Story-3272 Walkthrough

Outcome: the agents an Indian lawyer talks to are now named Kabir, Zoya, Neel and Vani. Nothing an engineer looks at moved: Logfire, agent keys, directories, envelope contracts and identity.py still say Blake, Vega, Indigo, Vani. That is not an oversight — it is the whole design.

The one idea: two namespaces that disagree on purpose

An agent has a permanent ROLE and a permanent code identity. Its PERSONA — the first name a lawyer sees or hears — is a regional presentation skin. India is the only live region, so there is exactly one skin.

Namespace	Spoken in	The cast
Engineering (global)	Logfire `service_name` + `gen_ai.agent.name`, `identity.py` `display_name`, `<agent>.yaml` `identity.name`, eval-suite names, dashboards, docs, PRs, agent keys, directory names	Blake · Vega · Indigo · Vani
Counsel-visible (regional)	system prompts, `SOUL.md`, model-facing tool descriptions, product copy	Kabir (registrar) · Zoya (framer) · Neel (indigo) · Vani (mira)

Do not "align" them. A PR that makes display_name agree with the map has broken one of the two namespaces — every saved Logfire query and every eval-kit lookup keys on the global name. The map itself carries this warning in its module docstring so nobody has to rediscover it.

Where the names live now

agents/_shared/personas.py          ← the ONLY place a persona first name is written
  COUNSEL_FACING_PERSONAS = {
      "mira": "Vani", "registrar": "Kabir", "framer": "Zoya", "indigo": "Neel",
  }
  persona(agent_key) -> str           ← called at IMPORT time, never per run

  agents/<agent>/SOUL.md              ← carries a {persona} slot, filled at load
  agents/<agent>/prompts/*.py         ← f-string {PERSONA}, or .replace() where the
                                        prompt contains literal JSON braces

Keyed by machine agent key (the directory name / identity.key), never by persona — so the key that resolves a name is the one thing that never changes.

What is deliberately absent

No region layer. No region parameter, no env var, no SSM lookup, no fallback chain. One plain dict. That module is the future region hook: a second region forks there and nowhere else, because no prompt inlines a literal. Building the layer now is exactly the speculative machinery AGENTS.md forbids.
No amira entry. The drafting colleague is a future persona with no runtime. A key for an agent that does not exist is dead machinery.
No pronoun edits. Genders match seat-for-seat; nothing needed changing.

Prompt caching: one byte change, fleet-wide, accepted

Every persona name is resolved once at import and interpolated into a module-level constant, so the stable prompt prefix is still byte-identical run to run (AGENTS.md CACHING-FIRST PROMPTS). What did change is the prefix itself, once, for every agent — a cold cache on the next deploy. Founder ruling: expected, note it, do nothing about it.

Re-enrichment: bumped, not suppressed

A rename is a prompt change, so the prompt versions bumped normally and staging re-enriches — free evidence that the renamed prompts work, with zero clients on the line.

Prompt	Version
Zoya proposer / critic / synthesizer / triage	v3 / v3 / v3 / v6
Kabir card (card_engine lineage)	`registrar-card-v3 → v4`
Kabir `CARD_SYSTEM` (agent.py lineage)	`registrar-card-v2 → v5` *
Kabir carve	`registrar-carve-v1 → v2`
Kabir segment	`registrar-segment-v2 → v3`
Kabir family proposer / judge	`…-v2 → v3` (both)
Vani converse	`mira.converse.v17 → v18`
Neel (all four passes)	auto — Indigo mints its version from the composed text

* _CARD_PROMPT_VERSION skips v3 and v4 because two different prompts share the registrar-card-vN string namespace (pre-existing); stepping past them keeps each version meaning one thing.

Not bumped, and why: registrar-classify-v1 and registrar-label-v1 are Pydantic Literal pins on four emitted artifact contracts (classify_models, govern_models, batch_models ×2, revert_models). This repo has never bumped them on a prompt edit — the classify prompt changed in #158 without one — so bumping now would be a wire-contract change, not the normal bump the ruling asked for. Flagged rather than done silently.

The one behaviour change counsel will notice

counsel: "tighten the issues section"
  → Vani calls request_frame_update
  → runtime answers exactly:  I've asked Zoya.      (was: I've asked Vega.)

Vani herself is unchanged. What changed is the name of the colleague she delegates to. FRAME_UPDATE_ACKNOWLEDGEMENT is now derived from the map, and the tests that pinned the old literal assert the constant instead — so the string can never drift from the prompt that teaches it.

A silent-drop this change found and fixed

Three registrar prompts carry literal JSON braces, so they cannot be f-strings and use "...".replace("{persona}", PERSONA). That turns the assignment into an ast.Call, and the lab's prompt inventory only recognised ast.Constant and ast.JoinedStr — so carve.py dropped out of the prompt lens entirely, silently. The governance test caught it. Fixed at the root in evalkit/lab.py: the parser now unwraps a .replace() chain, so the lens renders the filled prompt.

What the corpus kept

Adversarial .jsonl and dataset inputs still say Blake / Vega — the corpus is a ratchet, and an attack string is evidence of what was tried, not product copy. No golden needed regenerating: no counsel-facing generated-output assertion carried a persona name. Zero .jsonl files changed.

Verification

rg -n "You are (Blake|Vega|Indigo)" agents/     → 0 hits

ruff check agents/ evalkit/ tests/ scripts/    → All checks passed

pytest tests/                    457 passed
pytest agents/framer/tests       164 passed
pytest agents/indigo/tests       326 passed
pytest agents/registrar/tests    521 passed
pytest agents/mira/tests         395 passed
pytest agents/context_judge/tests 58 passed
pytest agents/atlas/tests        516 passed

Re-measured, not relaxed: the mira tool-wire token deltas moved 2505→2508 and 2278→2279, and the omitted-channel prompt hash was re-pinned. "Vega" and "Zoya" are both four characters but do not tokenize alike — the byte length of the prompt is unchanged at 20,990, only its bytes are.