observed failure: run_round_robin.sh --check repeatedly reported one or more workers are unhealthy while one qualified worker was serving an active
Oracle consultation.
authoritative /health evidence showed three idle profiles in healthy state
and profile-b-2 in the expected busy state with activeRun: true;
capacity was active=1, maximum=4.
the runner incorrectly required every profile state to equal healthy before
it evaluated the already-present active-run and availability consistency
checks. This contradicted the server's intentional busy state for an active
worker and the skill's documented allowance for busy workers.
all four Chrome root processes and CDP listeners were stable from the prior
service start. The visible browser activity was renderer/tab activity from the
active run, not root-Chrome crash/relaunch churn.
Correction and proof
the runner now requires exact state correspondence: an active worker must be
busy; an inactive worker must be healthy. Truly unhealthy workers, a busy
worker without an active run, and a healthy worker claiming an active run
remain rejected.
shell syntax and synthetic healthy/busy/unhealthy fixtures passed.
the public --check passed while profile-b-2 remained actively busy:
round_robin_check=ok, pinned build unchanged, four slots, capacity waiting
enabled, protocol v1 disabled, production listener PID 15286 unchanged.
no green service restart, Chrome termination, prompt submission, production
configuration change, or source-release change was required.
2026-07-15 cross-process FIFO and bounded-preflight requalification
Oracle recommendation and implementation
Oracle session: round-robin-queue-final.
chosen boundary: the server remains authoritative for one active run per
isolated worker, two per explicit account family, and four total; cooperating
local CLIs use a filesystem FIFO only to decide which fresh request may probe
allocation capacity next.
added src/remote/localAdmissionQueue.ts: hashed stable namespace, atomic JSON
state under the existing file lock, strict ordinal FIFO, mode-0700 queue root,
exact process-birth identity, dead/PID-reused waiter recovery, and release
immediately after allocation rather than after the full Oracle response.
src/remote/client.ts wraps only fresh protocol-v2 allocation. Existing routed
sessions bypass the queue and preserve sticky reattachment.
the runner passes one stable service namespace and queue root beneath
~/.oracle-round-robin/admission; dynamic tunnel ports, slugs, tokens, and
build hashes are excluded from queue identity.
the runner health preflight now has a 60-second overall deadline, five-second
request bounds, bounded SSH establishment, and a safe last-failure diagnostic.
the allocator has an explicit production policy mapping profile-a and
profile-a-2 to account family A, and profile-b and profile-b-2 to family
B, with two active leases per family and four total.
Deterministic and live proof
targeted queue/client/allocator checks passed after the final PID-reuse and
mode-0700 changes: 27 tests.
a real multi-process test launched independent Node processes against one queue
root and proved the second process could not enter before the first released.
full isolation suite: 172 tests passed.
TypeScript typecheck, production build, targeted formatting, shell syntax, and
local/remote artifact digest comparison passed.
no-submit live qualification allocated all four distinct workers with exactly
two from each account family; a fifth request returned only structured
409/no_capacity with retryAfterMs: 2000; all four allocations were then
abandoned and active capacity returned to zero.
the public --check passed on the deployed build with capacity four, protocol
v1 disabled, all workers healthy, and production listener PID 15286 with
start identity Tue Jul 14 12:27:33 2026 unchanged.
staging restart waited for an unrelated active Oracle run to become terminal;
no active allocation was interrupted.
two initial launchctl bootstrap attempts hit macOS error 5 while bootout and
bootstrap occurred in one SSH command. Rollback configuration was restored and
production remained unchanged. The cause was launchd deregistration timing,
compounded on the first attempt by a duplicate-label rollback plist under
~/Library/LaunchAgents.
the corrected deployment keeps candidates and rollback plists under the green
service home, leaves only the canonical plist in ~/Library/LaunchAgents, uses
mode 0644, waits after bootout, and retries exact bootstrap without broad
process termination. The new green release then started healthy.
2026-07-15 four-slot concurrency and capacity-wait requalification
Behavior and authority
requested behavior: two concurrent Oracle workers per account, four total;
callers beyond capacity wait instead of returning a capacity blocker.
safe design: four isolated Chrome workers, not two automations sharing one
Chrome data directory.
canonical runtime authority: protocol-v2 allocator/client plus the managed
worker pool in ~/dev/oracle-wt/dual-profile-isolation-scaffold.
dependent surfaces synchronized: green deployment config/release, runner,
SKILL.md, managed-staging documentation, tests, and this receipt.
Implementation
profile-pool configuration and allocation now accept any non-empty worker list
while retaining exactly one active lease per isolated worker.
green topology is profile-a:19222, profile-b:19223,
profile-a-2:19224, and profile-b-2:19225.
capacity exhaustion is the structured response
409 {error:"no_capacity", retryAfterMs:2000}. The client retries only that
response, indefinitely by default, with the identical request and affinity
identity. Upload and submission cannot begin before allocation.
runner serialization changed from one global lock and fixed port 29474 to
per-slug locks plus invocation-private dynamically allocated SSH forwards.
Normal invocations use a cheap immutable topology/build check so four callers
do not serialize on expensive readiness probes. --check retains the full
four-worker readiness check.
concurrent model/thinking-time hydration was slower than sequential use and
initially caused Thinking chip failures. Thinking-time selection now retries
for up to ten one-second intervals. Four-way live concurrency passed afterward.
the two new empty profiles were logged out. Per explicit user direction, the
service was stopped and the already-qualified isolated staging states were
cold-cloned profile-a -> profile-a-2 and profile-b -> profile-b-2, excluding
manager identity markers, singleton files, and DevTools state. No production
Chrome state or production cookies were read or copied.
targeted implementation proof: 49 profile/config/client tests passed; the later
browser/concurrency retry slice passed 70 tests with 1 intentional skip;
TypeScript typecheck and vendor build passed.
runner boundary proof: four concurrent synthetic dry-runs used four distinct
local loopback forwards and removed every per-slug lock.
live allocator proof: four held leases occupied four different workers; a
fifth allocation returned only structured no_capacity; release made capacity
reusable.
Live four-worker and wait proof
sequential live markers passed once on every worker:
profile-b: SEQUENTIAL-ONE-QUALIFIED-20260715
profile-a-2: SEQUENTIAL-TWO-QUALIFIED-20260715
profile-b-2: SEQUENTIAL-THREE-QUALIFIED-20260715
profile-a: SEQUENTIAL-FOUR-QUALIFIED-20260715
final four-way concurrent live markers all passed in one wave:
RETRY-FINAL-ONE/ TWO/ THREE/ FOUR-QUALIFIED-20260715, routed across all
four worker IDs with durable protocol-v2 receipts.
final five-caller wave passed all five markers. Allocation IDs 00000049
through 00000052 occupied the four slots. Allocation 00000053 reused
profile-b-2 only after its prior 00000051 run became terminal, proving the
fifth caller waited rather than failing or sharing an active worker.
final health: capacity.active=0, capacity.maximum=4; all workers healthy
and available.
production listener PID 15286 and start identity remained unchanged.
observed failure: protocol-v2 runs on both managed profiles reported a Cloudflare
challenge immediately after navigation, while ordinary oracle-packx remained healthy.
transport comparison: ordinary oracle-packx reaches tv:7333 directly over
Tailscale and attaches to the long-lived production Chrome; round-robin tunnels
control traffic over SSH to tv:19474 and attaches to service-owned Chrome on
19222/19223. Both browsers access ChatGPT from tv, so the tunnel was not
visible to Cloudflare and was not the cause.
real-DOM evidence: both managed profiles showed title ChatGPT, a recognized
ChatGPT app shell, and script[src*="/challenge-platform/"]. The old detector
treated that normal bot-management script as an interstitial by itself.
correction: app-shell evidence now overrides the script; real title/widget or
short verification-page evidence remains strong; script-only weak evidence must
persist through a 12-second hydration grace before classification.
Source and deterministic proof
source: src/browser/actions/navigation.ts in
~/dev/oracle-wt/dual-profile-isolation-scaffold.
green config/plist were atomically repointed; only the exact green LaunchAgent
was booted out/bootstrap-loaded. Profile-instance identities persisted.
No-submit and live qualification
both live profiles returned title ChatGPT, app-shell true,
challenge-script true, and patched classification not-blocked.
automatic no-submit allocation covered profile-a then profile-b; replay of the
first request preserved run/profile identity; both allocations were abandoned;
capacity returned to zero.
protocol-v1 GET and POST remained 404 unsupported_protocol.
terminal sticky reattachment returned the identical marker on each original
profile without resubmission.
final health: both profiles healthy and available, zero active runs. The runner's
authenticated health timeout is 30 seconds so active no-submit readiness probes
cannot spuriously fail the five-second transport budget.
isolation: no green listener/Chrome process opened protected production paths;
production listener PID 15286 and start identity
Tue Jul 14 12:27:33 2026 remained unchanged.
Demonstrated source case
case: isolated dual-ChatGPT-account Oracle green deployment on tv.
successful output: two healthy persistent workers, automatic A/B allocation,
sticky protocol-v2 routing, attachment handoff, restart persistence, and one
successful Pro Extended attachment response plus terminal reattach per account.
proof/receipt: green /health; allocation order profile-a, profile-b;
stable profile-instance IDs across restart; qualification answers
PROFILE-A-QUALIFIED-20260715 and PROFILE-B-QUALIFIED-20260715 with identical
reattach results; production PID 15286 and start identity unchanged.
failures and corrections mined: macOS ps command parsing lost executable paths
containing spaces; CDP reports Browser with uppercase B; concurrent Chrome
startup raced; login controls can flash before cookie restoration; profile B's
first login did not persist; protocol-v1 routes were initially still readable.
Product decision
chosen form: layered skill plus deterministic runner.
target runtime and install location:
~/.agents/skills/oracle-packx-round-robin.
existing owner or new surface: new specialized transport surface composing the
existing oracle-packx prompt/bundling owner.
public boundary: scripts/run_round_robin.sh --check and one prompt-file/bundle
invocation.
why expected reuse justifies the product: avoids repeatedly reconstructing SSH
tunneling, secret handling, release matching, round-robin semantics, isolated
session roots, and production guards for every expensive Oracle consult.
existing pattern reused: oracle-packx bundle/prompt/output contract and the
existing ~/.agents/skills/*/SKILL.md format.
Invariants and parameters
Invariant
Why stable
Verification
Browser mode, Pro, Extended
User/runtime requirement
Runner-owned CLI flags
Automatic profile selection
Required for persistent server round robin
Runner never passes --remote-profile; durable metadata records assignment
Sticky identity after allocation
Prevents cross-account replay
Protocol-v2 journal/allocation tests and live reattach proof
Production untouched
Critical safety boundary
PID/start identity checked before and after each invocation
Exact green release/client match
Prevents protocol drift
Build SHA and local/remote file SHA-256 comparisons
Isolated sessions
Prevents production session/config mixing
ORACLE_HOME_DIR=~/.oracle-round-robin
Exact cleanup only
Prevents collateral process termination
Tunnel PID command verification before SIGTERM
Environment detail
Parameter/default
Secret or safety handling
SSH target
pinned tv
Batch-mode SSH; no sudo
Green endpoint
pinned remote 127.0.0.1:19474
Temporary local SSH forward on 29474
Green token
remote mode-0600 token file
Read into process memory; never printed or persisted locally
Local matching client
~/dev/oracle-wt/dual-profile-isolation-scaffold
Dist file hashes must match pinned release
Mutable run state
~/.oracle-round-robin
mode-700 home and atomic run lock
Qualification case
case: clean public-boundary preflight plus synthetic dry-run and occupied-port
failure fixture.
material differences from source case: tiny synthetic JavaScript-shaped bundle
instead of the live qualification attachment; dry-run rather than paid browser
submission; disposable /tmp files; forced local tunnel collision failure.
isolation or sandbox: disposable mktemp directory; isolated client home; exact
synthetic Python listener PID cleaned after the failure test.
fresh HOME/profile/container or scrubbed-equivalent evidence: no browser prompt
submitted; no new Chrome profile; synthetic inputs removed; production config
was never loaded because the runner injects an isolated Oracle home and remote
environment.
relevant PATH, binaries, config, credentials, caches, and prior state excluded:
production ~/.oracle/config.json, production sessions, production cookies,
and production Chrome data were excluded; matching pinned CLI/client binaries
were intentionally retained as the qualified runtime.
qualification state: passed.
approval for any live or paid target: source-case live qualification was already
explicitly scoped to at most two prompt submissions; product boundary tests
submitted no additional prompts.
cleanup plan and result: exact SSH tunnel PID stopped; synthetic fixture files,
listener PID, and lock removed; cleanup passed.
command/invocation: run_round_robin.sh --check; synthetic --dry-run with
prompt and bundle; --check while an exact synthetic listener occupied port
29474.
success proof: preflight printed pinned build, two profiles, protocol v1
disabled, and production PID; dry-run selected browser gpt-5.6-pro, detected
remote host 127.0.0.1:29474, and previewed one attachment without submission.
failure-path proof: exited 1 with local tunnel port 29474 is already occupied.
revisions required: initial remote SHA command incorrectly expanded shell $1;
replaced with remote cut; added v1-disabled and post-check production guards;
all boundary checks rerun afterward.
Release boundary
final tested artifact digest: the two SHA-256 values above.
install/init: files live directly under the requested global ~/.agents/skills
location; no production Oracle configuration changes.
help/schema/API/skill surface: YAML frontmatter, SKILL.md, runner --help,
--check, live invocation, optional repeated --file, and --dry-run.
runtime state location: ~/.oracle-round-robin; remote mutable state remains
beneath ~/.oracle-staging/dual-profile-green on tv.
publication URL or package: none; local skill only.
proof rerun after final behavior change: required and completed.
Optional rollout handoff
rollout requested: no production cutover requested.
explicitly approved targets and actions: isolated green staging only.
credentials, branch, and public-boundary constraints: token remains remote and
ephemeral locally; worktree remains uncommitted; no publication.
cleanup and rollback: stop exact runner tunnel; green LaunchAgent can be stopped
only by its exact label if separately requested.
orchestration proof gates: not applicable.
unapproved targets intentionally excluded: production Oracle routing/config,
production Chrome state, and any commit.
Decision
reusable product because: the successful workflow has stable safety/routing
invariants and the runner removes repeated high-risk manual setup.
intentionally not generalized: host, label, ports, build SHA, release, worktree,
and profile IDs remain pinned; arbitrary targets would weaken fail-closed safety.
owner and maintenance trigger: update this skill, runner, and receipt together
only after a new immutable green release passes both-account qualification.
# Oracle PackX Round-Robin Productization Receipt
## 2026-07-16 active-worker preflight correction
### Observed false failure
- observed failure: `run_round_robin.sh --check` repeatedly reported `one or
more workers are unhealthy` while one qualified worker was serving an active
Oracle consultation.
- authoritative `/health` evidence showed three idle profiles in `healthy` state
and `profile-b-2` in the expected `busy` state with `activeRun: true`;
capacity was `active=1`, `maximum=4`.
- the runner incorrectly required every profile state to equal `healthy` before
it evaluated the already-present active-run and availability consistency
checks. This contradicted the server's intentional `busy` state for an active
worker and the skill's documented allowance for busy workers.
- all four Chrome root processes and CDP listeners were stable from the prior
service start. The visible browser activity was renderer/tab activity from the
active run, not root-Chrome crash/relaunch churn.
### Correction and proof
- the runner now requires exact state correspondence: an active worker must be
`busy`; an inactive worker must be `healthy`. Truly unhealthy workers, a busy
worker without an active run, and a healthy worker claiming an active run
remain rejected.
- shell syntax and synthetic healthy/busy/unhealthy fixtures passed.
- the public `--check` passed while `profile-b-2` remained actively busy:
`round_robin_check=ok`, pinned build unchanged, four slots, capacity waiting
enabled, protocol v1 disabled, production listener PID `15286` unchanged.
- no green service restart, Chrome termination, prompt submission, production
configuration change, or source-release change was required.
- current local digests:
- `SKILL.md`: `bcaec7ed0f30d1cf1af5caed5e29e6497776ebcef3c366cc68d410c62a201d95`
- `scripts/run_round_robin.sh`:
`88b7ecaf4adcf4fb18b3bfa07556cff167f23a4c6bf5350ab2208ebad91805bb`
## 2026-07-15 cross-process FIFO and bounded-preflight requalification
### Oracle recommendation and implementation
- Oracle session: `round-robin-queue-final`.
- chosen boundary: the server remains authoritative for one active run per
isolated worker, two per explicit account family, and four total; cooperating
local CLIs use a filesystem FIFO only to decide which fresh request may probe
allocation capacity next.
- added `src/remote/localAdmissionQueue.ts`: hashed stable namespace, atomic JSON
state under the existing file lock, strict ordinal FIFO, mode-0700 queue root,
exact process-birth identity, dead/PID-reused waiter recovery, and release
immediately after allocation rather than after the full Oracle response.
- `src/remote/client.ts` wraps only fresh protocol-v2 allocation. Existing routed
sessions bypass the queue and preserve sticky reattachment.
- the runner passes one stable service namespace and queue root beneath
`~/.oracle-round-robin/admission`; dynamic tunnel ports, slugs, tokens, and
build hashes are excluded from queue identity.
- the runner health preflight now has a 60-second overall deadline, five-second
request bounds, bounded SSH establishment, and a safe last-failure diagnostic.
- the allocator has an explicit production policy mapping `profile-a` and
`profile-a-2` to account family A, and `profile-b` and `profile-b-2` to family
B, with two active leases per family and four total.
### Deterministic and live proof
- targeted queue/client/allocator checks passed after the final PID-reuse and
mode-0700 changes: 27 tests.
- a real multi-process test launched independent Node processes against one queue
root and proved the second process could not enter before the first released.
- full isolation suite: 172 tests passed.
- TypeScript typecheck, production build, targeted formatting, shell syntax, and
local/remote artifact digest comparison passed.
- no-submit live qualification allocated all four distinct workers with exactly
two from each account family; a fifth request returned only structured
`409/no_capacity` with `retryAfterMs: 2000`; all four allocations were then
abandoned and active capacity returned to zero.
- the public `--check` passed on the deployed build with capacity four, protocol
v1 disabled, all workers healthy, and production listener PID `15286` with
start identity `Tue Jul 14 12:27:33 2026` unchanged.
- no prompt-bearing qualification was submitted.
### Immutable release
- qualified source archive/build:
`4a5f05a1d46857396cda46dab5913f7d57c303d8467167404a5b3e9c0d227c7a`.
- matching local/remote artifacts:
- CLI: `4e3ae1d769b14872e6be114220c21ef9eef65e1b6beb428a8b9385fbb8fa8064`
- protocol-v2 client: `7c764509723f49e7cb959b2f45379af96447fe6ad95d020fdaf5309c43bc0916`
- browser navigation: `e251bd0ce316653fbae6c99bdff20231a889fc9756ddf8051c3063482d2c7127`
- local admission queue: `e0c0fa70c44d761aa2b3ed31170129b3677156f80e7445584b06c4a2a49d871e`
- final skill digests:
- `SKILL.md`: `bcaec7ed0f30d1cf1af5caed5e29e6497776ebcef3c366cc68d410c62a201d95`
- `scripts/run_round_robin.sh`:
`67cd2bae1bc8b9f2a15cc245ac0fae57424be8a7b63385778014a95104decfd5`
### Deployment recovery evidence
- staging restart waited for an unrelated active Oracle run to become terminal;
no active allocation was interrupted.
- two initial `launchctl bootstrap` attempts hit macOS error 5 while bootout and
bootstrap occurred in one SSH command. Rollback configuration was restored and
production remained unchanged. The cause was launchd deregistration timing,
compounded on the first attempt by a duplicate-label rollback plist under
`~/Library/LaunchAgents`.
- the corrected deployment keeps candidates and rollback plists under the green
service home, leaves only the canonical plist in `~/Library/LaunchAgents`, uses
mode `0644`, waits after bootout, and retries exact bootstrap without broad
process termination. The new green release then started healthy.
## 2026-07-15 four-slot concurrency and capacity-wait requalification
### Behavior and authority
- requested behavior: two concurrent Oracle workers per account, four total;
callers beyond capacity wait instead of returning a capacity blocker.
- safe design: four isolated Chrome workers, not two automations sharing one
Chrome data directory.
- canonical runtime authority: protocol-v2 allocator/client plus the managed
worker pool in `~/dev/oracle-wt/dual-profile-isolation-scaffold`.
- dependent surfaces synchronized: green deployment config/release, runner,
`SKILL.md`, managed-staging documentation, tests, and this receipt.
### Implementation
- profile-pool configuration and allocation now accept any non-empty worker list
while retaining exactly one active lease per isolated worker.
- green topology is `profile-a:19222`, `profile-b:19223`,
`profile-a-2:19224`, and `profile-b-2:19225`.
- capacity exhaustion is the structured response
`409 {error:"no_capacity", retryAfterMs:2000}`. The client retries only that
response, indefinitely by default, with the identical request and affinity
identity. Upload and submission cannot begin before allocation.
- runner serialization changed from one global lock and fixed port `29474` to
per-slug locks plus invocation-private dynamically allocated SSH forwards.
Normal invocations use a cheap immutable topology/build check so four callers
do not serialize on expensive readiness probes. `--check` retains the full
four-worker readiness check.
- concurrent model/thinking-time hydration was slower than sequential use and
initially caused `Thinking chip` failures. Thinking-time selection now retries
for up to ten one-second intervals. Four-way live concurrency passed afterward.
- the two new empty profiles were logged out. Per explicit user direction, the
service was stopped and the already-qualified isolated staging states were
cold-cloned `profile-a -> profile-a-2` and `profile-b -> profile-b-2`, excluding
manager identity markers, singleton files, and DevTools state. No production
Chrome state or production cookies were read or copied.
### Immutable release and deterministic proof
- qualified source archive/build:
`8bff63d9681ab90fa4997467e1fddc80dd22de097676c469792e12ffa33ef5ff`.
- matching local/remote artifacts:
- CLI: `4b74dbc5b1d2f6166c52f15db4c7d5283179358f42d42e69eca87d3cb25e1ddd`
- protocol-v2 client: `b31e59079a0b2c3e5a10b64e6b98278853c75e9dee23329929318e67e2e53f97`
- browser navigation: `e251bd0ce316653fbae6c99bdff20231a889fc9756ddf8051c3063482d2c7127`
- targeted implementation proof: 49 profile/config/client tests passed; the later
browser/concurrency retry slice passed 70 tests with 1 intentional skip;
TypeScript typecheck and vendor build passed.
- runner boundary proof: four concurrent synthetic dry-runs used four distinct
local loopback forwards and removed every per-slug lock.
- live allocator proof: four held leases occupied four different workers; a
fifth allocation returned only structured `no_capacity`; release made capacity
reusable.
### Live four-worker and wait proof
- sequential live markers passed once on every worker:
- `profile-b`: `SEQUENTIAL-ONE-QUALIFIED-20260715`
- `profile-a-2`: `SEQUENTIAL-TWO-QUALIFIED-20260715`
- `profile-b-2`: `SEQUENTIAL-THREE-QUALIFIED-20260715`
- `profile-a`: `SEQUENTIAL-FOUR-QUALIFIED-20260715`
- final four-way concurrent live markers all passed in one wave:
`RETRY-FINAL-ONE/ TWO/ THREE/ FOUR-QUALIFIED-20260715`, routed across all
four worker IDs with durable protocol-v2 receipts.
- final five-caller wave passed all five markers. Allocation IDs `00000049`
through `00000052` occupied the four slots. Allocation `00000053` reused
`profile-b-2` only after its prior `00000051` run became terminal, proving the
fifth caller waited rather than failing or sharing an active worker.
- final health: `capacity.active=0`, `capacity.maximum=4`; all workers healthy
and available.
- production listener PID `15286` and start identity remained unchanged.
### Final local skill digests
- `SKILL.md`: `6ce51dd614a3694187ab3777b97970ef38d8157bb3bc1ab4883f1dfab858641d`
- `scripts/run_round_robin.sh`:
`98e3df8edfe30a3e44f248f907045e7fe8ee4726646b3711ceb5eb0dee9036d1`
## 2026-07-15 Cloudflare false-positive requalification
### Failure and root cause
- observed failure: protocol-v2 runs on both managed profiles reported a Cloudflare
challenge immediately after navigation, while ordinary `oracle-packx` remained healthy.
- transport comparison: ordinary `oracle-packx` reaches `tv:7333` directly over
Tailscale and attaches to the long-lived production Chrome; round-robin tunnels
control traffic over SSH to `tv:19474` and attaches to service-owned Chrome on
`19222`/`19223`. Both browsers access ChatGPT from `tv`, so the tunnel was not
visible to Cloudflare and was not the cause.
- real-DOM evidence: both managed profiles showed title `ChatGPT`, a recognized
ChatGPT app shell, and `script[src*="/challenge-platform/"]`. The old detector
treated that normal bot-management script as an interstitial by itself.
- correction: app-shell evidence now overrides the script; real title/widget or
short verification-page evidence remains strong; script-only weak evidence must
persist through a 12-second hydration grace before classification.
### Source and deterministic proof
- source: `src/browser/actions/navigation.ts` in
`~/dev/oracle-wt/dual-profile-isolation-scaffold`.
- tests: `tests/browser/pageActions.test.ts` covers app-shell + script, content-rich
pages, true title/widget/short-page challenges, and weak-evidence hydration.
- targeted result: 49 passed, 1 skipped across `pageActions.test.ts` and
`navigation.test.ts`.
- repository result: isolation and lint gates passed; full suite 793 passed,
4 skipped; TypeScript/vendor build passed; `git diff --check` passed.
### Immutable green release
- superseded rollback build:
`e56d7cccf7fdedc75189e8cc0f7ae3101684e02242254e8e85029f9bbc900e3b`.
- qualified source archive/build:
`2850837d0731411a8ebde3f9a2d65b27b107bcbbd97129477b647394e5cf508d`.
- matching local/remote artifacts:
- CLI: `4b74dbc5b1d2f6166c52f15db4c7d5283179358f42d42e69eca87d3cb25e1ddd`
- protocol-v2 client: `1ae5d15d0959dcac966cde24807bb5f173fac66a9da27a4ce782c262ce79467a`
- browser navigation: `e251bd0ce316653fbae6c99bdff20231a889fc9756ddf8051c3063482d2c7127`
- green config/plist were atomically repointed; only the exact green LaunchAgent
was booted out/bootstrap-loaded. Profile-instance identities persisted.
### No-submit and live qualification
- both live profiles returned title `ChatGPT`, app-shell `true`,
challenge-script `true`, and patched classification `not-blocked`.
- automatic no-submit allocation covered profile-a then profile-b; replay of the
first request preserved run/profile identity; both allocations were abandoned;
capacity returned to zero.
- protocol-v1 GET and POST remained `404 unsupported_protocol`.
- profile-a attachment response:
`PROFILE-A-CLOUDFLARE-FIX-QUALIFIED-20260715`.
- profile-b attachment response:
`PROFILE-B-CLOUDFLARE-FIX-QUALIFIED-20260715`.
- terminal sticky reattachment returned the identical marker on each original
profile without resubmission.
- final health: both profiles healthy and available, zero active runs. The runner's
authenticated health timeout is 30 seconds so active no-submit readiness probes
cannot spuriously fail the five-second transport budget.
- isolation: no green listener/Chrome process opened protected production paths;
production listener PID `15286` and start identity
`Tue Jul 14 12:27:33 2026` remained unchanged.
## Demonstrated source case
- case: isolated dual-ChatGPT-account Oracle green deployment on `tv`.
- successful output: two healthy persistent workers, automatic A/B allocation,
sticky protocol-v2 routing, attachment handoff, restart persistence, and one
successful Pro Extended attachment response plus terminal reattach per account.
- proof/receipt: green `/health`; allocation order `profile-a`, `profile-b`;
stable profile-instance IDs across restart; qualification answers
`PROFILE-A-QUALIFIED-20260715` and `PROFILE-B-QUALIFIED-20260715` with identical
reattach results; production PID `15286` and start identity unchanged.
- failures and corrections mined: macOS `ps` command parsing lost executable paths
containing spaces; CDP reports `Browser` with uppercase B; concurrent Chrome
startup raced; login controls can flash before cookie restoration; profile B's
first login did not persist; protocol-v1 routes were initially still readable.
## Product decision
- chosen form: layered skill plus deterministic runner.
- target runtime and install location:
`~/.agents/skills/oracle-packx-round-robin`.
- existing owner or new surface: new specialized transport surface composing the
existing `oracle-packx` prompt/bundling owner.
- public boundary: `scripts/run_round_robin.sh --check` and one prompt-file/bundle
invocation.
- why expected reuse justifies the product: avoids repeatedly reconstructing SSH
tunneling, secret handling, release matching, round-robin semantics, isolated
session roots, and production guards for every expensive Oracle consult.
- existing pattern reused: `oracle-packx` bundle/prompt/output contract and the
existing `~/.agents/skills/*/SKILL.md` format.
## Invariants and parameters
| Invariant | Why stable | Verification |
| --- | --- | --- |
| Browser mode, Pro, Extended | User/runtime requirement | Runner-owned CLI flags |
| Automatic profile selection | Required for persistent server round robin | Runner never passes `--remote-profile`; durable metadata records assignment |
| Sticky identity after allocation | Prevents cross-account replay | Protocol-v2 journal/allocation tests and live reattach proof |
| Production untouched | Critical safety boundary | PID/start identity checked before and after each invocation |
| Exact green release/client match | Prevents protocol drift | Build SHA and local/remote file SHA-256 comparisons |
| Isolated sessions | Prevents production session/config mixing | `ORACLE_HOME_DIR=~/.oracle-round-robin` |
| Exact cleanup only | Prevents collateral process termination | Tunnel PID command verification before SIGTERM |
| Environment detail | Parameter/default | Secret or safety handling |
| --- | --- | --- |
| SSH target | pinned `tv` | Batch-mode SSH; no sudo |
| Green endpoint | pinned remote `127.0.0.1:19474` | Temporary local SSH forward on `29474` |
| Green token | remote mode-0600 token file | Read into process memory; never printed or persisted locally |
| Local matching client | `~/dev/oracle-wt/dual-profile-isolation-scaffold` | Dist file hashes must match pinned release |
| Mutable run state | `~/.oracle-round-robin` | mode-700 home and atomic run lock |
## Qualification case
- case: clean public-boundary preflight plus synthetic dry-run and occupied-port
failure fixture.
- material differences from source case: tiny synthetic JavaScript-shaped bundle
instead of the live qualification attachment; dry-run rather than paid browser
submission; disposable `/tmp` files; forced local tunnel collision failure.
- isolation or sandbox: disposable `mktemp` directory; isolated client home; exact
synthetic Python listener PID cleaned after the failure test.
- fresh HOME/profile/container or scrubbed-equivalent evidence: no browser prompt
submitted; no new Chrome profile; synthetic inputs removed; production config
was never loaded because the runner injects an isolated Oracle home and remote
environment.
- relevant PATH, binaries, config, credentials, caches, and prior state excluded:
production `~/.oracle/config.json`, production sessions, production cookies,
and production Chrome data were excluded; matching pinned CLI/client binaries
were intentionally retained as the qualified runtime.
- qualification state: passed.
- approval for any live or paid target: source-case live qualification was already
explicitly scoped to at most two prompt submissions; product boundary tests
submitted no additional prompts.
- cleanup plan and result: exact SSH tunnel PID stopped; synthetic fixture files,
listener PID, and lock removed; cleanup passed.
- tested artifact digest:
- `SKILL.md`: `457e2c8e5889c97b4d964d333d1f4831bd6ba0fe964a1af2f1dc70f129537df9`
- `scripts/run_round_robin.sh`: `24e7e18f1d9450059bac584f052505de44500506b9583b74bb655d0878e2a65f`
- command/invocation: `run_round_robin.sh --check`; synthetic `--dry-run` with
prompt and bundle; `--check` while an exact synthetic listener occupied port
`29474`.
- success proof: preflight printed pinned build, two profiles, protocol v1
disabled, and production PID; dry-run selected browser `gpt-5.6-pro`, detected
remote host `127.0.0.1:29474`, and previewed one attachment without submission.
- failure-path proof: exited `1` with `local tunnel port 29474 is already occupied`.
- revisions required: initial remote SHA command incorrectly expanded shell `$1`;
replaced with remote `cut`; added v1-disabled and post-check production guards;
all boundary checks rerun afterward.
## Release boundary
- final tested artifact digest: the two SHA-256 values above.
- install/init: files live directly under the requested global `~/.agents/skills`
location; no production Oracle configuration changes.
- help/schema/API/skill surface: YAML frontmatter, `SKILL.md`, runner `--help`,
`--check`, live invocation, optional repeated `--file`, and `--dry-run`.
- runtime state location: `~/.oracle-round-robin`; remote mutable state remains
beneath `~/.oracle-staging/dual-profile-green` on `tv`.
- publication URL or package: none; local skill only.
- independent boundary check: shell syntax, help, authenticated topology check,
synthetic dry-run, occupied-port failure, and secret-string scan.
- proof rerun after final behavior change: required and completed.
## Optional rollout handoff
- rollout requested: no production cutover requested.
- explicitly approved targets and actions: isolated green staging only.
- credentials, branch, and public-boundary constraints: token remains remote and
ephemeral locally; worktree remains uncommitted; no publication.
- cleanup and rollback: stop exact runner tunnel; green LaunchAgent can be stopped
only by its exact label if separately requested.
- orchestration proof gates: not applicable.
- unapproved targets intentionally excluded: production Oracle routing/config,
production Chrome state, and any commit.
## Decision
- reusable product because: the successful workflow has stable safety/routing
invariants and the runner removes repeated high-risk manual setup.
- intentionally not generalized: host, label, ports, build SHA, release, worktree,
and profile IDs remain pinned; arbitrary targets would weaken fail-closed safety.
- owner and maintenance trigger: update this skill, runner, and receipt together
only after a new immutable green release passes both-account qualification.