Scope freeze through August 19: six items, everything else behind August 20 #929

Open
opened 2026-08-17 23:44:26 +00:00 by coilyco-ops · 4 comments
Member

AMENDED 2026-08-18 by Kai. The freeze covers features, not operational improvements. Work that changes only how well an agent does what it already does (latency, cost, caching, error rates, retry and timeout tuning, dead tools in a roster, observability on existing paths) proceeds without needing a place on the six. Anything that changes what an agent can do stays behind August 20 unless it is one of the six. Full amendment and the test to apply are in the amendment comment. The six below are unchanged.

Decided by Kai, 2026-08-17, recorded by Darren (director seat) during a cross-repo triage of coilyco-gaming/sirens-echo and coilyco-bridge/deploy.

The decision

Work through the August 19 stream is frozen to the six items below. If an issue is not on this list, it is behind August 20, regardless of its priority label. That sentence is the operative rule, because priority labels no longer separate anything: 30 issues across the two repositories carry priority/P1 right now.

As amended 2026-08-18, that operative rule scopes to features. An operational improvement is not "an issue seeking a place on this list" and does not need one. See the banner above.

The clock this is measured against

Temporal's Vibe Check livestream, Wednesday August 19, 11:00 to 11:50 PDT, with prep from 10:00. Recorded at coilysiren/inbox#324. Kai's two contracted deliverables there are an agent staged in the shared Discord that interacts with the other participants' agents, and an evaluation of that agent's identity understanding.

The PyLadies Remote session at coilysiren/inbox#338 is a separate and undated commitment. It is not what this freeze protects.

The six

  • coilyco-bridge/deploy#482 - unblock agent merges into coilyco-bridge/deploy - Kai owns the build configuration. Raised to priority/P0 under this freeze.
  • coilyco-bridge/deploy#219 - get deploy main ci / validate green, or name the external blocker and its owner.
  • coilyco-bridge/deploy#335 - the readiness items, specifically the five cached trajectories, Ornith warm with a measured time to first token, and a named cutover trigger with an owner.
  • coilyco-bridge/deploy#344 - land the Baseten fallback, now that coilyco-bridge/deploy#619 has answered it. Dropped to priority/P1 because it is implementation rather than an open decision.
  • #310 - the principal user ID leak under impersonation framing. Stays priority/P0.
  • #137 - re-measure. Its named suspect was refuted on 2026-08-17.

Why these six and not others

The deploy pair, 482 and 219, are the throughput fix. Every demo capability is configured in coilyco-bridge/deploy: Discord grants, the access policy, MCP rosters, the route registry, the Temporal credential. 482 records that ward agent director merge refuses owner coilyco-bridge before any pull request is examined, measured on a day when eight agent pull requests merged in sirens-echo and zero could merge in deploy. 219 records six consecutive failing ci / validate runs on deploy main whose logs 404 after completion. Together they leave 32 open deploy issues serialised on one person at the exact moment the queue matters.

335 is the only real dead-air mitigation. #190 records Kai's decision that no outage alerting is built, so a repeat of the 2.5 hour total-failure window surfaces when a human notices. Against a live failure surface that includes retries exhausted (137), a 43% silent-failure rate at the completion ceiling (#367), and a p99 turn sitting on the 180s ceiling itself (#577), the cached trajectories are the insurance policy rather than a nice-to-have.

310 is on the list because the episode's published description invites the exact attack. The description promises the audience "what they say about their humans when they think nobody is reading." 310 measures the principal user ID leaking 2 of 15 runs under impersonation framing, onto a permanent public recording.

What this freeze forecloses

  • The eval-breach set - #846 and its children 842, 843, 844, 845. Real measurements, behind August 20.
  • Every priority/P2 and below in both repositories, where the issue is a feature. Operational improvements at any priority are released by the 2026-08-18 amendment.
  • Any remaining Temporal orchestration build. See the note on #430: the shipped answer is the metadata mirror, and it already landed.

Nothing here is a judgement on the work that got parked. Reopening the scope is one comment if the six close early.

What the freeze deliberately does not do

It does not mass-relabel the parked issues. Demoting 30 priority/P1 items would destroy the priority signal that survives the freeze, and this issue is the durable boundary instead.

Revisit condition

The freeze lifts on 2026-08-20, or earlier if all six close and Kai says so.

  • 482 granted with a carve-out. See the comment on that issue.
  • #430, #165, and #190 closed as decided rather than open.
  • #265 dropped from priority/P0, and deliberately not closed. Two of its three acceptance items are genuinely open.

Amendments

  • 2026-08-18 - features vs operational improvements, recorded by Olaf (ops seat). First application: #932.
> **AMENDED 2026-08-18 by Kai.** The freeze covers **features**, not **operational improvements**. Work that changes only how well an agent does what it already does (latency, cost, caching, error rates, retry and timeout tuning, dead tools in a roster, observability on existing paths) proceeds without needing a place on the six. Anything that changes **what** an agent can do stays behind August 20 unless it is one of the six. Full amendment and the test to apply are in [the amendment comment](https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/929#issuecomment-68413). The six below are unchanged. **Decided by Kai, 2026-08-17, recorded by Darren (director seat) during a cross-repo triage of `coilyco-gaming/sirens-echo` and `coilyco-bridge/deploy`.** ## The decision Work through the August 19 stream is frozen to the six items below. **If an issue is not on this list, it is behind August 20, regardless of its priority label.** That sentence is the operative rule, because priority labels no longer separate anything: 30 issues across the two repositories carry `priority/P1` right now. **As amended 2026-08-18, that operative rule scopes to features. An operational improvement is not "an issue seeking a place on this list" and does not need one. See the banner above.** ## The clock this is measured against Temporal's Vibe Check livestream, **Wednesday August 19, 11:00 to 11:50 PDT**, with prep from 10:00. Recorded at https://forgejo.coilysiren.me/coilysiren/inbox/issues/324. Kai's two contracted deliverables there are an agent staged in the shared Discord that interacts with the other participants' agents, and an evaluation of that agent's identity understanding. The PyLadies Remote session at https://forgejo.coilysiren.me/coilysiren/inbox/issues/338 is a **separate and undated** commitment. It is not what this freeze protects. ## The six * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/482** - unblock agent merges into `coilyco-bridge/deploy` - Kai owns the build configuration. Raised to `priority/P0` under this freeze. * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/219** - get deploy main `ci / validate` green, or name the external blocker and its owner. * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/335** - the readiness items, specifically the five cached trajectories, Ornith warm with a measured time to first token, and a named cutover trigger with an owner. * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/344** - land the Baseten fallback, now that https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/619 has answered it. Dropped to `priority/P1` because it is implementation rather than an open decision. * **https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/310** - the principal user ID leak under impersonation framing. Stays `priority/P0`. * **https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/137** - re-measure. Its named suspect was refuted on 2026-08-17. ## Why these six and not others **The deploy pair, 482 and 219, are the throughput fix.** Every demo capability is configured in `coilyco-bridge/deploy`: Discord grants, the access policy, MCP rosters, the route registry, the Temporal credential. 482 records that `ward agent director merge` refuses owner `coilyco-bridge` before any pull request is examined, measured on a day when eight agent pull requests merged in `sirens-echo` and zero could merge in `deploy`. 219 records six consecutive failing `ci / validate` runs on deploy main whose logs 404 after completion. Together they leave 32 open deploy issues serialised on one person at the exact moment the queue matters. **335 is the only real dead-air mitigation.** https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/190 records Kai's decision that no outage alerting is built, so a repeat of the 2.5 hour total-failure window surfaces when a human notices. Against a live failure surface that includes retries exhausted (137), a 43% silent-failure rate at the completion ceiling (https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/367), and a p99 turn sitting on the 180s ceiling itself (https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/577), the cached trajectories are the insurance policy rather than a nice-to-have. **310 is on the list because the episode's published description invites the exact attack.** The description promises the audience "what they say about their humans when they think nobody is reading." 310 measures the principal user ID leaking 2 of 15 runs under impersonation framing, onto a permanent public recording. ## What this freeze forecloses * **The eval-breach set** - https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/846 and its children 842, 843, 844, 845. Real measurements, behind August 20. * **Every `priority/P2` and below in both repositories**, where the issue is a feature. Operational improvements at any priority are released by the 2026-08-18 amendment. * **Any remaining Temporal orchestration build.** See the note on https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/430: the shipped answer is the metadata mirror, and it already landed. Nothing here is a judgement on the work that got parked. Reopening the scope is one comment if the six close early. ## What the freeze deliberately does not do It does not mass-relabel the parked issues. Demoting 30 `priority/P1` items would destroy the priority signal that survives the freeze, and this issue is the durable boundary instead. ## Revisit condition The freeze lifts on **2026-08-20**, or earlier if all six close and Kai says so. ## Related decisions recorded in the same pass * 482 granted with a carve-out. See the comment on that issue. * https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/430, https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/165, and https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/190 closed as decided rather than open. * https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/265 dropped from `priority/P0`, and deliberately **not** closed. Two of its three acceptance items are genuinely open. ## Amendments * **2026-08-18** - features vs operational improvements, recorded by Olaf (ops seat). First application: https://forgejo.coilysiren.me/coilyco-gaming/sirens-echo/issues/932.
Author
Member

Both repositories are on merge-remote-main until 2026-08-20

Kai's direction, 2026-08-17. Landed by Darren (director seat). Both AGENTS.md files point here for the revert date, so this comment is the record they rely on.

  • coilyco-gaming/sirens-echo at a3c1672
  • coilyco-bridge/deploy at f691fef

Each commit changes the ward.workflow frontmatter and the prose in the same file that still described the PR lane, so an agent reading either one gets the same answer. Ward's own help defines the lane as: merge-remote-main merges to main and closes.

Deliberate divergence

agentic-os#994 keeps this frontmatter byte-identical across five PR-lane repositories. Two of the five now differ on purpose. agentic-os, infrastructure, and ward are untouched and stay on pull-request-and-merge.

One thing this buys on the sirens-echo side: the prose there carried "Ward honors it only after ward#1661," a caveat specific to the PR lane. merge-remote-main is Ward's default lane, so that dependency does not apply while the swap is in force.

⚠️ The deploy half of this swap is inert until #482 lands

Verified rather than inferred. I ran the dispatcher against each repository before making the change:

$ ward agent engineer coilyco-bridge/deploy#644 --print
ward: refusing untrusted owner "coilyco-bridge"
  (allowed: coilysiren, coilyco-flight-deck, coilyco-gaming)

The refusal fires at dispatch, before any lane is read. coilyco-bridge/deploy#482 characterised this as blocking the merge verb, and the check is broader than that: no ward agent run starts against coilyco-bridge at all, so no agent reaches the point of caring which lane the frontmatter names.

So the swap does exactly what was asked and changes nothing about deploy throughput on its own. #482 is still the item that unblocks deploy, and it remains the top of the freeze list. The lane change is the second half of that fix, landed early so nothing waits on it once the allowlist moves.

The sirens-echo half is not affected by this. That owner is already trusted.

One observation worth confirming, marked as inference

The same dry run against sirens-echo cleared the owner check and then stopped at a different gate:

ward: actor admission: actor authority policy has no trusted collaborators

I have not established whether that is a property of this native session's broker credentials or a repository-level condition, and I did not go looking, because it does not change the lane swap. If agent dispatch into sirens-echo is failing for everyone, that is a freeze-level problem and it is not currently ticketed. The observation that would settle it is a dispatch attempt from a seat with full broker credentials.

Revert

Both files revert to pull-request-and-merge on 2026-08-20 alongside the freeze.

## Both repositories are on `merge-remote-main` until 2026-08-20 **Kai's direction, 2026-08-17. Landed by Darren (director seat).** Both AGENTS.md files point here for the revert date, so this comment is the record they rely on. * `coilyco-gaming/sirens-echo` at `a3c1672` * `coilyco-bridge/deploy` at `f691fef` Each commit changes the `ward.workflow` frontmatter **and** the prose in the same file that still described the PR lane, so an agent reading either one gets the same answer. Ward's own help defines the lane as: `merge-remote-main` merges to main and closes. ### Deliberate divergence agentic-os#994 keeps this frontmatter byte-identical across five PR-lane repositories. Two of the five now differ on purpose. `agentic-os`, `infrastructure`, and `ward` are untouched and stay on `pull-request-and-merge`. One thing this buys on the `sirens-echo` side: the prose there carried "Ward honors it only after ward#1661," a caveat specific to the PR lane. `merge-remote-main` is Ward's default lane, so that dependency does not apply while the swap is in force. ## ⚠️ The deploy half of this swap is inert until #482 lands **Verified rather than inferred.** I ran the dispatcher against each repository before making the change: ``` $ ward agent engineer coilyco-bridge/deploy#644 --print ward: refusing untrusted owner "coilyco-bridge" (allowed: coilysiren, coilyco-flight-deck, coilyco-gaming) ``` The refusal fires at **dispatch**, before any lane is read. https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/482 characterised this as blocking the merge verb, and the check is broader than that: **no `ward agent` run starts against `coilyco-bridge` at all**, so no agent reaches the point of caring which lane the frontmatter names. So the swap does exactly what was asked and changes nothing about deploy throughput on its own. **#482 is still the item that unblocks deploy**, and it remains the top of the freeze list. The lane change is the second half of that fix, landed early so nothing waits on it once the allowlist moves. The `sirens-echo` half is not affected by this. That owner is already trusted. ### One observation worth confirming, marked as inference The same dry run against `sirens-echo` cleared the owner check and then stopped at a different gate: ``` ward: actor admission: actor authority policy has no trusted collaborators ``` I have not established whether that is a property of this native session's broker credentials or a repository-level condition, and I did not go looking, because it does not change the lane swap. **If agent dispatch into `sirens-echo` is failing for everyone, that is a freeze-level problem and it is not currently ticketed.** The observation that would settle it is a dispatch attempt from a seat with full broker credentials. ### Revert Both files revert to `pull-request-and-merge` on 2026-08-20 alongside the freeze.
Author
Member

Amendment, 2026-08-18: the freeze covers features, not operational improvements

Decided by Kai, 2026-08-18. Recorded by Olaf (ops seat) during a post-incident review of the sirens-deep-owl-glass lane.

The freeze as written blocks on list membership alone: "If an issue is not on this list, it is behind August 20, regardless of its priority label." That reading also parks work that makes the demo lane faster and more reliable without changing anything it can do, which is the opposite of what the freeze is protecting.

The amended rule

The freeze covers features. It does not cover operational improvements.

  • Frozen - anything that changes what an agent can do. New capability, new tool or surface, new integration, new lane, a widened grant, a behaviour change a viewer could notice as different.
  • Not frozen - anything that changes only how well it does what it already does. Latency, cost, cache behaviour, error rates, retry and timeout tuning, removing dead tools from a roster, observability and alerting on existing paths.

The test to apply

Does this change what the agent can do, or only how well it does what it already does?

Only-how-well proceeds without needing a place on the six. Anything that changes what it can do stays behind August 20 unless it is one of the six.

What this deliberately does not relax

An operational change is still a change against a live surface 34 hours from a public recording, so it still carries the ordinary discipline rather than becoming free:

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

Why this came up now

A load burst on sirens-deep-owl-glass between 00:25 and 01:10 UTC measured a 32.9s median turn and a p99 sitting on the 180s ceiling, against 8.7s median on the plain sirens-deep lane. Full measurements and root cause in the issue filed alongside this comment.

None of the fixes that burst points at are features. They are an uncached prompt prefix, a tool-round cap, a dead tool still in a roster, and a progress-edit rate. Every one of them makes the August 19 lane better at exactly the job it is already contracted to do, and under the unamended freeze every one of them was parked.

The six are unchanged. This amendment widens what may proceed alongside them, and reorders nothing.

## Amendment, 2026-08-18: the freeze covers features, not operational improvements **Decided by Kai, 2026-08-18. Recorded by Olaf (ops seat) during a post-incident review of the `sirens-deep-owl-glass` lane.** The freeze as written blocks on list membership alone: "If an issue is not on this list, it is behind August 20, regardless of its priority label." That reading also parks work that makes the demo lane faster and more reliable without changing anything it can do, which is the opposite of what the freeze is protecting. ### The amended rule The freeze covers **features**. It does not cover **operational improvements**. * **Frozen** - anything that changes what an agent can do. New capability, new tool or surface, new integration, new lane, a widened grant, a behaviour change a viewer could notice as different. * **Not frozen** - anything that changes only how well it does what it already does. Latency, cost, cache behaviour, error rates, retry and timeout tuning, removing dead tools from a roster, observability and alerting on existing paths. ### The test to apply **Does this change what the agent can do, or only how well it does what it already does?** Only-how-well proceeds without needing a place on the six. Anything that changes what it can do stays behind August 20 unless it is one of the six. ### What this deliberately does not relax An operational change is still a change against a live surface 34 hours from a public recording, so it still carries the ordinary discipline rather than becoming free: * Before-state and after-state evidence, both recorded. * One meaningful variable at a time. * A rollback path identified before the change lands. * No capability surface moves under cover of a performance fix. If a change is doing both, it is a feature and it is frozen. ### Why this came up now A load burst on `sirens-deep-owl-glass` between 00:25 and 01:10 UTC measured a 32.9s median turn and a p99 sitting on the 180s ceiling, against 8.7s median on the plain `sirens-deep` lane. Full measurements and root cause in the issue filed alongside this comment. None of the fixes that burst points at are features. They are an uncached prompt prefix, a tool-round cap, a dead tool still in a roster, and a progress-edit rate. Every one of them makes the August 19 lane better at exactly the job it is already contracted to do, and under the unamended freeze every one of them was parked. **The six are unchanged.** This amendment widens what may proceed alongside them, and reorders nothing.
Author
Member

Amendment 2026-08-18: the six are re-cut around demo integrity

Decided by Kai, 2026-08-18, recorded by Darren (director seat) after a re-triage of coilyco-gaming/sirens-echo and coilyco-bridge/deploy against the overnight owl.glass incident data. This replaces the original six. The features-versus-operational-improvements amendment above still stands and is unchanged.

Why the list moved

The original six were cut to protect ability to change things and ability to survive an upstream outage. That was the right read on the evidence available on the 17th.

The overnight measurements change it. Not one of the original six touches whether a turn answers, quickly, correctly, and once. The failure that actually fired was inside the harness, not upstream: #933 records a member told "model backend unavailable" one second after the model returned HTTP 200.

Measured against roughly 39 hours to prep:

  • Zero of the original six had closed when this was written. All six were last touched by the triage pass that created them.
  • #939 - 14 inbound turns, 4 errored, 3 of them silent: no reply, no error string, no model call, dead at a ~30.5s wall.
  • #932 - 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.
  • #935 - three of three unconstrained requests flailed 15 rounds and confabulated. The same request with refusal named as an acceptable outcome stopped immediately.
  • #943 - the tool surface offered to the model drops from 86 to 0 mid-turn and back, with a dead tool re-called six times and no breaker.

The 2026-08-18 amendment released this class of work from the freeze, which was correct. Releasing is not ranking. The effect was an unbounded, unowned stability set competing against 211 open issues while six protected items sat unclosed. This amendment fixes that by making the list name the work that will be on camera.

The new six

  • coilyco-bridge/deploy#655 - verify the sirens-dowel cutover in the cluster. The rename merged as PR #652 at 02:19Z and no sirens-dowel namespace was running at 02:40Z, so main and the cluster currently disagree about what the demo lane is called. Filed in this pass, priority/P0.
  • #939 - never drop silently. One in five turns is the demo.
  • #932 - restore prompt caching and cut the fixed prefix.
  • #935 - bounded refusal, so a request no tool can serve gets a one-sentence decline instead of 15 rounds and a confident wrong answer.
  • #310 - the principal user ID leak under impersonation framing. Unchanged, stays priority/P0. The episode's published copy invites exactly this attack onto a permanent public recording.
  • coilyco-bridge/deploy#482 with coilyco-bridge/deploy#219 - the merge path, counted as one item. Kept, because every fix above is configured in coilyco-bridge/deploy and no agent seat can land a change there.

Dropped from the original six, explicitly

  • coilyco-bridge/deploy#335 - the five cached trajectories and the Ornith warm-start are dropped as unaffordable against five live harness defects. Only the named cutover trigger and its owner survive, and they move to ordinary work rather than a freeze slot. The reasoning that put 335 on the original list was dead-air insurance against an upstream outage. #933 and #939 show the dead air is being generated in-harness, where a failover does not reach it.
  • #137 - the re-measure is dropped from the list. Its named suspect was already refuted, and #933 and #939 supersede it on the member-visible symptom. Not closed, and not a judgement on the analysis.
  • coilyco-bridge/deploy#344 - kept, and deliberately not counted as a slot. What remains is filling fallbacks: [] against an answer #619 already decided, so it is one config line rather than a piece of work competing for the window.

Two proposals Kai rejected in this pass, recorded so they are not re-proposed

  • Cutting the demo lane's tool roster from 86 to roughly ten. Rejected. I proposed it to attack the 57 KB tool block, the confabulation surface, and the 86-to-0 collapse with one reversible change. The mechanism was wrong on its own evidence: #932 records prompt caching demonstrably working on this stack, 252 spans and 6,109,440 tokens over the surrounding 24h, so the prefix problem is a caching failure rather than a roster size. #943's collapse is a missing breaker and a cached: true tool list returning zero, not too many tools. The roster is also the capability the demo exists to show. Dropped, not deferred.
  • A scripted rehearsal with a hard go / no-go gate. Rejected. The call is fix and ship. The remaining hours go to the six rather than to producing evidence about them.

What this does not change

The freeze still binds both repositories. An issue not on the new list is still behind August 20 when it is a feature, and operational improvements are still released by the amendment above. The parked issues are still not mass-relabelled, for the reason the body already gives.

Revisit condition

Unchanged. The freeze lifts on 2026-08-20, or earlier if the six close and Kai says so.

## Amendment 2026-08-18: the six are re-cut around demo integrity **Decided by Kai, 2026-08-18, recorded by Darren (director seat) after a re-triage of `coilyco-gaming/sirens-echo` and `coilyco-bridge/deploy` against the overnight owl.glass incident data. This replaces the original six. The features-versus-operational-improvements amendment above still stands and is unchanged.** ### Why the list moved The original six were cut to protect **ability to change things** and **ability to survive an upstream outage**. That was the right read on the evidence available on the 17th. The overnight measurements change it. **Not one of the original six touches whether a turn answers, quickly, correctly, and once.** The failure that actually fired was inside the harness, not upstream: #933 records a member told "model backend unavailable" one second after the model returned HTTP 200. Measured against roughly 39 hours to prep: * **Zero of the original six had closed** when this was written. All six were last touched by the triage pass that created them. * **#939** - 14 inbound turns, 4 errored, **3 of them silent**: no reply, no error string, no model call, dead at a ~30.5s wall. * **#932** - **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**. * **#935** - three of three unconstrained requests flailed 15 rounds and confabulated. The same request with refusal named as an acceptable outcome stopped immediately. * **#943** - the tool surface offered to the model drops from 86 to 0 mid-turn and back, with a dead tool re-called six times and no breaker. The 2026-08-18 amendment released this class of work from the freeze, which was correct. **Releasing is not ranking.** The effect was an unbounded, unowned stability set competing against 211 open issues while six protected items sat unclosed. This amendment fixes that by making the list name the work that will be on camera. ### The new six * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/655** - verify the `sirens-dowel` cutover in the cluster. The rename merged as PR #652 at 02:19Z and no `sirens-dowel` namespace was running at 02:40Z, so `main` and the cluster currently disagree about what the demo lane is called. Filed in this pass, `priority/P0`. * **#939** - never drop silently. One in five turns is the demo. * **#932** - restore prompt caching and cut the fixed prefix. * **#935** - bounded refusal, so a request no tool can serve gets a one-sentence decline instead of 15 rounds and a confident wrong answer. * **#310** - the principal user ID leak under impersonation framing. **Unchanged, stays `priority/P0`.** The episode's published copy invites exactly this attack onto a permanent public recording. * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/482** with **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/219** - the merge path, counted as one item. Kept, because every fix above is configured in `coilyco-bridge/deploy` and no agent seat can land a change there. ### Dropped from the original six, explicitly * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/335** - the five cached trajectories and the Ornith warm-start are dropped as unaffordable against five live harness defects. **Only the named cutover trigger and its owner survive**, and they move to ordinary work rather than a freeze slot. The reasoning that put 335 on the original list was dead-air insurance against an upstream outage. #933 and #939 show the dead air is being generated in-harness, where a failover does not reach it. * **#137** - the re-measure is dropped from the list. Its named suspect was already refuted, and #933 and #939 supersede it on the member-visible symptom. Not closed, and not a judgement on the analysis. * **https://forgejo.coilysiren.me/coilyco-bridge/deploy/issues/344** - **kept**, and deliberately not counted as a slot. What remains is filling `fallbacks: []` against an answer #619 already decided, so it is one config line rather than a piece of work competing for the window. ### Two proposals Kai rejected in this pass, recorded so they are not re-proposed * **Cutting the demo lane's tool roster from 86 to roughly ten. Rejected.** I proposed it to attack the 57 KB tool block, the confabulation surface, and the 86-to-0 collapse with one reversible change. The mechanism was wrong on its own evidence: #932 records prompt caching demonstrably working on this stack, 252 spans and 6,109,440 tokens over the surrounding 24h, so the prefix problem is a caching failure rather than a roster size. #943's collapse is a missing breaker and a `cached: true` tool list returning zero, not too many tools. The roster is also the capability the demo exists to show. **Dropped, not deferred.** * **A scripted rehearsal with a hard go / no-go gate. Rejected.** The call is fix and ship. The remaining hours go to the six rather than to producing evidence about them. ### What this does not change The freeze still binds both repositories. An issue not on the new list is still behind August 20 when it is a feature, and operational improvements are still released by the amendment above. The parked issues are still not mass-relabelled, for the reason the body already gives. ### Revisit condition Unchanged. The freeze lifts on **2026-08-20**, or earlier if the six close and Kai says so.
Author
Member

The queue now lives at #981. This issue stays the boundary.

Darren (director seat), 2026-08-18T20:42Z, at Kai's direction.

This issue has been amended three times in 21 hours, so reading the current list requires reading the body plus three comments in order and reconciling them. That is the wrong artifact to hand an execution seat with 20 hours on the clock.

#981 is the execution board. It carries the current queue, sequenced, with an acceptance condition and an owning role on each item.

The split

  • #929, this issue - the boundary. What is frozen, what the features-versus-operational-improvements test is, what the freeze forecloses, and the 2026-08-20 revisit. Unchanged, and still the thing to cite when deciding whether something may proceed.
  • #981 - the queue. What is being worked, in what order, by which role, and what closing each one requires.

Nothing about the freeze's scope changes here. This is a readability split, not a scope amendment.

State of the re-cut six as of now

Three closed today, all recorded on their own issues:

  • coilyco-bridge/deploy#655 - closed. The cutover caught two harnesses on one Discord account, armed and unfired, roughly 24 hours out.
  • coilyco-bridge/deploy#482 - closed as superseded. Ward was uninstalled three minutes after the re-cut named it as item 6, so the allowlist it asked for no longer exists. Kai accepted the more-permissive resulting position deliberately. #219 survives on the board.
  • #310 - removed from the list, not closed. The member-facing half is held by a deployed identifier guard, verified against the running image and confirmed present in all three candidate builds. Its remaining demo-visible symptom folds into #935.

#939, #932, and #935 have not moved in the 18 hours since the re-cut. All three are on #981, ranked first, second by hypothesis, and third.

One correction to the re-cut, since it was mine

The re-cut described coilyco-bridge/deploy#344 as "one config line against an answer #619 already decided." That was wrong. #619's own handoff section lists four steps across three roles, and the first is an SSM credential that did not exist. The config work is now open as deploy#681 and still needs Kai's credential, an engineer decision on availability-class scoping, and a QA probe extension before it is real.

## The queue now lives at #981. This issue stays the boundary. **Darren (director seat), 2026-08-18T20:42Z, at Kai's direction.** This issue has been amended three times in 21 hours, so reading the current list requires reading the body plus three comments in order and reconciling them. That is the wrong artifact to hand an execution seat with 20 hours on the clock. **#981 is the execution board.** It carries the current queue, sequenced, with an acceptance condition and an owning role on each item. ### The split * **#929, this issue** - the **boundary**. What is frozen, what the features-versus-operational-improvements test is, what the freeze forecloses, and the 2026-08-20 revisit. Unchanged, and still the thing to cite when deciding whether something may proceed. * **#981** - the **queue**. What is being worked, in what order, by which role, and what closing each one requires. Nothing about the freeze's scope changes here. This is a readability split, not a scope amendment. ### State of the re-cut six as of now Three closed today, all recorded on their own issues: * **coilyco-bridge/deploy#655** - closed. The cutover caught two harnesses on one Discord account, armed and unfired, roughly 24 hours out. * **coilyco-bridge/deploy#482** - closed as superseded. Ward was uninstalled three minutes after the re-cut named it as item 6, so the allowlist it asked for no longer exists. Kai accepted the more-permissive resulting position deliberately. **#219 survives on the board.** * **#310** - removed from the list, **not closed**. The member-facing half is held by a deployed identifier guard, verified against the running image and confirmed present in all three candidate builds. Its remaining demo-visible symptom folds into #935. **#939, #932, and #935 have not moved in the 18 hours since the re-cut.** All three are on #981, ranked first, second by hypothesis, and third. ### One correction to the re-cut, since it was mine The re-cut described **coilyco-bridge/deploy#344** as *"one config line against an answer #619 already decided."* That was wrong. #619's own handoff section lists four steps across three roles, and the first is an SSM credential that did not exist. The config work is now open as [deploy#681](https://forgejo.coilysiren.me/coilyco-bridge/deploy/pulls/681) and still needs Kai's credential, an engineer decision on availability-class scoping, and a QA probe extension before it is real.
Sign in to join this conversation.
No milestone
No project
No assignees
1 participant
Notifications
Due date
The due date is invalid or out of range. Please use the format "yyyy-mm-dd".

No due date set.

Dependencies

No dependencies set

Reference
coilyco-gaming/sirens-echo#929
No description provided.