docs(think): move developer content out of the LLM-facing docstring (#320)

[Ashutosh0x]

Addresses task 1 of #320 (@notowen333: "Clean up doc strings in community tools ... Task 1 is eligible for community contribution"), starting with the think tool that @jsgsamzn used as the reproducer.

Problem

ToolSpec.description is built by _extract_description_from_docstring (strands-py/src/strands/tools/decorator.py), which strips only the Args block - its own docstring says it preserves "Returns, Raises, Examples, and other sections". So Python usage examples, STRANDS_* environment variable docs, internal "How It Works" steps and developer notes all ship to the model on every request.

Measured on main by replicating that function against the current sources:

tool	description	approx tokens	Python examples reach the LLM
think	6,017 chars	~1,500	yes
swarm	5,423 chars	~1,355	yes
graph	3,511 chars	~877	yes
shell	1,721 chars	~430	no

shell is the outlier because it already does the right thing - developer usage lives in the module docstring, and the @tool docstring is written for the model. That is Solution 1 from the issue, and this PR applies it to think.

Change

Purely a docstring reorganisation in src/strands_tools/think.py:

• Moved to the module docstring - "Usage with Strands Agent" Python examples, "How It Works", "Model Selection Process", "Common Usage Scenarios", the STRANDS_* configuration block, and the developer notes. Nothing is deleted; it moves to where a developer reading the source will still find it.
• Kept in the @tool docstring - what the tool does, when a model should reach for it, the recursion constraint, all eight Args, and Returns.

Result: tool_spec description 6,017 -> 598 characters (-90%), with no Python examples and no environment variables left in it.

Parameter descriptions are unaffected. They are parsed separately into inputSchema via param_descriptions, so the model keeps full per-argument documentation - this only removes the prose the model was never meant to read.

While moving it I also fixed a broken snippet: the old thinking_system_prompt=Use Socratic questioning method: example was missing its quotes.

Validation

• ast.parse clean
• All 8 signature parameters still documented, no phantom entries (checked programmatically against the signature, so inputSchema is byte-identical)
• Before/after description sizes measured by replicating _extract_description_from_docstring exactly
• No test asserts on this docstring (tests/test_think.py has no description/tool_spec/__doc__ references)

I could not install strands-agents in my environment to render the real tool_spec, so the figures above come from replicating the extraction function rather than executing the decorator. Worth a sanity check in review, though the parity check on the signature means inputSchema cannot have shifted.

Scope

Deliberately one tool. swarm and graph have the same problem and I am happy to extend this PR to both, but the exact voice of an LLM-facing docstring is a judgement call and it seemed better to agree the shape on one tool first than to send three at once. Say the word and I will add them.