yxflc11 · Oct 10, 2026 · Oct 10, 2026 · Oct 10, 2026
diff --git a/docs/WORK_EXECUTION_CONTRACT.md b/docs/WORK_EXECUTION_CONTRACT.md
@@ -2,7 +2,7 @@
 
 Status: design contract, 2026-09-23; not an implemented API or schema. This refines the
 approved delivery plan, rather than creating another roadmap.
-The [durability experiment](../experiments/durable-execution/README.md) supplies narrow evidence;
+The [retired durability experiment](https://github.com/Peerframe/openbot/blob/50837bea7bf63290aab250f84fe1b705f942efdb/experiments/durable-execution/README.md) supplies narrow evidence;
 [Temporal is selected as the target recovery owner](decisions/0046-temporal-as-recovery-owner.md),
 while product integration and production activation remain open.
 

diff --git a/docs/decisions/0046-temporal-as-recovery-owner.md b/docs/decisions/0046-temporal-as-recovery-owner.md
@@ -33,6 +33,7 @@ Exact candidates and maintained upstream sources:
 Primary contracts and actual measurements are recorded in
 [durability qualification](../research/durable-execution-qualification.md),
 [Temporal review](../research/temporal-durability-review.md),
+[SDK composition](../research/sdk-durability-integration.md),
 [public work journey](../research/work-temporal-journey.md),
 [PostgreSQL operations](../research/temporal-postgres-operations.md),
 [transport](../research/temporal-transport-security.md), and

diff --git a/docs/decisions/0050-typescript-control-plane.md b/docs/decisions/0050-typescript-control-plane.md
@@ -249,10 +249,12 @@ semantics retained by the migration; their old setup commands are historical.
 
 The new Work path retains the [task authority](../research/python-task-authority.md),
 [model service](../research/python-model-services.md), [admission](../research/work-domain-admission.md),
+[native Task scope](../research/python-native-task-capabilities.md),
 [artifact publication](../research/work-artifact-publication.md) and
 [lookup-only reconciliation](../research/work-reconciliation-commands.md) decisions.
 The [former supervisor](../research/python-control-runtime-supervision.md),
-[runtime activation](../research/python-runtime-activation.md) and
+[runtime activation](../research/python-runtime-activation.md),
+[product Worker composition](../research/python-product-runtime.md) and
 [headless acceptance](../research/headless-runtime-acceptance.md) records document the retired
 process/wheel implementation; TS Work recovery, Linux native execution and packaged lifecycle checks
 replace those mechanisms without introducing a second authority or retrying unknown effects.
diff --git a/docs/research/durable-execution-qualification.md b/docs/research/durable-execution-qualification.md
@@ -83,7 +83,7 @@ these are implemented and independently verified.
 ## Executed evidence
 
 The original local harness passed all ten cases on 2026-09-23, including the deliberately
-unsafe negative control. See [the reproducible experiment](../../experiments/durable-execution/README.md).
+unsafe negative control. See [the reproducible experiment](https://github.com/Peerframe/openbot/blob/50837bea7bf63290aab250f84fe1b705f942efdb/experiments/durable-execution/README.md).
 This validates the narrow fixture behavior and leaves the selection gates above open.
 
 Additional controlled suspension cases keep the old process alive: SIGSTOP before external

diff --git a/docs/research/python-native-task-capabilities.md b/docs/research/python-native-task-capabilities.md
@@ -73,6 +73,6 @@ No legacy channel/Run was fabricated. Two sanitized actual histories replayed wi
 SDK and no registered Activities; root reran those replays after integration.
 
 Portable source, exact assertions and safe evidence are in
-[the native capability journey](../../experiments/work-journey/native-capabilities/README.md).
+[the native capability journey](https://github.com/Peerframe/openbot/blob/50837bea7bf63290aab250f84fe1b705f942efdb/experiments/work-journey/native-capabilities/README.md).
 This qualifies real product orchestration and recovery with a synthetic model transport;
 separate live Kimi and Linux/browser tests must not inherit a positive result from it.
diff --git a/docs/research/repository-hygiene-guards.md b/docs/research/repository-hygiene-guards.md
@@ -46,6 +46,8 @@ uses no new dependency, and every failure message says how to fix it. It enforce
    `check-developer-entrypoints.ts`.
 6. Every `experiments/*` directory is still run by code, scripts, `package.json` or CI.
 7. Source files open with a comment saying what they are for.
+8. No tracked `.py`, `requirements*.txt`, `pyproject.toml` or `.python-version` file, since P5
+   retired Python (ADR-0050); an explicit allow-list in the script is empty by default.
 
 Rules 5 and 7 start from a recorded baseline in `scripts/hygiene-baseline.json` that may only go
 down; the check asks for the baseline to be lowered as soon as files improve.

diff --git a/docs/research/typescript-control-plane-p4.md b/docs/research/typescript-control-plane-p4.md
@@ -844,7 +844,7 @@ therefore supplies the Docker-local name without changing content. The failed or
 owned runtime was removed and zero-container/firewall host baseline matched. P4 is still open.
 The private packet now gives the exact independently pinned loaded Squid manifest its original
 local tag before the unchanged config/manifest/platform guard. This follows the retained command
-[offline-load binding](../../experiments/linux-execution/deadline_probe.py); it changes no image
+[offline-load binding](https://github.com/Peerframe/openbot/blob/50837bea7bf63290aab250f84fe1b705f942efdb/experiments/linux-execution/deadline_probe.py); it changes no image
 content, source enforcement, routes, runtime budget or host daemon. Actual failure stderr was not
 retained publicly, so the transport-name attribution rests on the observed operation and loader
 contract, not an original daemon message. Fresh native browser and all current-head CI remain required.

diff --git a/docs/research/work-temporal-journey.md b/docs/research/work-temporal-journey.md
@@ -461,7 +461,7 @@ passed, with its Turbo lint/typecheck/test/build task output served from cache.
 
 ## One constructor-time Agent across two concurrent Runs (2026-09-24, revision 2)
 
-The [disposable probe](../../experiments/work-journey/multirun_port_probe.py) asks whether one
+The [disposable probe](https://github.com/Peerframe/openbot/blob/50837bea7bf63290aab250f84fe1b705f942efdb/experiments/work-journey/multirun_port_probe.py) asks whether one
 Agent built once **before** Worker startup can serve two concurrent Workflows with separate Run
 inputs. It requires a disposable Temporal frontend address via `--address` and pins the SDK
 versions at startup. Reviewed pins are unchanged:

diff --git a/experiments/browser-execution/README.md b/experiments/browser-execution/README.md
@@ -1,8 +1,9 @@
 # Chromium/runsc boundary experiment
 
-The CDP qualification entry is `probe.ts`, with typed transport and artifact modules pinned by
-`browser_a1.py`. The egress guest runs `egress_probe.ts`; both use existing guest Node type
-stripping. No host Node prerequisite or new compilation layer is added to the Python qualifiers.
+The CDP qualification entry is `probe.ts`, with typed transport and artifact modules. The egress
+guest runs `egress_probe.ts`; both use existing guest Node type stripping. The Python qualifiers
+that pinned and launched them (`browser_a1.py`, `qualify_egress.py`) were deleted after P5; read
+them at [the last revision that had them](https://github.com/Peerframe/openbot/tree/50837bea7bf63290aab250f84fe1b705f942efdb/experiments/browser-execution).
 The TS entry at `d426715` passed the authorized single fixed Linux/runsc case on 2026-09-30
 (`deadline-a1-ts0930a`): both browsers closed, the profile reopened, synthetic DOM and 1280×800 PNG
 were independently checked, and original 180-second Invocation expiry emptied the cgroup and