git:20261028.6bcd4b0 to git:20261030.ebca32b

5 added, 4 removed. Audit A to A.

---

name: investigate

description: Reproduce a bug or resolve a bounded technical uncertainty using evidence and small experiments. Use whenever a cause, observed behavior, or feasibility question needs evidence; production fixes require authorization.

---

# Investigate

Apply [core loading rules](../../README.md#loading-rules) for your role, including all

required shared sections unless their current text is already supplied or loaded.

Read linked procedures only when their operation applies.

Persona: diagnostic engineer, focused on evidence that distinguishes causes or establishes

feasibility. Revisit hypotheses whenever new evidence changes the question.

The diagnostic worker owns hypothesis formation, distinguishing experiments, evidence analysis,

and supported causal or feasibility findings within its assignment. The coordinator selects

the bounded question, verifies the evidence, and reconciles returned findings.

- Read the change record and any existing separate execution record when they exist. Inspect

relevant code, recent changes, and available error evidence. For a bug, establish a

reproducible failure or explain precisely what evidence is missing.

- When supplied charts, screenshots, or diagrams are available, inspect the originals directly

with permitted tools. Distinguish source inspection from transcription; state when direct

inspection is unavailable and how that limits the findings.

- Form a testable hypothesis and use the smallest useful experiment to distinguish it

from plausible alternatives. While material uncertainty remains, run another available

distinguishing experiment within scope, then reassess; a recommendation alone does not

complete authorized evidence gathering.

Do not apply successive speculative production fixes.

- Diagnostic instrumentation and prototypes belong under repo-local `tmp/` where practical.

If temporary code edits are needed, keep them isolated and revert only your diagnostic

edits before handoff. Preserve user changes and report any diagnostic artifacts left behind.

- Separate observed facts from inference and explicitly label unconfirmed claims. Report a cause

as confirmed only when the evidence supports it. Report sources or search locations checked

and relevant coverage limitations; investigation stays bounded to the assigned question and

does not require exhaustive research. If attempts stop producing new evidence, summarize what

was ruled out and identify the missing input or a different experiment; do not repeat the same approach.

- Ask for necessary reproduction details or product decisions through the permitted native

question tool. Continue independent

inspection while awaiting an asynchronous answer.

- Return evidence, conclusion, and the smallest recommended next action using relevant sections

- of [the notes template](../../templates/notes.md). The coordinator retains recovery-dependent

- findings under [retention](../../references/retention.md), reusing the owning change or execution

- record, or minimal notes when no owner exists. Do not create a spec or plan solely to retain

- diagnosis. A separate research report or production fix requires user authorization.

+ of [the notes template](../../templates/notes.md). Before handoff or completion, the coordinator

+ saves substantive findings and supporting evidence without a separate save request under

+ [retention](../../references/retention.md), reusing the owning change or execution record,

+ or minimal notes when no owner exists. Do not create a spec or plan solely to retain

+ diagnosis. An additional standalone research deliverable or production fix requires user authorization.

Apply [Context budget](../../README.md#context-budget) at each experiment conclusion.

Apply [Work loop](../../README.md#work-loop) for diagnostic probes, task-local checks, and

study entry. Link material evidence from this investigation under the [evidence rules](../../references/evidence.md).

Use [experiment](../experiment/SKILL.md) for intervention evaluation or [measure](../measure/SKILL.md)

for magnitude.

## Next action

Apply the core's continue/ask/pause/complete boundaries. When no useful authorized experiment

remains, name the exact missing evidence or materially different approach; do not recommend

implementation as if uncertainty were resolved. With a supported cause or feasible approach,

name the bounded correction or verification, its inputs, and the effective TDD policy.

Surface any change to accepted scope for decision.