Story-2083 Walkthrough

Change: This is S-8 of the agent prompt-caching build (Epic-2050, RFD docs/notes/rfd-agent-caching-strategy.md §7.3 / §4). S-2 landed the wiring primitive (model_settings(seat), cache_point(seat), session_id plumbing) but left the panel call sites on the plain ZDR_SETTINGS. S-8 flips Mira's command panel onto that wiring: the four panel seats (the proposer roster, the critic, the synthesizer) now carry the within_run cache policy on every call, keyed by session_id = run_id. Ambient agents (Vega/Blake) are untouched — they never pay cache writes (that is S-9).

What within-run buys (honesty over hope, §7.3)

Caches are per-provider-per-model, and a panel run is one call per model — so cross-seat reads on today's diverse roster are ~zero. The two real within-run wins are: a same-model pair (e.g. minimax as both a proposer and the synthesizer) reading the shared context prefix warm, and a retried seat re-sending its identical prompt. This story ships the wiring for both; it does not promise a hit-rate (the S-4 live probe measures realized reads). What pinning always buys the panel is correctness-shaped: ONE nonce and ONE session_id per run, so every seat pins one frame — provenance never forks between draft and final.

What ships

File	Role
`agents/mira/panel_prompts.py`	The three seat builders (`build_proposer_prompt`/`build_critic_prompt`/`build_synthesizer_prompt`) now return a `list[str \| CachePoint]` sandwich: the shared context prefix, a `cache_point(seat)` breakpoint, then the per-seat suffix (drafts / critique / instruction). `prompt_text()` is the plain-string view for structural tests.
`agents/mira/panel.py`	Each seat call passes `model_settings(seat, session_id=run_id)` (ZDR + the seat's cache breakpoint + the sticky `x-session-id` header); the `@cache`d role-agents bake `model_settings(role)` instead of `ZDR_SETTINGS`. `run_command_panel(data, *, session_id=None)` threads the run's sticky key to every seat and its retries.
`agents/mira/command.py`, `agents/mira/runtime.py`	`run_command(data, *, run_id=None)` passes the `agent_run_id` down as the panel `session_id`. The broker seam already carries `run_id`; it just reaches the panel now.
`agents/mira/tests/test_panel_cache.py` (new)	The acceptance suite: a CachePoint on every seat prompt (absent when the seat is `off`); `session_id = run_id` + ZDR on every call; a retried seat re-sends the same cacheable prompt under the same session; a different run gets a different session (cold across runs); the shared context prefix is byte-identical across seats.
`agents/mira/tests/conftest.py`, `agents/mira/evals/adversarial.py`	`_user_text` now flattens a `[str, CachePoint, str]` UserPromptPart (pydantic-ai stores a sequence prompt as one part with list content), so the panel stub + the injection-wall check still read all walled text through the sandwich shape.

The sandwich breakpoint (§5.1 placement)

The breakpoint sits at the stable/volatile boundary: the shared context (matter meta, command, current frame, locks, notes, docs) is byte-identical across every seat and across a seat's retries; the per-seat suffix is not. cache_point(seat) returns None when the seat is cold (policy off), so the prompt shape is unchanged when caching is off — caching is economics, not correctness.

def _sandwich(seat, context_parts, suffix_parts):
    seq = ["\n\n".join(context_parts)]
    bp = cache_point(seat)          # None when the seat is cold -> no marker, no drift
    if bp is not None:
        seq.append(bp)             # closes the shared, cacheable prefix
    seq.append("\n\n".join(suffix_parts))
    return seq

session_id = run_id: warm within, cold across

One session_id per run (the agent_run_id) rides the sticky x-session-id header on every seat call and its retries, so a re-send lands on the same provider endpoint and reads the cached prefix warm. It is never carried across runs — each run has a fresh agent_run_id — so cross-run seats stay cold. That is the within_run policy, not designed cross-run caching (§4). ZDR rides the same settings object, so cache can never be split from ZDR.

Acceptance criteria → tests

AC Proof

provenance still one version_hash across seats test_shared_context_prefix_is_identical_across_seats — one nonce, one context prefix for proposer/critic/synth, so every seat pins one frame.

measured within-run reads reported (even if ~0 — honesty over hope) The wiring is asserted, not a hit-rate: test_session_id_is_run_id_on_every_seat (marker + sticky key + ZDR on every seat) and test_retry_re_sends_the_same_cacheable_prompt_within_the_run (the retry re-uses the identical CachePoint prompt under the same session). Realized reads are the S-4 live probe's job; the S-5 model.usage cache fields (merged, pramaan-agents#201) carry them on the wire when they land.

AC	Proof
provenance still one `version_hash` across seats	`test_shared_context_prefix_is_identical_across_seats` — one nonce, one context prefix for proposer/critic/synth, so every seat pins one frame.
measured within-run reads reported (even if ~0 — honesty over hope)	The wiring is asserted, not a hit-rate: `test_session_id_is_run_id_on_every_seat` (marker + sticky key + ZDR on every seat) and `test_retry_re_sends_the_same_cacheable_prompt_within_the_run` (the retry re-uses the identical CachePoint prompt under the same session). Realized reads are the S-4 live probe's job; the S-5 `model.usage` cache fields (merged, pramaan-agents#201) carry them on the wire when they land.

Verification

uv run --package mira pytest agents/mira/tests -q          # 139 passed
uv run --package mira python agents/mira/evals/run.py adversarial   # 13 held / 0 failures
uv run pytest tests -q                                    # 175 passed (governance)
uv run ruff check agents/mira tests/ scripts/             # All checks passed!

Enablement only, no new egress: no CodeZip rebuild or AgentCore redeploy needed for the wiring; the panel emits cache markers on the wire now, but reads are earned only where a provider endpoint serves a cache (S-4). Ambient agents (Vega/Blake) are untouched by this story — their no-paid-writes audit is S-9. llm.yaml already carries the panel seats' within_run policy from S-1; this story consumes it.