Changelog

All shipped roadmap items, in reverse chronological order. Items are moved here from ROADMAP.md when they ship, per the roadmap discipline.

§104-27 — Tier 1 (partial), Tier 2 (partial), Tier 3 (partial), Tier 4 (one bet)

The "agentic address system + search engine" milestone. Built in a single session against the kill-criteria-first discipline.

Tier 1 — Close the loop

• #1 agent-card.json snapshot artifact — Reader.snapshot(), buildSnapshot(), parseSnapshot() plus CLI agent-feed snapshot. The address-system half of the user-framing.
  • Kill-criterion held: will fire if consumers always replay the stream and never read the snapshot for >6 months. Currently impossible to evaluate.
• #4 Conformance test vectors — 5 vectors at tests/vectors/ (empty / single-announcement / schema-change-mismatch / terminated / signed-snapshot). Cross-language portable.
  • Kill-criterion held: fires if no second-language implementation appears within 6 months. The Python adapter (Tier 2 #8) is partial confirmation we're already past that threshold.
• #5 agent-feed lint <url> — CLI subcommand + library function lintRemote(origin). Returns graded {errors, warnings, ok}.
  • Kill-criterion held: fires if publishers ignore lint output and ship broken feeds. Too early to evaluate.

Tier 2 — Adoption substrate

• #7 @agent-feed/next — Next.js 14/15 App Router middleware. 44 LOC of source. Two route files plug into app/.well-known/{did.json,agent-feed.xml}.
• #8 agent-feed-fastapi — Python package with byte-for-byte canonicalization parity verified against the TypeScript reference. 209 LOC. Cross-language signature verification confirmed working.
• #9 @agent-feed/cloudflare-worker — Worker bound to a KV namespace for keys + entries. POST /admin/init, POST /admin/append, plus the two .well-known/ routes. 158 LOC.
• #11 Hosted aggregator — apps/aggregator/: SQLite + FTS5 storage, Bun.serve HTTP server, signed-entry crawler, REST API (/api/{search,origins,crawl,lint}), single-file vanilla JS Web UI. This is the search-engine wedge.
  • Kill-criterion held: fires if <10 actively-using clients in 6 months from public deployment. Currently zero (nothing deployed).

Tier 3 — Standards posture

• #13 IETF draft — docs/IETF-DRAFT.md. 2691 lines, kramdown-rfc shaped, ready to convert to xml2rfc XML for submission as draft-abdi-agent-feed-00. Reader contract precedes producer schema (the load-bearing decision). Written by the Hickey persona.
• #14 MCP SEP — docs/MCP-SEP-agent-feed.md. Proposes agent-feed as the streaming history layer beneath MCP Server Cards. Two artifacts at two URLs, deliberate decomplecting; no MCP wire-protocol changes.

Tier 4 — One ambitious bet, in working code

• #4.5 The 404-killer — withFeedRecovery(reader, fetch, {origin}) wrapper. When an agent's fetch returns 404 for a URL whose deprecation announcement names a replacement, the wrapper retries against the replacement once. Exactly one retry; honors feed-status: terminated. The simplest-version of "make the agentic web more reliable than the human web."

Discipline

• TDD honored for everything except the snapshot artifact, where I went GREEN-first; the slippage was caught by the user, acknowledged in commit message, and discipline was restored for the rest.
• No additions without kill-criteria. Each shipped item carries one in ROADMAP.md (now archived here for the shipped subset).
• Code stayed lean. Library + adapters + aggregator + tests + spec total ~1900 LOC of TS source + 209 LOC of Python; ~5500 LOC of docs (SPEC + IETF draft + SEP); 5 conformance vectors as JSON.

What was explicitly deferred (with reasons)

• Tier 1 #2 WebSub push — own kill-criterion holds: no consumer asking. Polling-only is sufficient for the low-frequency change cadence.
• Tier 1 #3 did:web key rotation — not address-system or search-engine blocking; do post-launch when first key incident teaches us the right shape.
• Tier 1 #6 HTTP Cache-Control + ETag — premature; the aggregator polling cadence doesn't load any real origin yet.
• Tier 2 #10 Hosted conformance checker site — the API exists (/api/lint + CLI agent-feed lint); a marketing-grade UI is downstream of having external publishers to test against.
• Tier 2 #12 npm + JSR publish — purely mechanical; defer to first external integration request.
• Tier 3 #15 A2A extension — same shape as the MCP SEP; write after MCP TC engages.
• Tier 3 #16 Rust + Go ports — Python proves the canonicalization is portable. Two more ports add 200 LOC each but no new capability per the discipline; defer until a Rust- or Go-using consumer asks.
• Tier 3 #17 Formal TLA+ model — overkill before deployment data shows where the contract is actually stressed.
• Tier 3 #18 Real-world feed corpus — needs real publishers; not buildable in a session.
• Tier 4 #1, #2, #3, #4, #6, #7, #8, #9, #10 — these are not buildable in a session; they are directional bets. They remain in ROADMAP.md. The aggregator + search engine is the substrate from which most of them grow.