Prompt Stack

Where this fits

Prompt Stack is the before-model observability surface in the conversation quality lifecycle. Use it when you need to see what BMO assembled before the model ran — not after a draft exists.

Changes live output?

No. The ring and status JSON are read-only metadata. They never change what the model receives on a future turn.

Reach for it when

A response surprises you and you want to confirm what context the agent was given — for example checking whether a mode, tool, or memory segment was actually included before blaming the model.

Inspect

Job	Route	What it shows
Live TUI state	/prompt-stack or sidebar Prompt stack	Current process ring, last-build actions, recent build rows, and the next route to Context Lineage.
Full trace inspection	TUI Context Lineage	Source classes, included/shed/empty groups, boundary details, compaction, and segment text for the current or last turn.
CLI process view	bmo config show-prompt-stack	The same human-readable status for the current process.
Agent JSON status	get_prompt_stack_status	Bounded posture/status JSON for automation.
MCP JSON status	bmo_get_prompt_stack_status	Same bounded JSON shape as the agent status tool.
HTTP JSON status	GET /v1/prompt-stack/status	App-owned status snapshot for live runtimes.
Recent-build export	list_recent_prompt_stack_builds	Strict recent-build metadata for agent workflows.

The trust boundary is intentional: the ring, slog events, HTTP status, MCP tools, and agent-tool JSON carry metadata only. Use Context Lineage for richer trace inspection; do not expect those status surfaces to export full prompt bodies or durable cross-process history.

Prompt-stack segment actions are a finite reviewed taxonomy: included, shed, empty, disabled, denied, mode_conflict, context_escape, and phase_error. Segment reasons are bounded to reviewed families such as shed:token_budget, disabled:policy, denied:policy, and unreviewed; status surfaces do not copy raw provider errors, user text, tool output, home paths, or prompt bodies into reason fields.

Prompt Stack inspector

Live BMO TUI: the /prompt-stack readout reports the current process ring, last-build action summary, recent build rows, and the next Context Lineage route. It is the bounded, metadata-only view of what BMO assembled in front of the model, in what order, and why.

Related quality surfaces

• Conversation quality — lifecycle map and pick-your-job tree
• Quality Gates — jump here after a draft exists when you need a rubric verdict
• Shadow Evals — jump here when prompt or routing changes need bmo eval compare before/after evidence; semantic-ready prompt-stack scenarios live under eval/scenarios/prompt-stack/

Deep dive

Every turn records prompt-stack builds through the system prompt assembler. Telemetry uses prompt_stack.fired and prompt_stack.action slog events plus a process-scoped ring (capacity 12). Implementation detail: prompts topic.

See also: Prompt Enhancer (user-message context injection).