audit metadata.
docsfab composition list --json shows only the retirement reason,
effectiveAt, and optional successorCompositionId.
The selected model ID, endpoint identity, transport, and credential family must
match the installed composition. DocsFab rejects missing or ambiguous
credential sources and does not fall back to another composition.
Composition-execution writes stay local unless you pass --draft-mr.
Composition-execution results permanently report technicalApproval: not-established,
roleQualification: not-established, and releaseEligible: false. A run that
names --composition can also use --release-gate, --comment-mr, and
--resume-from. Without --composition, a run uses the composition that
host.yaml names, or the configured model profile; it does not make a new
approval or release claim.
Moved Topic Links
Earlier section bookmarks lead to the canonical topic below.
Refer to Override Model Settings in DocsFab.
Refer to Read-Loop Supervision in DocsFab.
Refer to Model Setting Precedence in DocsFab.
Refer to Model Setting Precedence in DocsFab.
Refer to Model Setting Precedence in DocsFab.
Refer to Model Setting Precedence in DocsFab.