🌱 Plan rung: viewer wording, and the model without its routing envelope

[ibolton336]

What a viewer reads today

Watching a coolstore run on the ROKS demo cluster, the middle rung of the
plan ladder says:

Agent works its standing prompt (us.anthropic.claude-sonnet-4-5-20250929-v1:0, turn 17 of 200)

Two problems, both only in the text:

• "standing prompt" is our word, not the viewer's — and it collides with
  the run's own prompt, which is the thing they might expect to be quoted.
  CONTEXT.md already defines Agent.spec.prompt as "standing instructions
  for how the agent operates", so the rung now uses the glossary's word.
  (The quoted-task branch, Agent works the task: "…", is unchanged.)
• The model id is a routing envelope. us. is the cross-region inference
  profile's geo group, anthropic. the vendor namespace, -v1:0 the API
  version — none of it says which model is answering, and the rung is the only
  place a viewer sees the model while a run is going (the console shows it at
  Gateway-selection time, not on the run page).

After

Agent works its standing instructions (claude-sonnet-4-5-20250929, turn 17 of 200)

modelDisplayName keeps the model's identity — the name and the snapshot
date — and drops only the envelope. It is deliberately narrow:

in	out
us.anthropic.claude-sonnet-4-5-20250929-v1:0	claude-sonnet-4-5-20250929
anthropic.claude-3-5-sonnet-20241022-v2:0	claude-3-5-sonnet-20241022
gemini-2.5-pro, gpt-4.1	unchanged — a version in the name is not a vendor namespace (the segment must be all letters to come off)
ai21.jamba-1-5-large-v1:0	ai21.jamba-1-5-large — digits in the segment, so the namespace stays
publishers/anthropic/models/… (Vertex)	unchanged — path-shaped refs are left alone

Tradeoff worth a reviewer's eye: the geo prefix is information, and two
profiles that differ only by geo now render identically. I dropped it because
the run already pins one Gateway, whose spec.model.name and endpoint carry
the full id, as does the pod's KONVEYOR_LLM_MODEL. Say the word and I'll keep
it.

Wording alternative

If "standing instructions" still reads as ours rather than theirs, the other
candidate is plain Agent is working — the no-instructions branch has nothing
to quote, and the rung already carries the stage, model and turn. I went with
the glossary term because it keeps the distinction between "this run was given
a task" and "it is working from the Agent's own instructions". Easy to switch.

No changelog fragment

The plan ladder itself (#230, merged 2026-09-15) is not in v0.11.0-alpha.4
(cut 2026-09-11) — planTaskRung does not exist in that tag. This wording has
therefore never been in a release, so there is nothing to tell users they need
to un-learn; 230-harness-plan-rung.yaml still describes the shipped behaviour
accurately.

Verification

• cd harness && go test ./... — pass; TestPlanTaskRung gains a
  cross-region-Bedrock case (the exact ROKS string above), new
  TestModelDisplayName covers the table above plus "" and a bare "us."
• go vet ./..., make lint — clean
• Not seen in a console against a live run: the ROKS cluster runs the published
  image, not this branch.

🤖 Generated with Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 6569441d-4df2-4e13-bc2a-be6a554920e3

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}