Round 3 of the dogfood confirmed that every multi-branch CLI command
works end-to-end, but the single biggest remaining friction is still
checkpoint payload construction. Each agent hands off a ~15-18 KB
markdown document; turning that into the Smriti schema (decisions,
assumptions, tasks, open_questions, entities, artifacts) is three
minutes of hand-written JSON per checkpoint and adds no product value.
This build adds an LLM-powered extractor that collapses that work into
one pipe command:
cat /tmp/r3_agent_a_output.md | smriti checkpoint create my-project \
--extract --author-agent codex-A
The CLI reads stdin as freeform markdown, calls the new
POST /api/v5/checkpoint/extract endpoint, maps the returned fields
into a commit payload, and writes the checkpoint. --dry-run prints the
extracted payload without committing so users can review first.
--extract and --from-json are mutually exclusive.
Backend architecture mirrors the existing review endpoint: stateless
LLM call (no session or commit ID required), uses the same background
intelligence provider (cfg.background.provider / cfg.background.model)
as draft and review, same JSON-mode prompt shape, same 502-on-parse-
error pattern. The extractor endpoint differs in one small way: it
passes allow_mock=True to get_adapter so unconfigured test envs fall
back to MockAdapter without raising 500. Production envs always have
a real provider configured and never hit this fallback.
The extractor is the first LLM-backed endpoint that gets tested
against a real mock response. To make that work, MockAdapter.send now
detects response_format={"type": "json_object"} in kwargs and returns
a canned JSON blob covering every field any current Smriti endpoint
looks for (title, objective, summary, decisions, assumptions, tasks,
open_questions, entities, artifacts, issues, suggestions). Existing
chat.send text-mode tests are unaffected because they don't pass
response_format. This also unblocks future tests for draft and review.
Manual verification against a real OpenAI provider: piped a realistic
23-line handoff markdown with 4 decisions, 3 assumptions, 3 tasks,
2 open questions, and a python code block. The extractor returned
exactly those items in the right fields (4/3/3/2/1) and produced a
valid checkpoint with all fields populated. Round 4's load-bearing
claim — zero hand-written JSON per checkpoint — is now achievable.
153/153 backend tests pass (149 pre-existing + 4 new extract tests).
