docs: changelog and blog post for v0.10.0 - #4245

Changelog entry and release blog post for v0.10.0, prepared ahead of the release — the tag does not exist yet and no release was triggered.

generate-changelog refuses to run without the tag, so v0.10.0 was created locally on origin/main, used to generate the entry, and deleted immediately. It was never pushed, so release.sh can still create it normally.

Contents

• hindsight-docs/src/pages/changelog/index.md — generated 0.10.0 entry: 3 breaking changes, 27 features, 30 improvements, 105 fixes, and one migration (e2f4a6c8b0d1, attachments + document_attachments — both new and empty, so quick on any deployment size).
• hindsight-docs/blog/2026-09-09-version-0-10-0.md — hand-written release notes.
• skills/hindsight-docs/references/changelog/index.md — the skill's changelog mirror.

What the blog post covers

• Images and files as first-class retain content (feat(retain): inline images and files as first-class content #4077), and recall results carrying the attachments behind each fact (feat(api,control-plane): recall results carry the attachments behind each fact #4277)
• Throughput: pure-ASGI middleware (perf(api): replace both BaseHTTPMiddleware with pure ASGI (3.2x on cheap routes) #4235), concurrent embedding batches (perf(embeddings): issue a remote provider's batches concurrently (#4039) #4043), token counting on toktok (perf(tokenizer): replace tiktoken with quicktok and default to o200k_base #3788, perf(tokenizer): move token counting from quicktok to toktok-rs #4022)
• Prompt preview for every operation + retain prompt tester (feat(api,control-plane): a prompt tester for retain, and prompt preview for every operation #4140)
• Fuzzy tag matching (feat(recall): fuzzy tag matching on tag_groups leaves (#4026) #4028) and multi-text entity labels (feat(entity-labels): add an open-vocabulary multi-valued label type (multi-text) #4027)
• Knowledge pages and mental models in document transfers (feat: export knowledge base in document transfers #3855)
• Also: extensions registry + Supabase unbundling (feat(extensions): add hindsight-extensions registry and unbundle Supabase #3988) with its ModuleNotFoundError upgrade step, StaticKeysTenantExtension (feat(extensions): add StaticKeysTenantExtension — env-configured per-user API keys with per-schema isolation #3675), Meta Model API provider (feat(llm): add Meta Model API as a first-class provider #4081), worker bank rotation (fix(worker): rotate slots across banks so bulk ingest stops starving them (#3861) #3980), mental-model list defaulting to metadata (feat(mental-models): default list to metadata; MCP list returns metadata only #4225), reflect grounding and delta-refresh fixes (fix(reflect): grounding defects in reflect and knowledge-page delta refresh, plus system-evals #4304)
• Breaking: bank profile and background endpoints answer 410 Gone (feat(api)!: retire the bank profile and background endpoints #4127); curl is gone from the Docker images (fix(docker): stop installing curl into the cp-only image #4199, fix(docker): drop curl from the API runtime images #4202), which breaks any user-defined healthcheck that execs curl inside the container

The per-bank text-search switch (#3819) and webhook signatures (#3986) are in the changelog but not the blog post.

Manual edits to the generated entry

• Removed the lines for free-threaded CPython 3.14 and --event-loops (feat: run Hindsight on free-threaded CPython 3.14 (-py3.14t image) #4037, feat(api): serve from several event loops in one process #4067, and their follow-ups) — reverted by revert: drop free-threaded CPython 3.14 support (#4037, #4067) #4234 before release.
• Added two bug-fix entries for fix(reflect): grounding defects in reflect and knowledge-page delta refresh, plus system-evals #4304, which the generator dropped: its subject reads "grounding defects … plus system-evals", and both user-facing fixes are only in the body.

Second commit: a generator fix

The first pass produced 37 entries for 221 core commits, against 65 for the 145-commit release before it — whole user-facing fixes went unlisted. Three causes, measured separately on that range:

• max_completion_tokens was a shared budget on a reasoning model — reasoning bills against it, so a long commit list spends most of it before the first entry.
• The prompt rules read as licence to merge unrelated fixes and to drop anything whose commit title sounds internal. Largest single effect.
• One LLM call for the whole release. Commits are now batched by token count and summarized concurrently, then deduped. Holding model and prompt fixed, one call gives 120 entries against 152 for five batches: 40 commits get an entry only when batched, 8 only when not.

The input was never the constraint. A 245-commit release is ~8500 tokens of commit subjects in total, so every arrangement fits the context window many times over. The batch budget is 2000 tokens; what it bounds is how many commits the model weighs at once before it starts summarizing them away.

The dedup result is re-ordered by the release's own commit order and filtered to commit_ids that were in the input, so a dropped or invented id can't reorder the changelog. Default model is now gpt-5.6-terra. tests/test_generate_changelog_batching.py covers the batching arithmetic and the dedup guarantees with the LLM stubbed.

Latest run, against main at 630c3a6: 240 core commits, 173 entries in 5 batches, 4 deduped.

Before merging

• The changelog is generated against the tag, so it needs a regenerate at release time for whatever lands between now and then — treat this entry as a draft of the shape, not the final list. Re-apply the two manual edits above after regenerating.
• No hero image. Recent release posts carry /img/blog/version-X-Y-Z-release.png; this one still needs one.
• openapi.json in the skill still reads 0.9.2 — correct today; it needs a re-sync once release.sh bumps the version.