git:20260921.41cb25e to git:20261020.a2aa9ef

1 added, 1 removed. Audit A to A.

---

name: fable-understand

description: Source-grounded codebase understanding for behavior, architecture, boot flow, data flow, integration wiring, or "how does this work" questions. Use when the user asks Codex to explain a system, trace a path, map a subsystem, or answer behavior questions with citations and unknowns.

---

# Fable Understand

Answer from implementation evidence, not memory or stale docs.

For ECF-style governed runs, use `../../references/ecf-run-contract.md`. For large or high-risk understanding tasks, use real Codex subagents when the runtime exposes a subagent tool and the user has not opted out; treat cross-module architecture, boot flow, data flow, integration wiring, security/privacy/money/data/API, or many-file mapping questions as large by default. Otherwise run `single-agent multi-lens` and say why no subagents were used when workflow trace is requested.

- For large or high-risk Fable tasks, use `gpt-5.6-sol` with Ultra (`model_reasoning_effort = "ultra"`) for the parent coordinator when it is selected or can be configured before the run. Record the requested and actual coordinator model and effort plus any fallback; never imply Sol Ultra was active when it was not. Request parallel delegation for disjoint Fable lenses. Use `gpt-5.6-luna` with `medium` reasoning for every delegated subagent by default, setting the model and effort explicitly on each spawn when the runtime supports it. Keep load-bearing synthesis and final verification with the Sol Ultra coordinator. If the runtime cannot honor the worker selection, report the requested and actual worker model plus the fallback; never claim Luna was used without runtime evidence. Otherwise use `single-agent multi-lens` and report the reason.

+ Use the centralized capability-aware profiles in `../../references/model-support.md` and `../../config/model-profiles.json`. Keep model and reasoning selections configurable, preserve explicit overrides, and record the requested and actual coordinator model and effort and requested and actual worker model plus any fallback. Request parallel delegation only for useful disjoint lenses; report `single-agent multi-lens` when no workers ran. The coordinator owns synthesis and final verification. Never claim a model, effort or subagent ran without runtime evidence.

## Workflow

1. Restate the question and scope.

2. Restate authority boundaries and declare the ECF run mode when the user asks for ECF, subagents, or a receipt.

3. Read repo instructions and the most direct source files.

4. Trace from entrypoint to effects:

- route/command/UI entry

- service/module boundaries

- data reads/writes

- external calls

- errors, retries, and fallbacks

- tests and docs that confirm or contradict behavior

5. Inspect callers and importers before answering behavior questions.

6. Use a small diagram or ordered flow when it improves clarity.

7. Include unknowns, assumptions, and stale-doc risks.

## Evidence Safety

Never print raw secrets, tokens, private keys, wallet keys, credential files, or `.env` values. Redact secret-like values and cite only the file/path/key name needed to explain the issue.

## Output

Prefer this shape:

- direct answer

- source-backed flow

- important edge cases

- unknowns or verification gaps

- useful next probe, only if needed

Every non-obvious claim should have a file, line, command, artifact, or runtime citation.

When requested, include a compact `Workflow Trace` with mode, ECF contract status, lenses covered, spawned agents or no-subagent reason, verification method, and coverage gaps.