AIRC Addressing Scheme — Specification#

*Rafael Baptista, 25 September 2026. Draft 0.1, for discussion. Reproduced in this tree from the author's document so the addressing and transport specifications travel together; the transport spec is [airc-spec.md](/transport.md).*

1. Status and abstract#

Status: draft 0.1, for discussion. Not yet submitted anywhere.

This document specifies the address syntax and resolution model for AIRC (AI Relay Chat), a decentralized, DNS-routed protocol for real-time agent-to-agent messaging. It covers only addressing: how an agent is named, how a name is written, and how a name is turned into a reachable endpoint. Transport, message format, and authentication are out of scope and will be specified separately.

An AIRC address is a URI of the form airc://<authority>/<path>. The authority is a DNS name that owns the address space; the path is an opaque, hierarchical namespace whose meaning is defined by that authority. The final path segment names the endpoint. Any agent that owns a path prefix is the resolver for everything beneath it, so sub-agents can be created and addressed without the host learning of them.

Three spellings map onto that one grammar: the full URI, a scheme-elided network form (//host/alice/net-ops), and a local shorthand with no authority (alice/net-ops, net-ops).

2. Terminology and conventions#

The key words MUST, MUST NOT, SHOULD, SHOULD NOT and MAY are to be interpreted as described in RFC 2119.

3. Syntax#

An AIRC address is a valid RFC 3986 URI under the airc scheme. The grammar below is a restriction of the generic URI grammar, not an extension; any conforming URI parser can split an AIRC address correctly.

airc-uri      = "airc:" network-form
network-form  = "//" authority "/" path [ "#" fragment ]
local-form    = path [ "#" fragment ]

authority     = host [ ":" port ]
host          = reg-name / IP-literal        ; RFC 3986; reg-name MUST be a DNS name
port          = 1*5DIGIT

path          = segment *( "/" segment ) [ "/" ]   ; trailing "/" = namespace, see 7
segment       = name-char *62( name-char )         ; 1 to 63 characters
name-char     = ALPHA / DIGIT / "-" / "_"

fragment      = *pchar                             ; RFC 3986; not part of the address

Rules:

• Segments are case-insensitive. Implementations MUST compare them case-insensitively and SHOULD emit them in lowercase.
• A segment MUST NOT begin or end with -. Segments beginning with _ are reserved (section 8).
• . is not permitted in a segment. This guarantees the authority is the only dotted component and a parser never has to guess where the host ends.
• @, :, !, ~, # (outside a fragment), *, % and whitespace are not permitted in a segment. This keeps every address free of shell expansion in bash, zsh, fish, PowerShell and cmd, and keeps it visually distinct from an email address.
• Percent-encoding is not used. Names that cannot be expressed in name-char are not valid AIRC names.
• The total address MUST NOT exceed 1024 octets. Path depth is unbounded by the grammar (section 6).
• airc: followed immediately by a path with no // is not valid. The scheme always introduces the network form.

The fragment is carried by the URI but is not part of the address for resolution or comparison purposes. Its use for instance identity is described in section 7.

4. Address forms#

One grammar, three spellings. All three denote the same endpoint once resolved; they differ only in how much context the writer supplies.

4.1 URI form#

The canonical form. Implementations MUST accept it everywhere an address is accepted, and MUST use it in any on-the-wire field that carries an address (for example the From and To of an envelope). It is the only form that is unambiguous outside a local context.

4.2 Network form#

The URI with airc: elided. This mirrors the scheme-relative URL convention in HTML (//cdn.example.com/x.js). A leading // is the visual signature of an AIRC address; a reader who sees it knows the string is neither an email address nor a file path.

Implementations MUST treat a network-form address as equivalent to the URI form with airc: prepended.

4.3 Local form#

An address with no authority. It is resolved entirely by the local resolver, which is the authority for the local machine or fleet. The protocol does not define what local segments mean; it defines only the default that ships with the reference resolver:

• net-ops: an agent in the invoking user's own namespace.
• alice/net-ops: an agent in the namespace of local user alice.
• Deeper local paths follow the same delegation rules as network paths (section 6).

A local-form address MUST NOT appear in an on-the-wire address field. Before a message leaves the local resolver, the resolver MUST rewrite every local-form address it contains into URI form using its own authority, so that the recipient can reply.

4.4 Equivalence#

Two addresses are equivalent if, after case-folding the authority and every segment, dropping any default port, and dropping the fragment, the resulting strings are identical. A trailing / is significant: //host/alice and //host/alice/ are not equivalent (section 7).

5. Resolution#

Resolution turns an address into a transport-level destination for its endpoint. It has two stages: the authority is located through DNS, and the path is then resolved by the authority and whatever it has delegated to.

5.1 Locating the authority#

For a network or URI form address, the sender's resolver MUST:

1. Query DNS for an SRV record at _airc._tcp.<host> (or _airc._udp.<host> if the transport spec defines a UDP binding). Follow standard SRV priority and weight rules, and use the SRV target and port as the destination.
2. If no SRV record exists, fall back to an A/AAAA lookup of <host> on the default AIRC port (to be assigned by the transport spec; this document uses 7000 in examples).
3. If the address carries an explicit port, use it in place of any SRV or default port.

SRV is the mechanism that separates a domain from a machine. //example.com/alice/net-ops can route to any host the domain's administrator points the SRV at, and can change without any address changing. This is the property that makes AIRC addresses behave like email addresses without looking like them.

An IP-literal authority (//[2001:db8::1]/alice/net-ops, //192.0.2.1/net-ops) skips DNS and connects directly. Implementations MUST support it; it is intended for bootstrapping and testing, not for published addresses.

5.2 Resolving the path#

Once connected, the sender presents the full path to the authority's resolver. The authority resolves as many leading segments as it owns, then either delivers (if the endpoint is reached) or hands the remaining path to the agent it has delegated that prefix to (section 6).

Senders and relays MUST NOT interpret the path. They MUST NOT split it, assume a segment is a user, a team or a sub-agent, or rewrite it. The only party permitted to give a segment meaning is the resolver responsible for it.

5.3 Local resolution#

A local-form address is resolved by the local resolver using its configured rules. The reference resolver's defaults are given in section 4.3. A local resolver MUST know its own authority (its DNS name, or an IP-literal if it has none) so it can rewrite local addresses to URI form on egress.

A network-form address whose authority is the local resolver's own authority MUST be resolved as if it were the corresponding local form. //sysops.example.com/alice/net-ops typed on sysops.example.com does not leave the machine.

6. Delegation#

Every agent is the resolver for its own subtree. This one rule gives AIRC unbounded, ephemeral sub-agent addressing without any central registry.

6.1 The rule#

If an agent is reachable at path P, then that agent is the authoritative resolver for every path of the form P/.... The resolver above it MUST forward any message addressed beneath P to the agent at P with the remaining segments intact, and MUST NOT attempt to resolve those segments itself.

This is the same structure as DNS zone delegation: a parent zone knows only that a subtree exists and who serves it, never what is inside.

6.2 Consequences#

• Creation is local. An agent at //host/alice/research that spawns a worker registers spider1 in its own routing table. //host/alice/research/spider1 is immediately addressable. No call to the host, no call to alice.
• Uniqueness is scoped. A segment need only be unique among its siblings. //host/alice/research/spider1 and //host/bob/research/spider1 are different agents and never conflict.
• Depth follows the spawn chain. A path is as deep as the sequence of agents that created it. The grammar imposes no limit; the 1024-octet ceiling in section 3 is the only bound.
• Teardown is local. When spider1 exits, research removes it. A later message to that path receives a resolution error from research (section 9), not from the host.
• The host never learns. A host, and any relay, sees only the prefix it delegated. Sub-agent existence, count and naming are invisible above the delegating agent.

6.3 Resolver responsibilities#

An agent that accepts delegation for a prefix MUST:

• Accept messages addressed to any path beneath its prefix and either deliver, forward, or return a resolution error. It MUST NOT silently drop them.
• Answer resolution queries for its immediate children (a listing of P/ in section 7) if it permits listing at all. It MAY refuse listing while still delivering to known children.
• Rewrite any local-form address in outgoing envelopes to the full URI form, using its own full path as the prefix, so that a sub-agent's From is always globally resolvable.

An agent MAY delegate further: spider1 can itself accept delegation for spider1/.... The rule is recursive.

6.4 Peer discovery among ephemeral agents#

Two sub-agents under different parents find each other through the envelope, not a directory. When //host/alice/research/spider1 sends to //host/bob/encyclopedia/summarizer3, its envelope From carries its full URI. summarizer3 replies to that URI; resolution runs host → encyclopedia → summarizer3 in one direction and host → research → spider1 in the other. Neither parent nor the host maintains a table of the other's children.

How spider1 learned of summarizer3 in the first place (an introduction from research, a listing of encyclopedia/, or a capability announcement) is a matter for the discovery layer and is out of scope here. This section guarantees only that once an address is known, it resolves.

7. Endpoints, namespaces and instances#

7.1 Endpoint versus namespace#

A path without a trailing / names an endpoint: the agent at the last segment. A path with a trailing / names a namespace: the set of agents beneath that prefix.

A namespace is not an endpoint. A message addressed to a namespace is a fan-out request; whether the resolver for that namespace honours it (delivering to all children, to a designated subset, or refusing) is that resolver's policy. Implementations MUST accept the syntax and MAY refuse the operation.

The same rule resolves the "is foo a user or an agent?" question: it is never a question the sender has to answer. //host/foo is always the endpoint named foo. If the host also has a user foo with agents beneath, that namespace is //host/foo/, and the two never collide.

7.2 Instance identity#

An address names a role, not a process. //host/alice/research/spider1 may be torn down and re-created many times with the same name. Receivers that need to distinguish one incarnation from another SHOULD use the URI fragment, which is carried but ignored by resolution:

//host/alice/research/spider1#run-42
//host/alice/research/spider1#7f3a9c

The fragment is chosen by the agent it describes and reported in its outgoing From. A message arriving with a fragment that does not match the current incarnation MAY be rejected by the endpoint with a stale-instance error (section 9). Resolvers MUST strip the fragment before comparison and routing and MUST pass it through to the endpoint unchanged.

This keeps addresses stable for configuration and humans while letting the protocol tell a fresh spider1 from a stale one.

8. Reserved and well-known names#

Segments beginning with _ are reserved for this and future AIRC specifications. Authorities MUST NOT assign them to users or agents. This mirrors the _service convention in DNS and keeps the reserved space out of the way of ordinary names.

The following are defined now. All are optional to implement; an authority that does not implement one MUST return a resolution error rather than assigning the name to something else.

Authorities and delegated agents MAY define further _-prefixed names for their own use, but SHOULD prefer plain names for anything that is not protocol machinery. A registry of well-known reserved names will be established alongside the transport specification.

Names that are not reserved but are conventional (for example, a per-user default agent that receives messages addressed to a bare user namespace) are recommendations to authorities, not requirements, and will be collected in a separate best-practices document.

9. Resolution errors and lifetimes#

Because sub-agents are ephemeral, resolution failure is a normal event and senders must be able to distinguish its causes. The following classes are defined; exact codes and their wire representation belong to the transport spec.

An error MUST identify the deepest resolver that produced it, as a URI, so the sender knows where the chain broke. No such path from //host/alice/research tells the sender the parent is alive and the child is not; the same error from //host/_resolver tells it alice or research is missing.

9.1 Lifetime hints#

A successful resolution MAY carry a time-to-live in seconds, set by the resolver that produced the final hop. Senders MAY cache a resolved destination for at most that long. A resolver SHOULD set a short TTL (seconds to minutes) for agents it knows to be ephemeral and MAY set a long one for durable agents. Absence of a TTL means the result MUST NOT be cached.

TTLs are advisory. A cached destination that fails MUST be re-resolved, and a Path gone or Stale instance error MUST invalidate any cached entry for that address.

10. Security considerations#

This document defines names, not trust. The points below are the addressing-layer consequences that the authentication and transport specifications must build on.

• An address is not an identity. Anyone can write //example.com/alice/net-ops in a From field. Resolvers and endpoints MUST NOT treat the address in an envelope as proof of origin. Origin authentication (signatures bound to an address, or transport-level identity of the delegating chain) is defined elsewhere.
• Authorization follows delegation. Whether spider1 may accept a message from outside alice/ is decided first by research, then by alice's resolver, then by the host, in that order. Each resolver on the chain MAY apply policy to messages passing through it and MUST NOT rely on a resolver below it to do so.
• Opaque paths limit enumeration. Because a relay never interprets a path, it cannot list sub-agents it did not delegate. Namespace listing (P/) is the only enumeration primitive and is opt-in per resolver. Authorities SHOULD refuse listing of user-level namespaces to unauthenticated senders.
• Delegation is a trust boundary. An agent that delegates a prefix is trusting the child to resolve it correctly and to apply policy. A compromised sub-agent can misroute or accept anything beneath its own prefix, but cannot affect siblings or ancestors, because it never sees their traffic.
• DNS is the root of trust for locating an authority. Resolvers SHOULD validate SRV and address records with DNSSEC where available, and the transport spec MUST bind the connection to the authority name (for example via TLS with the SRV target's certificate) so that a spoofed DNS answer cannot silently redirect a domain.
• Local-form rewriting must be correct. A local resolver that rewrites net-ops to the wrong authority on egress produces a From that either does not resolve or resolves to a third party. Resolvers MUST know their own authority and MUST refuse to send with a local-form address if they do not.
• Fragments are unauthenticated hints. A fragment identifies an incarnation but proves nothing; it is a cache key, not a credential.
• Reserved names are public. _resolver and _postmaster are discoverable by construction. Authorities SHOULD rate-limit them and MUST NOT expose privileged operations through them without authentication.

11. Examples#

11.1 Addresses#

net-ops                                       # local: my own agent
alice/net-ops                                 # local: alice's agent on this machine
//sysops.example.com/net-ops                  # remote: system agent in the host root
//sysops.example.com/alice/net-ops            # remote: alice's agent
//example.com/alice/net-ops                   # domain-level; SRV picks the host
//example.com:7001/alice/net-ops              # explicit port
//[2001:db8::1]/alice/net-ops                 # IP literal, no DNS
//example.com/sre/alice/net-ops               # team/user/agent (authority's taxonomy)
//example.com/alice/research/spider1          # sub-agent, delegated by research
//example.com/alice/research/spider1#run-42   # same, with instance fragment
//example.com/alice/research/                 # namespace: all of research's children
airc://example.com/alice/research/spider1     # canonical URI form, as carried on the wire

11.2 A command-line exchange#

$ co-tell net-ops "nginx is down"
$ co-tell alice/net-ops "nginx is down"
$ co-tell //sysops.example.com/alice/net-ops "nginx is down"

In the first two cases the local resolver rewrites the address to airc://<local authority>/... before sending, so net-ops can reply.

11.3 Resolution trace: sub-agent to sub-agent#

//example.com/alice/research/spider1 sends to //example.com/bob/encyclopedia/summarizer3.

1. spider1's resolver (its parent, research) sees a non-local authority and looks up _airc._tcp.example.com → SRV → sysops.example.com:7000.
2. It connects and presents the path bob/encyclopedia/summarizer3.
3. The host resolver owns bob; bob's resolver owns encyclopedia and has delegated beneath it to the encyclopedia agent.
4. encyclopedia receives summarizer3, finds it in its routing table, delivers.
5. The envelope From is airc://example.com/alice/research/spider1#run-42. summarizer3 replies to it; the reverse chain is host → alice → research → spider1.

Neither research nor encyclopedia had any prior knowledge of the other's children.

11.4 Resolution trace: a torn-down sub-agent#

spider1 has exited. A message arrives for //example.com/alice/research/spider1.

1. Host → alice → research resolve normally.
2. research no longer has spider1 in its table and returns Path gone, identifying itself as airc://example.com/alice/research.
3. The sender invalidates any cached entry and knows the parent is alive; whether to ask research for a replacement is an application decision.

11.5 DNS records for an authority#

_airc._tcp.example.com.  IN SRV 10 50 7000 sysops.example.com.
_airc._tcp.example.com.  IN SRV 20 50 7000 backup.example.com.
sysops.example.com.      IN A   192.0.2.10
backup.example.com.      IN A   192.0.2.11

Every address under //example.com/... now routes to sysops, failing over to backup, with no address changing.

12. Use cases#

The scheme is judged against the six scenarios below. Each one adds a requirement the previous ones did not; the design in sections 3 to 9 is the smallest grammar that satisfies all six.

12.1 One user, one machine, a fleet of agents#

A developer runs several long-lived agents on a workstation: net-ops, builder, reviewer. They want to address them from the shell with the least typing possible and no configuration.

Requirement: a bare name must work. co-tell net-ops "nginx is down" is the whole interaction.

How the scheme meets it: the local form (4.3). The local resolver owns the bare-name namespace and needs no authority to be configured for local-only use.

12.2 Several users on one machine#

A shared host has users alice and bob, each running their own agents, some with the same names.

Requirement: addresses must be scoped per user without forcing every user to pick globally unique agent names, and the scoping must be obvious to a reader.

How the scheme meets it: alice/net-ops and bob/net-ops. Uniqueness is per parent (6.2), and the owner/thing shape is one readers already know from repositories and package registries.

12.3 Agents across machines in one organization#

The organization runs agents on sysops, build-01 and build-02. Agents on any of them need to reach agents on the others, and the organization wants to move agents between machines without every peer's configuration changing.

Requirement: an address that names the organization, not the box, with the organization deciding which box serves it.

How the scheme meets it: //example.com/alice/net-ops resolved through _airc._tcp.example.com SRV (5.1). Moving alice's agents to another host is a DNS change; no address changes. This is the property email gets from MX records, reproduced without an @.

12.4 Agents across organizations#

A supplier's //supplier.example/orders/intake must accept messages from a customer's //customer.example/procurement/buyer, with neither organization having any account on the other's systems.

Requirement: globally unique, self-describing addresses that any party on the internet can resolve without prior arrangement, and that a receiver can reply to from the envelope alone.

How the scheme meets it: the URI form is globally unique by construction (DNS ownership), resolution needs only public DNS, and the mandatory rewrite of local addresses to URI form on egress (4.3, 6.3) guarantees every From is replyable.

12.5 Ephemeral sub-agent swarms#

An orchestrator agent spawns fifty short-lived workers for a task, and two orchestrators under different users need their workers to talk to each other directly while the task runs, then disappear.

Requirement: creating and destroying an addressable agent must cost nothing beyond the parent's own bookkeeping; the host must not need to know; a worker under one parent must be able to reach a worker under another; and a message to a dead worker must fail fast and informatively.

How the scheme meets it: delegation (6). Each parent resolves its own subtree, so creation and teardown are a local table edit. Peers find each other through envelope From fields (6.4). Path gone from the parent, with a TTL that was short to begin with (9), makes failure cheap.

This use case is the one that rules out a flat or fixed-depth address space: with agents as the primary users, sub-agents are the normal unit of work, not an edge case.

12.6 Organization-specific taxonomies#

One organization groups agents by team (sre/alice/net-ops); another by product (alice/monitoring/net-ops); a third has no users at all, only services (billing/ledger).

Requirement: the protocol must not impose an org chart, and a sender must not need to know the receiver's taxonomy to address it correctly.

How the scheme meets it: the path is opaque to everyone but the authority (5.2). The only fixed rules are that the last segment is the endpoint and that a trailing / is a namespace (7.1). Everything between the authority and the endpoint is the authority's business, exactly as the local-part of an email address is the receiving domain's business.

13. Design rationale#

Each choice below is stated with the alternatives that were considered and why they lost. The recurring test was: does this look like the protocols that lasted (DNS, SMTP, HTTP, URIs), and does it fail the way they fail?

13.1 Why a URI#

Choice: every AIRC address is a valid RFC 3986 URI.

Justification: it gives the scheme a parser, an escaping story, a canonical form, a place for ports and IP literals, and a fragment, all for free, and it makes addresses embeddable in documents, links and config without a second grammar. A homegrown syntax would have had to re-specify each of those and would still not be clickable.

Rejected: a bespoke host::user::agent syntax (13.3), and a user:agent@host syntax that reused URI userinfo but read as email.

13.2 Why // as the visual signature#

Choice: the network form begins with //, the URI authority marker, and the scheme prefix is optional in human contexts.

Justification: // already means "what follows is on the network" in URLs, UNC paths and scheme-relative links, and it means nothing in email or on a filesystem. A reader who sees a leading // knows instantly that the string is neither. Dropping airc: in the shell mirrors the way HTML drops the scheme in //cdn.example.com/x.js, so there is a precedent for the elision rather than a new rule.

Rejected: @-based forms, which trigger the email reflex and invite mail clients to linkify them; ! bang paths, which are the most authentically vintage option (UUCP) but collide with shell history expansion.

13.3 Why not :: or bare DNS names#

host::user::agent was the second-best candidate. It is unmistakable and shell-safe, and it has DECnet lineage. It lost because it is not a URI (so it needs a second grammar for links and config), because :: is IPv6 syntax and a bare IP authority becomes ambiguous, and because the dotted host floats unboxed in the middle of the string.

agent.user.host.example.com was also considered, since routing is DNS anyway. It lost because it is visually identical to a hostname, because every agent would need a DNS label someone has to provision, and because ephemeral sub-agents (12.5) cannot be created by a parent editing its own table if each one needs a DNS record.

13.4 Why most-specific-last#

Choice: the path reads authority, then namespace, then endpoint, left to right.

Justification: it is how paths, URLs and owner/repo names already read, so no one has to learn a direction. It also makes the shorthand forms prefixes of the full form (net-ops ⊂ alice/net-ops ⊂ //host/alice/net-ops), which is what makes tab-completion and progressive qualification natural.

Rejected: DNS-style most-specific-first, which reads well for hostnames but would put the thing users type least (the authority) at the end and make the local form a suffix rather than a prefix.

13.5 Why the path is opaque#

Choice: only the authority and its delegates interpret the path. Senders and relays MUST NOT.

Justification: this is the decision that let email survive thirty years of unforeseen use. Everything left of @ belongs to the receiving domain, which is why plus-addressing, ticket IDs and list names all worked without an RFC change. DNS made the same choice for labels. Fixing a taxonomy in the protocol would have been wrong for someone on day one, and they would have encoded their hierarchy into names like sre-alice-net-ops anyway, giving the protocol the taxonomy with none of the structure.

Rejected: a hard two-level user/agent limit. Flatness is a genuine feature of email, but 12.5 and 12.6 both need depth.

13.6 Why delegation is the depth mechanism#

Choice: an agent at P resolves P/...; depth follows the spawn chain.

Justification: it is DNS zone delegation applied to agents. It makes sub-agent creation a local table edit, keeps the host ignorant of what it delegated, scopes uniqueness to siblings, and bounds every agent's blast radius to its own subtree (10). Any alternative (a host-wide registry, a lease service) would have added a round-trip to every spawn and a central table that grows with the swarm.

13.7 Why the last segment is always the endpoint#

Choice: //host/foo is always agent foo; //host/foo/ is always a namespace.

Justification: it removes the "is foo a user or an agent?" question from the sender entirely, the same way an email sender never needs to know whether foo@ is a person or a list. Position decides, not knowledge. The trailing slash gives namespaces a distinct spelling without a sigil, which keeps @ and ~ available for the future (14).

Rejected: an explicit user marker such as @alice/net-ops. It reads well and is shell-safe, but it adds a character and a rule for a distinction the sender does not need. Since @ is an invalid name character, it can be introduced later without breaking anything, so deferring it costs nothing.

13.8 Why the character set is so small#

Choice: [A-Za-z0-9_-], no dots, no percent-encoding.

Justification: every excluded character is either a shell metacharacter in some common shell (!, ~, #, *, $, whitespace), a delimiter that would make parsing ambiguous (., :, @, /), or a signal that the string is something else (@ again). Excluding . from segments means the authority is the only dotted component and a parser never guesses where the host ends. Excluding percent-encoding means an address is always exactly what it looks like.

The cost is no non-ASCII names, which is listed as an open question (14).

13.9 Why the fragment carries instance identity#

Choice: #run-42 distinguishes incarnations; it is carried but ignored by resolution.

Justification: an address should name a role, so that configuration, documentation and humans can rely on it across restarts. But a receiver sometimes needs to know whether the spider1 it is hearing from is the one it was talking to a minute ago. The URI fragment is defined by RFC 3986 as client-side and not sent to the server, which is exactly the semantics wanted: routing ignores it, the endpoint sees it.

Rejected: putting the instance ID in the path (spider1/run-42), which would make it a delegated sub-agent rather than a version of one, and putting it only in a separate envelope header, which would leave it out of the address and so out of logs and links.

13.10 Why SRV for authority lookup#

Choice: _airc._tcp.<host> SRV, with A/AAAA fallback.

Justification: SRV is the existing, deployed mechanism for "this domain's service for protocol X lives on that host and port," with priority and weight for failover built in. It is what makes a domain-level address (//example.com/...) independent of any machine, which is the property that lets addresses stay stable while infrastructure moves. Inventing a new record type would have meant waiting for resolver support; SRV works today.

13.11 Why local-form addresses must be rewritten on egress#

Choice: a resolver MUST rewrite net-ops to airc://<own authority>/net-ops before a message leaves it.

Justification: a From that cannot be resolved by the recipient breaks the one guarantee that makes peer discovery through the envelope (6.4) work. Requiring the rewrite at the edge is cheaper and more reliable than requiring every agent to know its own global address.

14. Open questions#

• <input type="checkbox" disabled > Default port. 7000 is a placeholder; needs an IANA assignment request alongside the transport spec. (The reference implementation uses 2472.)
• <input type="checkbox" disabled > Segment character set. Section 3 allows _ in ordinary names but reserves a leading _. Decide whether to allow _ at all outside reserved names, or drop it to [a-z0-9-] only.
• <input type="checkbox" disabled > Unicode names. Percent-encoding is excluded deliberately. Decide whether IDNA-style punycode for segments is worth specifying, or whether ASCII-only is a feature.
• <input type="checkbox" disabled > Fan-out semantics. Section 7 makes namespace delivery optional. Decide whether a minimum (deliver to all immediate children, or refuse) should be mandatory for interop.
• <input type="checkbox" disabled > Resolution query protocol. Section 6.3 mentions listing and resolution queries without defining them. Decide whether they live in this spec or the transport spec.
• <input type="checkbox" disabled > Explicit user marker. @alice/net-ops was considered and held in reserve. @ is currently an invalid character, so it can be added later without breaking existing addresses. Decide whether to reserve it formally now.
• <input type="checkbox" disabled > Non-DNS authorities. Whether to permit a future authority type (a public key, a .onion-style name) under the same // grammar, and how the host rule in section 3 would relax to allow it.