[bernalalexis-try]

Fixes #1743.

The hook reference in docs/docs/concepts/plugins.mdx had drifted from the HookType enum, which defines 27 hooks in 8 groups.

• Add reference sections and policy table rows for the 9 missing hooks: generation_batch_pre_call, generation_batch_post_call, generation_batch_error, generation_event, adapter_function_invocation_complete, adapter_function_phase_complete, streaming_start, streaming_event, streaming_end.
• Add the Adapter function lifecycle and Streaming pipeline categories to the overview and the reference.
• Fix the counts ("eight hook categories", "all 27 hook types").
• Add test/plugins/test_plugin_docs.py, which fails when a HookType member has no #### section or policy table row, or when the count text goes stale.

Writable fields in the policy table were taken from mellea/plugins/policies.py; the new hooks other than generation_batch_pre_call are observe-only.

Checked that every HookType value has a section and a table row. I haven't run the new pytest file locally.

[ajbozarth]

I am still on the fence about codifying this, but overall this is a solid PR

Some feedback from Claude. Verified the inventory against HookType, the payload dataclasses, and policies.py — counts, sections, table rows, and writable fields all check out, and the guard test passes.

One optional note: the sampling_loop_start "Fires:" line (existing, not touched by this PR) still reads as firing for any strategy, but stream() runs never fire sampling_loop_* — they use streaming_* + RetryEvent. Now that the Streaming pipeline section sits right below it, a one-line caveat there would close the reader-confusion that motivated #1743. Optional, since every explicit ask in the issue is satisfied.

[bernalalexis-try]

Thanks for the review. Pushed 089da12, which merges current main and:

• Fixes the red code-checks / quality runs: ruff format wanted PLUGINS_DOC in test/plugins/test_plugin_docs.py on a single line. ruff check and ruff format --check are clean now, and the new test file passes locally (55 passed).
• Adds the caveat you suggested under sampling_loop_start: stream runs do not fire the sampling_loop_* hooks and use the streaming pipeline hooks and RetryEvent instead, linked to the streaming section.

I could not run markdownlint or the full mypy pass locally, so I am relying on CI for those.

[AngeloDanducci]

Thanks for the update, started CI, one small nit about resolve.

[AngeloDanducci]

Suggested change

	PLUGINS_DOC = Path(__file__).parents[2] / "docs" / "docs" / "concepts" / "plugins.mdx"
	PLUGINS_DOC = Path(__file__).resolve().parents[2] / "docs" / "docs" / "concepts" / "plugins.mdx"

Probably okay without this but the other path walks we have use resolve.

[ajbozarth]

nit: perhaps we should make this count generic so we don't need to update it in the future (less thrash)

Suggested change

	Plugins are organized into eight hook categories:
	Plugins are organized into different hook categories:

[ajbozarth]

@bernalalexis-try soft ping in case you missed our above feedback notifications :)

[bernalalexis-try]

Thanks both. Pushed fdd4bb5 with the two nits:

• test/plugins/test_plugin_docs.py: PLUGINS_DOC now goes through .resolve() before walking up the parents.
• docs/docs/concepts/plugins.mdx: the intro now says "different hook categories" instead of a number, so it won't need updating when a category is added.

test_plugin_docs.py passes (55) and ruff check / ruff format --check are clean.

[ajbozarth]

This LGTM, pending @AngeloDanducci approval we can merge this. If you want to fix the DCO error you can follow Claude advice pasted below, otherwise I can bypass it since its correct for at least one commit (the actual requirement)

---

Thanks, both nits and the sampling_loop_start caveat look good. One thing left: DCO is failing because fdd4bb51 is missing its Signed-off-by line (it also looks like it was committed with a different email than the earlier commits). Running git commit --amend --signoff --no-edit followed by git push --force-with-lease should fix it.