sift/docs/plans/p5a/07-digest-cli-docs.md

P5a task 07: page citations, metrics, CLI flags, skill, unit, README

Branch: p5a (git switch p5a; git status --short must be empty, otherwise stop) Commit subject: P5a: digests cite pages; sift_pdf_* metrics; sift fetch --pages/--grep; skill, unit and README

Goal

Finish P5a:

• digests of PDFs cite (p. N);
• the PDF path is visible in /metrics;
• sift fetch can ask for pages and grep;
• the agents' skill tells them PDFs work;
• the unit gets the memory the child needs;
• the README describes all of it.

Files

• Copy: internal/digest/pages_test.go, internal/metrics/pdf_test.go, cmd/sift/fetchpdf_test.go
• Modify:
  • internal/digest/digest.go (prompt line, PromptVersion)
  • internal/metrics/metrics.go
  • internal/api/pdf.go (the metric calls)
  • cmd/sift/fetchcmd.go
  • skill/SKILL.md
  • nix/module.nix (only MemoryMax)
  • README.md
  • docs/implementer-log.md

Rules

1. Digest.
   • Add one sentence to systemPrompt: When the document has "--- page N ---" markers, cite the page of each fact as (p. N).
   • Set PromptVersion = "3". The digest cache keys on it, so digests made with the old prompt are not served.
   • Split already keeps markers whole; the given test pins that.
2. Metrics (metrics.go):

   • func (s *Set) ObservePDF(outcome string, d time.Duration, pages int). It feeds:

     • the counter sift_pdf_extract_total{outcome};
     • the histogram sift_pdf_extract_seconds;
     • the counter sift_pdf_pages_total.

     outcome must be one of ok, no_text, encrypted, unreadable, timeout, too_large, busy; anything else is counted as other. The label never carries raw text: the given test sends blocked: evil.example.

   • func (s *Set) ObservePDFWait(d time.Duration) feeds the histogram sift_pdf_pool_wait_seconds.

   • func (s *Set) PDFCacheHit() feeds the counter sift_pdf_cache_hits_total.

   • Register them on the Set's own registry, like sift_requests_total.

   • In internal/api/pdf.go:

     • call ObservePDF once per extraction attempt, with the outcome and the pages extracted;
     • call ObservePDFWait with the time spent waiting in PDFPool.Do before the job started (measure around Do: total minus the job's own time);
     • call PDFCacheHit for every answer served from the cache, fresh or after a 304.
3. CLI (fetchcmd.go).
   • Parsing becomes a struct, and today's tuple function stays as a wrapper so the existing tests keep compiling:

     type fetchOpts struct {
     	URL, Mode, JQ, Question, Pages, Grep string
     	NoContext, Recover, JSON               bool
     }
     func parseFetchOpts(args []string) (fetchOpts, error) // the parser; knows --pages and --grep (both take a value)
     func parseFetchArgs(args []string) (…the same 8 results as today…) // calls parseFetchOpts
     func fetchBody(o fetchOpts) map[string]any            // the /v1/fetch request body
     

   • fetchBody builds today's body: url, mode (defaulting to digest when there is a question, else clean), jq, question, context (question set and not --no-context), recover. It adds "pages" and "grep" only when set.
   • The command uses parseFetchOpts and fetchBody.
   • Add [--pages a-b] [--grep RE] to fetchUsage.
   • The given cmd/sift/fetchpdf_test.go and the existing flag tests must pass.
4. Skill (skill/SKILL.md):
   • Remove the 415 unsupported type row's mention of PDF.
   • Add a short "PDFs" section:
     • PDFs work;
     • a long one answers with a map (outline with page numbers, first pages);
     • then use --pages 41-60 or --grep 'pattern', or --question for a digest that cites pages;
     • scanned PDFs return no text layer;
     • never download a PDF and run pdftotext yourself.
5. Unit (nix/module.nix): MemoryMax = "1G"; (was "512M"). Nothing else in the file changes.
6. README:
   • a "PDF documents" section covering: detection, limits (100 MB, 120 s download, 60 s extraction, 2,000 pages), the child process, the pool (2 slots, 4 waiting, 20 s), the cache (24 h, conditional GET), the answers (full / map / pages / grep / digest with citations), the errors table from task 06, and the metrics;
   • note that the engine is go-pdfkit/extract v0.3.0, chosen by the bake-off in docs/P5a.md §3.

Steps

• 1. git switch p5a; copy the three given tests. 2. See them fail.
• 3. digest, 4. metrics and the calls in pdf.go, 5. CLI, 6. skill, unit, README.
• 7. gofmt -w; make gate.
• 8. make build && bin/sift fetch 2>&1 | grep -c -- '--pages' prints 1.
• 9. Row p5a/07-digest-cli-docs; commit.

git add internal/digest internal/metrics internal/api/pdf.go cmd/sift skill nix/module.nix README.md docs/implementer-log.md
git commit -F .state/commit-msg.txt

Done when

• All tests pass; make gate ok; given files byte-identical; the nix/module.nix diff is the one line.