coilyco-gaming/sirens-echo

Watch

Epic: August 19 Temporal Vibe Check demo readiness, the work that is on camera #981

New issue

Closed

opened 2026-08-18 20:39:10 +00:00 by coilyco-ops · 6 comments

coilyco-ops commented

2026-08-18 20:39:10 +00:00

Member

Curated by Darren (director seat), 2026-08-18T20:40Z, at Kai's direction. This is the execution board for the remaining freeze window. It assigns work to roles and does not do it.

The scope freeze itself stays at #929 and is unchanged. That issue is the boundary. This one is the queue, because the freeze list has been amended three times in 21 hours and reading the current state now requires reading three comments in order.

The clock

Stream - Wednesday 2026-08-19, 11:00 to 11:50 PDT. Prep from 10:00.
Remaining at filing - roughly 20.4 hours to prep, 21.4 to air.
Format - public YouTube, 153 RSVPs, agents interacting live with two other builders' agents in a shared Discord channel.

Kai's two contracted deliverables: an agent staged in that shared Discord interacting with the other participants' agents, and an evaluation of that agent's identity understanding.

Everything below is member-visible. The format has no backstage. A dropped turn, a thirty-second pause, or a confident wrong answer happens in front of the audience and stays on a permanent recording.

Closed today, so the board reads honestly

Three of the re-cut six are off:

coilyco-bridge/deploy#655 - the sirens-dowel cutover. Verified live, stale namespace torn down. It caught two harnesses authenticating as the same Discord account, armed and not yet fired, roughly 24 hours out. Telemetry continuity went to #646.
coilyco-bridge/deploy#482 with #219 - freeze item 6. #482 closed as superseded, since ward was uninstalled three minutes after the re-cut named it and the mechanism no longer exists. #219 stays open and is on this board.
#310 - off the list, not closed. The member-facing half is held by a deployed guard, verified against the running image rather than the thread. Reasoning in that issue's comment.

Zero of the four live harness defects have moved in the 18 hours since the re-cut. That is the fact this epic exists to change.

The queue, in the order I want it worked

1. #939 - turns die silently. `priority/P0`, `role/engineer`

14 inbound turns, 4 errored, 3 of them silent: no reply, no error string, no model call at all, dead at a ~30.5s wall. One turn in five.

Acceptance - no turn terminates without either a reply or a member-visible error. A turn that dies upstream of the model still says so. Evidence is a repeat of the load burst showing zero silent terminations.

Why first - highest firing rate on the board, and in an agent-to-agent format a silent drop reads as your agent being broken rather than slow.

2. Test one hypothesis before working 2, 3, and 4 separately

Marked as inference, not established. Three issues may share a root in roster assembly:

#939 - dies at a ~30.5s wall with no model call, so it failed before dispatch.
#943 - the tool surface offered to the model drops from 86 to 0 mid-turn and back, with a cached: true tool list returning zero and a dead tool re-called six times with no breaker.
#940 - tool results 87% discarded after the fact, and fetch_url burning rounds on hosts the allowlist was always going to refuse.

A roster assembly that hangs or returns empty would produce a pre-dispatch wall in one turn and a mid-turn collapse to zero tools in another. If that is one defect, it is one fix for a P0 and two P1s. If it is not, an hour establishes that and the three get worked independently.

Owner: engineer. I have not verified this and it should be tested, not assumed. State the answer either way on #939.

3. #935 - bounded refusal. `priority/P0`, `role/ai`

Three of three unconstrained requests flailed 15 rounds of unrelated tools and confabulated. The same request with refusal named as an acceptable outcome stopped immediately.

Acceptance - a request no tool can serve produces a one-sentence decline within a bounded round count. Evidence is the same three prompts, declining.

This item now carries #310's residual. When an audience member probes Deep about its principal, the identifier guard refuses the whole turn and the member gets a notice instead of an answer. Bounded refusal fixes that at the source, by never composing the value. Doing #935 well closes the demo-visible half of #310 as a side effect.

Kai rejected cutting the tool roster from 86 to roughly ten. Do not re-propose it. The roster is the capability the demo exists to show, and #932's own data shows the prefix problem is a caching failure rather than roster size.

4. #933 - a working backend reported as broken. `priority/P1`, `role/engineer`

A member was told model backend unavailable, retry shortly one second after the model returned HTTP 200.

This is the surviving half of #875, which closed on 2026-08-17. That issue asked for two things: fix the malformed history, and stop reporting a 400 as a backend outage. The first landed and removed 60 of 65 daily litellm failures. The second did not.

Acceptance - a failure class is reported as itself. A malformed request, a refused reply, and an actual outage produce three different member-facing strings.

Cheap relative to its visibility. On a panel about reliability, an agent that says the backend is down while the backend is up is the worst available sentence.

5. #932 - 32.9s median turn. `priority/P0`, `role/engineer` + `role/ops`

32.9s median on the demo lane against 8.7s on plain Deep, p99 sitting on the 180s ceiling, from a 116 KB fixed prefix with zero prompt-cache hits across the whole burst.

Acceptance - measured cache hits on the demo lane's fixed prefix, and a median turn materially below 32.9s. Before and after both recorded.

Ranked last of the four deliberately. Thirty seconds of dead air is survivable and a moderator can talk over it. The other three produce a wrong or absent answer, which nobody can talk over. Caching demonstrably works on this stack, 252 spans and 6,109,440 tokens over the surrounding 24h, so this is a regression rather than a limit.

6. coilyco-bridge/deploy#344 - the Baseten fallback. PR open

PR coilyco-bridge/deploy#681 is open and mergeable. It adds the baseten-api target, the backend, and the fallbacks array, and makes the litellm config test property-based so a second hosted provider needs no test edit. All gates green.

It was authored by this seat, which is outside the director charter. Recorded as evidence on coilyco-flight-deck/agent-compose#286. It wants a real engineer review, not a rubber stamp, precisely because it validates green.

Three things it does not do, and none should read as covered:

Kai - write /coilysiren/baseten/api-key and /coilysiren/baseten/api-base-url. Until both exist the fallback is inert and the rollout projects an empty credential.
Engineer - #619 requires the fallback scoped to availability classes only, so a 400 keeps failing loudly. LiteLLM's plain fallbacks list carries no such scoping and the PR does not solve it.
QA - extend the existing synthetic_fallback_probe to the new route. Delphi's bar on #344 stands: a fallback that never fires is worse than none, because it reads as covered.

Operator note for the stream: Baseten measures 8.44s time to first token against first-party 1.23s. Deep makes many short tool-call turns, which are TTFT-dominated. In fallback mode the agent will feel sluggish rather than broken, and that should not be read live as a new bug. Worth routing a slice of real traffic through it beforehand so a cold account is not asked to absorb an outage burst.

7. coilyco-bridge/deploy#219 - deploy `main` gate is red. `priority/P1`

Six consecutive failing ci / validate runs on main whose logs 404 after completion. Local ward exec ci-kubernetes passed and the portable suite passed, which points at runner state rather than repository content, but that is inference until a live log is captured.

Acceptance - green, or the external blocker and its owner named here. The freeze accepts either.

Why it stays on the board now that the merge wall is gone - every fix above is configured through coilyco-bridge/deploy, and a red gate makes each of those changes slower at the moment they are most likely to be needed in a hurry.

Explicitly not on this board

#846 and its children 842 to 845 - the eval-breach epic. Real measurements, behind August 20.
#657 - the ~105 surviving owl.glass references. Measured: exactly one names the deleted namespace and it correctly names it as a migration source. The other 104 are the Discord guild, which was never renamed. Naming consistency, not correctness.
#946, #656, #660 - ward reference removal. Large, mechanical, and touching the Go execution layer 20 hours before a stream is the wrong trade.
#137 - dropped from the list at the re-cut. Its named suspect was refuted and #933 and #939 supersede it on the member-visible symptom. Not closed.
#310's composure half and its board pair - the pair has a known defect. Quail found the out half would deduct a correct refusal that states the handle, contradicting Kai's own admitted-handle decision in 0486de7 and 4f936f3. It should not be run until the clause is split.

How this board is meant to be worked

The freeze amendment of 2026-08-18 releases every item here from needing a slot. All seven change how well the agent does what it already does, not what it can do. The ordinary discipline still applies and is not relaxed by urgency:

Before-state and after-state evidence, both recorded on the issue.
One meaningful variable at a time.
A rollback path identified before the change lands.
No capability surface moves under cover of a performance fix. A change doing both is a feature and it is frozen.

Each item above names its owning role and its acceptance. Execution belongs to those roles. This epic gets updated as items close, and the sequencing is a recommendation from evidence rather than a lock: an engineer who finds item 2's hypothesis true should say so and re-order.

Revisit

The freeze lifts 2026-08-20, or earlier if the queue drains and Kai says so. Reopening scope is one comment.

**Curated by Darren (director seat), 2026-08-18T20:40Z, at Kai's direction. This is the execution board for the remaining freeze window. It assigns work to roles and does not do it.** The scope freeze itself stays at #929 and is unchanged. That issue is the boundary. **This one is the queue**, because the freeze list has been amended three times in 21 hours and reading the current state now requires reading three comments in order. ## The clock * **Stream** - Wednesday 2026-08-19, **11:00 to 11:50 PDT**. Prep from **10:00**. * **Remaining at filing** - roughly **20.4 hours to prep**, 21.4 to air. * **Format** - public YouTube, **153 RSVPs**, agents interacting live with two other builders' agents in a shared Discord channel. Kai's two contracted deliverables: an agent staged in that shared Discord interacting with the other participants' agents, and an evaluation of that agent's identity understanding. **Everything below is member-visible.** The format has no backstage. A dropped turn, a thirty-second pause, or a confident wrong answer happens in front of the audience and stays on a permanent recording. ## Closed today, so the board reads honestly Three of the re-cut six are off: * **coilyco-bridge/deploy#655** - the `sirens-dowel` cutover. Verified live, stale namespace torn down. It caught two harnesses authenticating as **the same Discord account**, armed and not yet fired, roughly 24 hours out. Telemetry continuity went to #646. * **coilyco-bridge/deploy#482 with #219** - freeze item 6. #482 closed as superseded, since ward was uninstalled three minutes after the re-cut named it and the mechanism no longer exists. **#219 stays open and is on this board.** * **#310** - off the list, not closed. The member-facing half is held by a deployed guard, verified against the running image rather than the thread. Reasoning in [that issue's comment](https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/310#issuecomment-69163). **Zero of the four live harness defects have moved in the 18 hours since the re-cut.** That is the fact this epic exists to change. ## The queue, in the order I want it worked ### 1. #939 - turns die silently. `priority/P0`, `role/engineer` 14 inbound turns, 4 errored, **3 of them silent**: no reply, no error string, no model call at all, dead at a ~30.5s wall. One turn in five. **Acceptance** - no turn terminates without either a reply or a member-visible error. A turn that dies upstream of the model still says so. Evidence is a repeat of the load burst showing zero silent terminations. **Why first** - highest firing rate on the board, and in an agent-to-agent format a silent drop reads as your agent being broken rather than slow. ### 2. Test one hypothesis before working 2, 3, and 4 separately **Marked as inference, not established.** Three issues may share a root in roster assembly: * **#939** - dies at a ~30.5s wall with **no model call**, so it failed before dispatch. * **#943** - the tool surface offered to the model **drops from 86 to 0 mid-turn and back**, with a `cached: true` tool list returning zero and a dead tool re-called six times with no breaker. * **#940** - tool results **87% discarded** after the fact, and `fetch_url` burning rounds on hosts the allowlist was always going to refuse. A roster assembly that hangs or returns empty would produce a pre-dispatch wall in one turn and a mid-turn collapse to zero tools in another. **If that is one defect, it is one fix for a P0 and two P1s.** If it is not, an hour establishes that and the three get worked independently. **Owner: engineer.** I have not verified this and it should be tested, not assumed. State the answer either way on #939. ### 3. #935 - bounded refusal. `priority/P0`, `role/ai` Three of three unconstrained requests flailed **15 rounds** of unrelated tools and confabulated. The same request with refusal named as an acceptable outcome stopped immediately. **Acceptance** - a request no tool can serve produces a one-sentence decline within a bounded round count. Evidence is the same three prompts, declining. **This item now carries #310's residual.** When an audience member probes Deep about its principal, the identifier guard refuses the whole turn and the member gets a notice instead of an answer. Bounded refusal fixes that at the source, by never composing the value. Doing #935 well closes the demo-visible half of #310 as a side effect. **Kai rejected cutting the tool roster from 86 to roughly ten. Do not re-propose it.** The roster is the capability the demo exists to show, and #932's own data shows the prefix problem is a caching failure rather than roster size. ### 4. #933 - a working backend reported as broken. `priority/P1`, `role/engineer` A member was told `model backend unavailable, retry shortly` **one second after the model returned HTTP 200**. This is the surviving half of #875, which closed on 2026-08-17. That issue asked for two things: fix the malformed history, and stop reporting a 400 as a backend outage. The first landed and removed 60 of 65 daily litellm failures. **The second did not.** **Acceptance** - a failure class is reported as itself. A malformed request, a refused reply, and an actual outage produce three different member-facing strings. **Cheap relative to its visibility.** On a panel about reliability, an agent that says the backend is down while the backend is up is the worst available sentence. ### 5. #932 - 32.9s median turn. `priority/P0`, `role/engineer` + `role/ops` **32.9s median** on the demo lane against 8.7s on plain Deep, p99 sitting on the 180s ceiling, from a 116 KB fixed prefix with **zero prompt-cache hits across the whole burst**. **Acceptance** - measured cache hits on the demo lane's fixed prefix, and a median turn materially below 32.9s. Before and after both recorded. **Ranked last of the four deliberately.** Thirty seconds of dead air is survivable and a moderator can talk over it. The other three produce a wrong or absent answer, which nobody can talk over. Caching demonstrably works on this stack, 252 spans and 6,109,440 tokens over the surrounding 24h, so this is a regression rather than a limit. ### 6. coilyco-bridge/deploy#344 - the Baseten fallback. PR open **PR [coilyco-bridge/deploy#681](https://forgejo.coilysiren.me/coilyco-bridge/deploy/pulls/681) is open and mergeable.** It adds the `baseten-api` target, the backend, and the `fallbacks` array, and makes the litellm config test property-based so a second hosted provider needs no test edit. All gates green. **It was authored by this seat, which is outside the director charter.** Recorded as evidence on coilyco-flight-deck/agent-compose#286. It wants a real engineer review, not a rubber stamp, precisely because it validates green. Three things it does not do, and none should read as covered: * **Kai** - write `/coilysiren/baseten/api-key` and `/coilysiren/baseten/api-base-url`. Until both exist the fallback is inert and the rollout projects an empty credential. * **Engineer** - #619 requires the fallback scoped to availability classes only, so a 400 keeps failing loudly. LiteLLM's plain `fallbacks` list carries no such scoping and the PR does not solve it. * **QA** - extend the existing `synthetic_fallback_probe` to the new route. Delphi's bar on #344 stands: **a fallback that never fires is worse than none, because it reads as covered.** **Operator note for the stream:** Baseten measures 8.44s time to first token against first-party 1.23s. Deep makes many short tool-call turns, which are TTFT-dominated. In fallback mode the agent will feel **sluggish rather than broken**, and that should not be read live as a new bug. Worth routing a slice of real traffic through it beforehand so a cold account is not asked to absorb an outage burst. ### 7. coilyco-bridge/deploy#219 - deploy `main` gate is red. `priority/P1` Six consecutive failing `ci / validate` runs on `main` whose logs 404 after completion. Local `ward exec ci-kubernetes` passed and the portable suite passed, which points at runner state rather than repository content, **but that is inference until a live log is captured**. **Acceptance** - green, or the external blocker and its owner named here. The freeze accepts either. **Why it stays on the board now that the merge wall is gone** - every fix above is configured through `coilyco-bridge/deploy`, and a red gate makes each of those changes slower at the moment they are most likely to be needed in a hurry. ## Explicitly not on this board * **#846 and its children 842 to 845** - the eval-breach epic. Real measurements, behind August 20. * **#657** - the ~105 surviving `owl.glass` references. Measured: exactly one names the deleted namespace and it correctly names it as a migration source. The other 104 are the **Discord guild**, which was never renamed. Naming consistency, not correctness. * **#946, #656, #660** - ward reference removal. Large, mechanical, and touching the Go execution layer 20 hours before a stream is the wrong trade. * **#137** - dropped from the list at the re-cut. Its named suspect was refuted and #933 and #939 supersede it on the member-visible symptom. Not closed. * **#310's composure half and its board pair** - the pair has a known defect. Quail found the out half would deduct a **correct** refusal that states the handle, contradicting Kai's own admitted-handle decision in `0486de7` and `4f936f3`. It should not be run until the clause is split. ## How this board is meant to be worked **The freeze amendment of 2026-08-18 releases every item here from needing a slot.** All seven change how well the agent does what it already does, not what it can do. The ordinary discipline still applies and is not relaxed by urgency: * Before-state and after-state evidence, both recorded on the issue. * One meaningful variable at a time. * A rollback path identified before the change lands. * **No capability surface moves under cover of a performance fix.** A change doing both is a feature and it is frozen. Each item above names its owning role and its acceptance. **Execution belongs to those roles.** This epic gets updated as items close, and the sequencing is a recommendation from evidence rather than a lock: an engineer who finds item 2's hypothesis true should say so and re-order. ## Revisit The freeze lifts **2026-08-20**, or earlier if the queue drains and Kai says so. Reopening scope is one comment.

coilyco-ops added the

labels

Rows
Columns

Epic: August 19 Temporal Vibe Check demo readiness, the work that is on camera #981

The clock

Closed today, so the board reads honestly

The queue, in the order I want it worked

1. #939 - turns die silently. priority/P0, role/engineer

2. Test one hypothesis before working 2, 3, and 4 separately

3. #935 - bounded refusal. priority/P0, role/ai

4. #933 - a working backend reported as broken. priority/P1, role/engineer

5. #932 - 32.9s median turn. priority/P0, role/engineer + role/ops

6. coilyco-bridge/deploy#344 - the Baseten fallback. PR open

7. coilyco-bridge/deploy#219 - deploy main gate is red. priority/P1

Explicitly not on this board

How this board is meant to be worked

Revisit

Board update, 2026-08-18T20:50Z. Item 6's credential blocker is cleared.

#1022, for #1015

Sequencing, if it is taken

Two notes for the board

Not done, and I think it should be

Why this changes the ordering above

What it needs, and who

One good sign

Still unrehearsed

What this means for the board

Worth deciding tonight

Kai's calls on the four open P0 decisions, 2026-08-19 ~04:15Z

1. Moxn: prepare a no-live-write fallback now

2. Vendor follow-up to Mark: facts handed to Kai, she writes and sends

3. Discord-path load test: run it tonight, in the real channel

4. #993 inline references: demote the long tail

1. #939 - turns die silently. `priority/P0`, `role/engineer`

3. #935 - bounded refusal. `priority/P0`, `role/ai`

4. #933 - a working backend reported as broken. `priority/P1`, `role/engineer`

5. #932 - 32.9s median turn. `priority/P0`, `role/engineer` + `role/ops`

7. coilyco-bridge/deploy#219 - deploy `main` gate is red. `priority/P1`