DIAGNOSTIC ONLY — do not merge: bare tip on GitHub-hosted runners

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.

[jack-champagne]

Do not merge. Do not review. This exists only to get a CI signal and will be closed as soon as it reports.

It is local/amicode at f43233e6 with one change: test.yml's runs-on swapped from blacksmith-4vcpu-* to ubuntu-latest / windows-latest (plus the Node 24.15 pin for the known Playwright/Chromium hang). No source changes, no spec changes, no timeout changes.

Why: the fork has 0 registered runners for the blacksmith-4vcpu-* labels it requests, so no test job has ever run on the default branch — the 106 commits since v1.17.3-amicode.9 have never been CI-verified as a whole. #2 (which fixes the runner labels for real) came back 4/5 red after rebasing onto the tip, and I need to know which of those failures belong to the tip and which belong to #2's own e2e accommodations.

Specifically this separates:

• the 4 session.llm native recorded fixture mismatches (a system-prompt change added a "Communicating with the user" block without re-recording fixtures) — expected to reproduce here, which would confirm they are the tip's, not fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's
• unit exit 137 / OOM / SIGILL — expected to reproduce, would confirm it's the hosted-runner memory ceiling rather than anything fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 did
• the 2 e2e failures (regression-session-list … directory path request resolves, smoke-session-timeline … while paging through history) — if these pass here, they are fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's stale selectHomeProject accommodation and fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 needs updating; if they fail here, the tip has a real regression and the amicode.10 cut should wait

[jack-champagne]

Done its job — closing.

It established the bare-tip baseline that let us tell pre-existing failures apart from regressions: e2e (linux) 2 failed / 5 passed, unit (linux) = stale LLM fixtures + v2 location + the exit-137 segfault.

Since then, on that evidence:

• fix(ci): rewire typecheck/test workflows for local/amicode default branch #2 landed the real runner fix (and v2 location now passes — that was fix(ci): rewire typecheck/test workflows for local/amicode default branch #2's b37d0f5b9)
• test(server): isolate the keychain in the connections-route suite #91 fixed the segfault (keychain isolation) and dropped the windows unit lane
• ci: fix amicode-release UI gate check for the unconditional layout memo #81 fixed the release gate check, verified against a real 160MB build
• fix(provider): port upstream provider/model transform for Claude 5 #73 merged with CI provably at baseline

Superseded by real CI on the default branch, which now actually runs.