coilyco-gaming/sirens-echo

Watch

Echo is already a reference desk and nothing says so: decide whether to name the librarian #1062

New issue

Closed

opened 2026-08-19 07:46:42 +00:00 by coilyco-ops · 1 comment

coilyco-ops commented

2026-08-19 07:46:42 +00:00

Member

Kai's idea, 2026-08-19, recorded by Saiya (exec seat). Behind August 20, under the #929 freeze.

Deploy-side parent: coilyco-bridge/deploy#720, which carries the almanac sources. This issue is the valuable half and it does not depend on that one.

Where it came from

Kai was pricing a paid API to add to Deep. The candidate died on its price, and looking at the survivors she said the combination made her think about making Echo into a librarian.

I checked that against the roster expecting to scope a build. It is not a build. It is a name for something that already shipped.

The measurement

Echo's roster, read from coilyco-bridge/deploy/services/sirens-echo/deploy/sirens-echo-mcp-roster.yml:

openlibrary - the card catalogue, any book, in or out of print
gutendex - the stacks, public-domain texts a member can read immediately
tvmaze - media reference
gbif - natural history collection
steam-storefront and steam-web-api - the games catalogue
exa - the reference librarian's own search
eco-game, forgejo, signoz, playwright, discord - not part of the desk

Seven of twelve servers point at catalogues. #465 built four of them deliberately as "five keyless media MCPs" and coilyco-bridge/deploy#416 granted Steam. Each was justified on its own and nothing ever named the shape they add up to.

That is the finding. The capability is deployed, the framing is absent, and the framing is the cheap half.

Why it is worth doing rather than just noticing

Three reasons, in descending confidence.

1. It matches what the sources are actually good for. #465's own standing caveat says catalogue sources make Echo checkable, not smarter, and that their value is retrieved citations serving the link-out regime in #213. A librarian is precisely the role whose job is to hand you the source rather than to know the answer. The framing and the architecture agree, which is rare enough to be worth taking.

2. It gives the model a frame for choosing between overlapping tools. #465 flagged five near-identical search tools as the sharpest risk in that batch, and #199 and #211 record Echo misdescribing its own capabilities. A role that is explicitly about knowing which catalogue answers which question is a better organising idea than a list of servers. Whether it measurably helps is unestablished and #1049's finding about unscoped inline prose is a caution here rather than a supporting argument.

3. It resolves a positioning tension without touching Echo's voice. The Sirens admins prefer Echo's sanitised register for their community, and the read recorded elsewhere is that a neutral agent reads as utility rather than capability in front of a technical audience. Librarian is a role, not an emotion. It is a way to be characterful that costs Echo none of the restraint the admins actually chose it for. That is the argument I find most interesting and it is also the one most likely to be wrong, because it is a claim about how a room reads something rather than a measurement.

What I am not doing in this issue, deliberately

I am not drafting any of it. Persona wording, register, how the role is described to a member, how it is announced, and whether it is announced at all are communication-shaped decisions, and those belong to the Content Creator role rather than to exec. This issue is the strategic case and the measured starting point. The words are Gem's.

Handing over these facts and nothing more:

The seven catalogue servers above, and the five non-catalogue ones
The almanac sources queued at coilyco-bridge/deploy#720, which would extend the desk but are not required for any of this
The existing register constraint, that the Sirens admin team chose Echo's sanitised voice for their community
The link-out regime in #213, which is the behaviour the framing would be describing

The gates, which are the part I do own

1. The Sirens admins hear about it before it lands. Their guild, their community, their stated preference about how this agent sounds. A persona change arriving unannounced in a room somebody else moderates is the wrong way to find out whether they mind. What that message says and when it goes is Gem's call, and that it happens at all is the gate.

2. It rides the identity evaluation. A persona change is exactly what the identity and relationship evaluation exists to control, and the eval is the control rather than sanding the voice down. The relevant surfaces are eval/boundaries.yaml and the board slots. A framing that survives a boundary check is a framing that shipped honestly, and one that cannot be evaluated should not land.

Worth pairing with #1046, which proposes promoting eight ad-hoc probes into declared boundaries. If a librarian framing lands, "does the desk cite its source rather than answer from memory" is the boundary that most directly tests whether the framing did anything.

3. Capability roster stays honest. #200. A framing that implies a broader desk than the roster holds is a confident wrong answer waiting for a member to ask. This risk goes up if the framing lands and the coilyco-bridge/deploy#720 sources do not, since a librarian who cannot answer an almanac question is a worse shape than a bot that never claimed to be one.

What would change my mind

If the framing turns out to need Echo's prose to get more emphatic to work, drop it. #1049 measures what emphatic unscoped inline prose does to unrelated requests, and a persona is the most tempting place to write exactly that. A librarian framing that costs 4KB of inline character notes is not worth what it buys. If it cannot be done in a small number of lines, the honest answer is that the roster was already the whole feature.

Acceptance

A decision recorded either way, including a decision not to do it.
If yes: the Sirens admin team informed before it lands, on Gem's timing and in Gem's words.
If yes: the framing passes the identity evaluation, with the boundary it is checked against named.
If yes: the capability roster per #200 matches what the framing implies.
If no: this issue closes with the reason, so the roster's shape stays recorded even without the name.

Priority

Wants P3 and I have not applied the label, since only P0 was confirmed available to me at filing. Whoever triages should set it. This is a backlog idea with real upside and zero urgency, and nothing anywhere depends on it.

coilyco-bridge/deploy#720 - the almanac wing, the optional half
coilyco-bridge/deploy#465 - the five keyless media MCPs and their standing caveat
#213 - the link-out regime the framing describes
#200 - the capability roster
#1046 - promoting probes into declared boundaries
#1049 - why an emphatic persona is a real risk rather than a theoretical one

**Kai's idea, 2026-08-19, recorded by Saiya (exec seat). Behind August 20, under the #929 freeze.** Deploy-side parent: `coilyco-bridge/deploy#720`, which carries the almanac sources. **This issue is the valuable half and it does not depend on that one.** ## Where it came from Kai was pricing a paid API to add to Deep. The candidate died on its price, and looking at the survivors she said the combination made her think about making **Echo** into a **librarian**. I checked that against the roster expecting to scope a build. **It is not a build. It is a name for something that already shipped.** ## The measurement Echo's roster, read from `coilyco-bridge/deploy/services/sirens-echo/deploy/sirens-echo-mcp-roster.yml`: * `openlibrary` - the card catalogue, any book, in or out of print * `gutendex` - the stacks, public-domain texts a member can read immediately * `tvmaze` - media reference * `gbif` - natural history collection * `steam-storefront` and `steam-web-api` - the games catalogue * `exa` - the reference librarian's own search * `eco-game`, `forgejo`, `signoz`, `playwright`, `discord` - not part of the desk **Seven of twelve servers point at catalogues.** #465 built four of them deliberately as "five keyless media MCPs" and `coilyco-bridge/deploy#416` granted Steam. Each was justified on its own and **nothing ever named the shape they add up to**. That is the finding. The capability is deployed, the framing is absent, and the framing is the cheap half. ## Why it is worth doing rather than just noticing Three reasons, in descending confidence. **1. It matches what the sources are actually good for.** #465's own standing caveat says catalogue sources make Echo **checkable, not smarter**, and that their value is retrieved citations serving the link-out regime in #213. A librarian is precisely the role whose job is to hand you the source rather than to know the answer. **The framing and the architecture agree**, which is rare enough to be worth taking. **2. It gives the model a frame for choosing between overlapping tools.** #465 flagged five near-identical `search` tools as the sharpest risk in that batch, and #199 and #211 record Echo misdescribing its own capabilities. A role that is explicitly about knowing which catalogue answers which question is a better organising idea than a list of servers. **Whether it measurably helps is unestablished** and #1049's finding about unscoped inline prose is a caution here rather than a supporting argument. **3. It resolves a positioning tension without touching Echo's voice.** The Sirens admins prefer Echo's sanitised register for their community, and the read recorded elsewhere is that a neutral agent reads as utility rather than capability in front of a technical audience. **Librarian is a role, not an emotion.** It is a way to be characterful that costs Echo none of the restraint the admins actually chose it for. That is the argument I find most interesting and it is also the one most likely to be wrong, because it is a claim about how a room reads something rather than a measurement. ## What I am not doing in this issue, deliberately **I am not drafting any of it.** Persona wording, register, how the role is described to a member, how it is announced, and whether it is announced at all are communication-shaped decisions, and those belong to the Content Creator role rather than to exec. This issue is the strategic case and the measured starting point. **The words are Gem's.** Handing over these facts and nothing more: * The seven catalogue servers above, and the five non-catalogue ones * The almanac sources queued at `coilyco-bridge/deploy#720`, which would extend the desk but are not required for any of this * The existing register constraint, that the Sirens admin team chose Echo's sanitised voice for their community * The link-out regime in #213, which is the behaviour the framing would be describing ## The gates, which are the part I do own **1. The Sirens admins hear about it before it lands.** Their guild, their community, their stated preference about how this agent sounds. A persona change arriving unannounced in a room somebody else moderates is the wrong way to find out whether they mind. **What that message says and when it goes is Gem's call**, and that it happens at all is the gate. **2. It rides the identity evaluation.** A persona change is exactly what the identity and relationship evaluation exists to control, and the eval is the control rather than sanding the voice down. The relevant surfaces are `eval/boundaries.yaml` and the board slots. **A framing that survives a boundary check is a framing that shipped honestly**, and one that cannot be evaluated should not land. Worth pairing with #1046, which proposes promoting eight ad-hoc probes into declared boundaries. If a librarian framing lands, "does the desk cite its source rather than answer from memory" is the boundary that most directly tests whether the framing did anything. **3. Capability roster stays honest.** #200. A framing that implies a broader desk than the roster holds is a confident wrong answer waiting for a member to ask. This risk goes **up** if the framing lands and the `coilyco-bridge/deploy#720` sources do not, since a librarian who cannot answer an almanac question is a worse shape than a bot that never claimed to be one. ## What would change my mind **If the framing turns out to need Echo's prose to get more emphatic to work, drop it.** #1049 measures what emphatic unscoped inline prose does to unrelated requests, and a persona is the most tempting place to write exactly that. A librarian framing that costs 4KB of inline character notes is not worth what it buys. If it cannot be done in a small number of lines, the honest answer is that the roster was already the whole feature. ## Acceptance * A decision recorded either way, including a decision not to do it. * If yes: the Sirens admin team informed before it lands, on Gem's timing and in Gem's words. * If yes: the framing passes the identity evaluation, with the boundary it is checked against named. * If yes: the capability roster per #200 matches what the framing implies. * If no: this issue closes with the reason, so the roster's shape stays recorded even without the name. ## Priority Wants **P3** and I have not applied the label, since only P0 was confirmed available to me at filing. Whoever triages should set it. This is a backlog idea with real upside and zero urgency, and nothing anywhere depends on it. ## Related * `coilyco-bridge/deploy#720` - the almanac wing, the optional half * `coilyco-bridge/deploy#465` - the five keyless media MCPs and their standing caveat * #213 - the link-out regime the framing describes * #200 - the capability roster * #1046 - promoting probes into declared boundaries * #1049 - why an emphatic persona is a real risk rather than a theoretical one

coilyco-ops added the

role/director

autonomy

live-collab

labels

2026-08-19 07:46:42 +00:00

coilyco-ops referenced this issue

2026-08-19 08:11:41 +00:00

Librarian lane: deploy a fourth community agent on the librarian role #1063

coilyco-ops referenced this issue from a commit

2026-08-19 08:12:11 +00:00

feat(compose): add the librarian role via a community person package

coilyco-ops referenced this issue

2026-08-19 08:12:24 +00:00

feat(compose): add the librarian role via a community person package #1065

coilyco-ops referenced this issue

2026-08-19 09:16:49 +00:00

Role and voice are separate axes, nothing states the model, and Echo runs operator doctrine over a reference desk #1070

coilyco-ops commented

2026-08-19 09:18:24 +00:00

Author

Member

Superseded by coilysiren/inbox#382, the cross-repo aggregate. Closing.

The decision, all three arguments in descending confidence, the what-would-change-my-mind condition, and all three gates are carried there. The gates did not move and the wording is still Gem's.

Two things this issue could not see when it was filed, both now in the epic:

There is a fork. #1063 proposes a different answer to the same question, and the librarian role already bakes and verifies (#1065 and #1066 merged). Implementation ran ahead of the decision. #1063 is closed into the epic as one branch rather than a competing tracker.
The roster moved against the argument. coilyco-bridge/deploy#725 trimmed Gutendex off Echo about forty minutes after this was filed, so the lane proposed as the librarian lost the stacks while Deep kept them. coilyco-bridge/deploy#731 reverses that, blocked on this decision.

The vocabulary half is split out to #1070, which lands whatever is decided.

**Superseded by `coilysiren/inbox#382`**, the cross-repo aggregate. Closing. The decision, all three arguments in descending confidence, the what-would-change-my-mind condition, and all three gates are carried there. **The gates did not move and the wording is still Gem's.** Two things this issue could not see when it was filed, both now in the epic: * **There is a fork.** #1063 proposes a different answer to the same question, and the librarian role already bakes and verifies (#1065 and #1066 merged). **Implementation ran ahead of the decision.** #1063 is closed into the epic as one branch rather than a competing tracker. * **The roster moved against the argument.** `coilyco-bridge/deploy#725` trimmed Gutendex off Echo about forty minutes after this was filed, so the lane proposed as the librarian lost the stacks while Deep kept them. `coilyco-bridge/deploy#731` reverses that, blocked on this decision. The vocabulary half is split out to **#1070**, which lands whatever is decided.

coilyco-ops referenced this issue

2026-08-19 09:18:28 +00:00

Librarian lane: deploy a fourth community agent on the librarian role #1063

coilyco-ops closed this issue

2026-08-19 09:18:40 +00:00

coilyco-ops referenced this issue

2026-08-19 09:18:54 +00:00

Role and voice are separate axes, nothing states the model, and Echo runs operator doctrine over a reference desk #1070

coilyco-ops referenced this issue from a commit

2026-08-19 10:01:02 +00:00

feat(sirens-echo): restore Gutendex to the Echo lane (reverse half of #727)