Harness-independent execution backend architecture (D-056, #146) #147

Merged

adrian.tkacz merged 3 commits from arch/harness-independent-backend into develop 2026-09-27 20:49:51 +00:00

Conversation 3 Commits 3 Files changed 6 +614 -21

kilo commented 2026-09-27 16:35:19 +00:00

Member

Copy link

Closes #146 (architecture-only child of program #132). READY_FOR_REVIEW — left open for the human merge gate; the owner will have the result reviewed independently with GPT-5.6 Sol before merge. Workers never merge and this PR is not self-approved.

What

Makes explicit in the authoritative documents that Scarcity Router is a harness-independent execution backend — not a backend for ZCode. ZCode is the first demanding representative harness that exposed gaps in the execution contract; the intended client population also includes Kilo, Cline, other coding-agent harnesses, OpenAI-compatible SDK clients, simple scripts and our own future agent/orchestration algorithms. One new decision, D-056 (next free number verified against docs/decisions.md), amends existing documents; no new design document, no parallel architecture, no production behavior.

Documents changed (all amendments in place)

• docs/decisions.md — new D-056 with the ten settlement points and explicit reconciliation/boundary notes (see below).
• docs/architecture.md — Context framing + new section “Harness-independent execution backend (D-056, #146)”: responsibility-boundary table (HARNESS / SCARCITY ROUTER / EXECUTION SOURCE), semantic execution model and protocol-adapter inventory, model/execution identity invariant, effective-capability intersection, client-owned tool lifecycle, transport-tier summary, layered acceptance strategy, and an explicit current-vs-accepted-architecture split.
• docs/execution-surface.md — new “Harness compatibility contract (D-056)” section: the 16 semantic areas with per-area implementation status (implemented / contract frozen with #137 pending / architecture accepted with #136 pending / partially implemented), the tool lifecycle, and the effective-capability subsection. Listener-defaults bullet reframed: loopback plain HTTP is a supported local transport, not an insecure-debug exception.
• docs/security.md — new “Client-transport tiers (D-056)” section (A loopback supported/implemented; B same-host container architecturally permitted, not implemented, never a global non-loopback plaintext listener; C cross-host/LAN/VPN TLS required; D public-CA normal path, private CA advanced, no verify=false/TOFU/silent fallback; E worker transport unchanged) + the verified-TLS bullet amended accordingly.
• docs/product.md — harness-independent positioning, client population, and the normal-configuration target (base URL + API key + logical model id).
• docs/worker-protocol.md — boundary note only: D-056's local-HTTP tiers govern external inference-API client traffic and deliberately do not weaken worker transport.

Architectural decisions made (the PR's twelve settlements)

1. Harness-independent execution backend identity (D-056 pt 1, product.md, architecture.md).
2. Harness vs Router vs execution-source responsibility (pt 2; boundary table).
3. Semantic harness compatibility contract — semantic, not payload quirks; edge-only bounded normalization; no if zcode/if kilo/if cline in the core (pt 4, execution-surface.md).
4. Client-owned tool lifecycle; backend-native tools a distinct capability domain; source-level ineligibility for tool-requiring requests — a source limitation, never an architecture limitation (pt 5; #137 investigates Codex under this rule).
5. Execution-source capability gating via the compatibility matrix, unchanged fail-closed admission (pt 5/6).
6. Multi-protocol adapter direction: one semantic execution model; Chat Completions today; future explicit Responses adapter; SR machine/control interfaces; first-party algorithms; nothing promised that does not exist (pt 7).
7. Exact model/effort invariant restated; D-054/D-055 untouched; sr-pin: escape hatch; logical ids normal (pt 3).
8. Effective capability = model ∩ channel ∩ administrator allowance; UNKNOWN stays UNKNOWN; catalog maxima are never per-source evidence — architecture for #136 (pt 6).
9. Local HTTP vs remote HTTPS policy: plain HTTP as supported local transport, tiers A–E (pt 8; security.md).
10. Local Docker plaintext boundary: permitted only under deployment isolation (dedicated/internal: true network, no published plaintext port), explicit mode required, hard to accidentally expose LAN plaintext (pt 8B).
11. Public-CA normal remote path / private-CA advanced path; private CA never required for localhost (pt 8D; aligns with #139).
12. Layered representative-harness acceptance: Tier 1 generic protocol fixtures (#133 exists), Tier 2 ZCode representative, Tier 3 Kilo/Cline as evidenced; no compatibility claims without evidence (pt 10).

Discipline

Current implemented behavior, newly accepted architecture and deferred implementation are separated everywhere (status table in execution-surface.md; “Current state versus accepted architecture” in architecture.md; D-056 Boundary). Nothing describes a target as already implemented; no Responses implementation; no #136/#137/TLS-automation/worker changes; main untouched.

Validation

• uv sync --only-dev ✓
• make check — 2100 tests OK ✓ (known pre-existing load flake #143 did not recur)
• uv run basedpyright — 0 errors / 0 warnings / 0 notes ✓
• git diff --check ✓
• Cross-document consistency reviewed against D-053/D-054/D-055, security and execution surface; D-042/D-043/D-045 explicitly reconciled in D-056's Reconciliation note.

UX impact

Positive, documentation-level: the architecture now commits to a harness needing only base URL + API key + logical model id, with loopback HTTP a first-class supported local transport — no CA hacks or pins in the normal local path. No user-facing behavior changes in this PR.

Review

Independent review requested (owner will route to GPT-5.6 Sol). Not security-critical: no code, no new trust boundary, no credential path; D-056 tightens rather than relaxes transport semantics everywhere except the explicitly bounded, not-yet-implemented container tier, which is recorded as requiring its own implementation issue before any non-loopback plaintext bind exists.

Closes #146 (architecture-only child of program #132). **READY_FOR_REVIEW** — left open for the human merge gate; the owner will have the result reviewed independently with GPT-5.6 Sol before merge. Workers never merge and this PR is not self-approved. ## What Makes explicit in the authoritative documents that **Scarcity Router is a harness-independent execution backend** — not a backend for ZCode. ZCode is the first demanding representative harness that exposed gaps in the execution contract; the intended client population also includes Kilo, Cline, other coding-agent harnesses, OpenAI-compatible SDK clients, simple scripts and our own future agent/orchestration algorithms. One new decision, **D-056** (next free number verified against `docs/decisions.md`), amends existing documents; no new design document, no parallel architecture, no production behavior. ## Documents changed (all amendments in place) - **`docs/decisions.md`** — new **D-056** with the ten settlement points and explicit reconciliation/boundary notes (see below). - **`docs/architecture.md`** — Context framing + new section “Harness-independent execution backend (D-056, #146)”: responsibility-boundary table (HARNESS / SCARCITY ROUTER / EXECUTION SOURCE), semantic execution model and protocol-adapter inventory, model/execution identity invariant, effective-capability intersection, client-owned tool lifecycle, transport-tier summary, layered acceptance strategy, and an explicit current-vs-accepted-architecture split. - **`docs/execution-surface.md`** — new “Harness compatibility contract (D-056)” section: the 16 semantic areas with per-area implementation status (implemented / contract frozen with #137 pending / architecture accepted with #136 pending / partially implemented), the tool lifecycle, and the effective-capability subsection. Listener-defaults bullet reframed: loopback plain HTTP is a **supported local transport**, not an insecure-debug exception. - **`docs/security.md`** — new “Client-transport tiers (D-056)” section (A loopback supported/implemented; B same-host container architecturally permitted, **not** implemented, never a global non-loopback plaintext listener; C cross-host/LAN/VPN TLS required; D public-CA normal path, private CA advanced, no verify=false/TOFU/silent fallback; E worker transport unchanged) + the verified-TLS bullet amended accordingly. - **`docs/product.md`** — harness-independent positioning, client population, and the normal-configuration target (base URL + API key + logical model id). - **`docs/worker-protocol.md`** — boundary note only: D-056's local-HTTP tiers govern external inference-API client traffic and deliberately do not weaken worker transport. ## Architectural decisions made (the PR's twelve settlements) 1. Harness-independent execution backend identity (D-056 pt 1, product.md, architecture.md). 2. Harness vs Router vs execution-source responsibility (pt 2; boundary table). 3. Semantic harness compatibility contract — semantic, not payload quirks; edge-only bounded normalization; no `if zcode`/`if kilo`/`if cline` in the core (pt 4, execution-surface.md). 4. Client-owned tool lifecycle; backend-native tools a distinct capability domain; source-level ineligibility for tool-requiring requests — a source limitation, never an architecture limitation (pt 5; #137 investigates Codex under this rule). 5. Execution-source capability gating via the compatibility matrix, unchanged fail-closed admission (pt 5/6). 6. Multi-protocol adapter direction: one semantic execution model; Chat Completions today; future explicit Responses adapter; SR machine/control interfaces; first-party algorithms; nothing promised that does not exist (pt 7). 7. Exact model/effort invariant restated; D-054/D-055 untouched; `sr-pin:` escape hatch; logical ids normal (pt 3). 8. Effective capability = model ∩ channel ∩ administrator allowance; UNKNOWN stays UNKNOWN; catalog maxima are never per-source evidence — architecture for #136 (pt 6). 9. Local HTTP vs remote HTTPS policy: plain HTTP as supported local transport, tiers A–E (pt 8; security.md). 10. Local Docker plaintext boundary: permitted only under deployment isolation (dedicated/`internal: true` network, no published plaintext port), explicit mode required, hard to accidentally expose LAN plaintext (pt 8B). 11. Public-CA normal remote path / private-CA advanced path; private CA never required for localhost (pt 8D; aligns with #139). 12. Layered representative-harness acceptance: Tier 1 generic protocol fixtures (#133 exists), Tier 2 ZCode representative, Tier 3 Kilo/Cline as evidenced; no compatibility claims without evidence (pt 10). ## Discipline Current implemented behavior, newly accepted architecture and deferred implementation are separated everywhere (status table in execution-surface.md; “Current state versus accepted architecture” in architecture.md; D-056 Boundary). Nothing describes a target as already implemented; no Responses implementation; no #136/#137/TLS-automation/worker changes; `main` untouched. ## Validation - `uv sync --only-dev` ✓ - `make check` — 2100 tests OK ✓ (known pre-existing load flake #143 did not recur) - `uv run basedpyright` — 0 errors / 0 warnings / 0 notes ✓ - `git diff --check` ✓ - Cross-document consistency reviewed against D-053/D-054/D-055, security and execution surface; D-042/D-043/D-045 explicitly reconciled in D-056's Reconciliation note. ## UX impact Positive, documentation-level: the architecture now commits to a harness needing only base URL + API key + logical model id, with loopback HTTP a first-class supported local transport — no CA hacks or pins in the normal local path. No user-facing behavior changes in this PR. ## Review Independent review requested (owner will route to GPT-5.6 Sol). Not security-critical: no code, no new trust boundary, no credential path; D-056 tightens rather than relaxes transport semantics everywhere except the explicitly bounded, not-yet-implemented container tier, which is recorded as requiring its own implementation issue before any non-loopback plaintext bind exists.

kilo added 1 commit 2026-09-27 16:35:19 +00:00

docs: harness-independent execution backend architecture (D-056, #146) ... 74e6768d93

Amend the authoritative documents so Scarcity Router is explicitly a
harness-independent execution backend rather than a ZCode-shaped backend:

- D-056: responsibility boundary (harness / router / execution source),
  semantic harness compatibility contract, client-owned tool lifecycle,
  capability-layer intersection (model ∩ channel ∩ admin; UNKNOWN stays
  UNKNOWN), protocol adapters around one semantic execution model, exact
  model/effort invariant restated, local-transport tiers (loopback
  supported; bounded same-host container tier accepted not implemented;
  cross-host TLS; public-CA normal path; worker transport unchanged),
  harness-independent configuration target, layered acceptance strategy.
- architecture.md: new D-056 section (boundary table, adapter inventory,
  effective capability, tool lifecycle, transport tiers, acceptance
  tiers, current-vs-accepted split).
- execution-surface.md: harness compatibility contract with per-area
  implementation status; effective-capability subsection; listener
  defaults reframed (loopback = supported local transport).
- security.md: client-transport tiers section; verified-TLS bullet
  amended per D-056.
- product.md: harness-independent positioning and client population.
- worker-protocol.md: boundary note that D-056 does not weaken worker
  transport.

Architecture-only: no production behavior; #136/#137 implementation and
the same-host container tier remain future explicit issues.

kilo referenced this pull request 2026-09-27 17:42:58 +00:00

G. TLS + certificate lifecycle (public-CA direction, private-CA fallback) #139

adrian.tkacz referenced this pull request from a commit 2026-09-27 18:08:57 +00:00

docs: remediate D-056 review findings (PR #147)

adrian.tkacz added 1 commit 2026-09-27 18:08:57 +00:00

docs: remediate D-056 review findings (PR #147) ... 5d415be221

Findings 1-4 plus the two clarifications from the REQUEST_CHANGES review:

1. Docker bridge loopback-publish (-p 127.0.0.1:8787:8787) plaintext is
   NOT currently supported (published traffic reaches the container on a
   non-loopback interface); loopback support is precise: native loopback
   and host-network containers. Bridge support is part of the accepted
   container tier, implemented under #139.
2. The same-host container tier is a property of the ENTIRE composed
   HTTP listener (execution, machine-interface, control and admin ride
   one listener, D-041), never a per-surface mix; an execution-only
   plaintext listener would be a separate explicit decision. Worker
   transport unchanged.
3. Tool-call/tool-result rows are source-capability-aware: implemented
   on evidenced server-direct channels (PASS/PARTIAL cells); #137 is
   specifically the Codex worker-source gap (both cells UNSUPPORTED).
4. Protocol layering corrected: recommendation/machine/control surfaces
   are consumers of the ONE selection/routing core, not execution-
   protocol adapters; Chat Completions, future Responses and first-party
   execution consumers are the adapters around the semantic execution
   model.
5. Model-level aggregation of differing per-route effective ceilings is
   explicitly #136 scope; D-056 fixes only the per-route intersection.
6. Issue #139 carries an additive scope note assigning the container
   tier there (frozen scope text unchanged; no new issue).

kilo commented 2026-09-27 18:19:40 +00:00

Author

Member

Copy link

Remediation + fresh exact-head review record (bounded review lifecycle: one initial review → one remediation → narrow final verification).

Remediation commit: 5d415be (74e6768 → 5d415be) — resolves all four REQUEST_CHANGES findings and both clarifications: (1) Docker bridge loopback-publish plaintext marked NOT currently supported, loopback support stated precisely (native loopback + host-network containers), bridge support assigned to the accepted container tier; (2) composed-listener boundary settled explicitly — the same-host container tier applies to the ENTIRE composed HTTP listener (D-041 one-process composition), never a per-surface mix; (3) tool-call/tool-result status is source-capability-aware (server-direct PASS/PARTIAL evidence; #137 = Codex worker-source gap only); (4) protocol layering corrected (recommendation/machine/control = consumers of the ONE selection core, not execution-protocol adapters); (5) model-level aggregation of per-route ceilings explicitly deferred to #136; (6) additive scope comment on existing #139 carries the container tier (no new issue).

Gates at 5d415be: uv sync --only-dev ✓; make check — one transient failure on the first run that did not reproduce in two consecutive full-suite runs (both 2100 OK; consistent with filed load flake #143); uv run basedpyright 0/0/0 ✓; git diff --check ✓.

Independent exact-head review (GLM-5.3-Flash, reviewed 5d415be vs a573af5): VERDICT: APPROVED. All five remediation points verified against code (gateway_server bind/TLS refusal, control_server composed listener, openai_http_evidence PASS/PARTIAL cells, codex evidence UNSUPPORTED cells, m10 Docker notes); all original settlements confirmed present; no new overclaims. UX impact: none.

DEFER findings (recorded for the human merge gate, no action required on this PR):

1. D-044's own Status line lacks a one-line amendment annotation pointing at D-056 (repo precedent: D-017/D-028) — suggested for a follow-up docs commit.
2. Issue #136's frozen scope should name the model-level aggregation settlement explicitly when next amended, so the D-056 delegation is traceable on both ends.
3. Observation only: promoting loopback plaintext from "bounded exception" to "supported local transport" is a recorded relaxation of a D-044 rule — accepted because it documents long-implemented default behavior and the container tier's implementation stays gated at security-critical #139.

PR left open, not merged, not self-approved; awaiting the owner's independent GPT-5.6 Sol review at the human merge gate.

Remediation + fresh exact-head review record (bounded review lifecycle: one initial review → one remediation → narrow final verification). **Remediation commit:** `5d415be` (`74e6768` → `5d415be`) — resolves all four REQUEST_CHANGES findings and both clarifications: (1) Docker bridge loopback-publish plaintext marked NOT currently supported, loopback support stated precisely (native loopback + host-network containers), bridge support assigned to the accepted container tier; (2) composed-listener boundary settled explicitly — the same-host container tier applies to the ENTIRE composed HTTP listener (D-041 one-process composition), never a per-surface mix; (3) tool-call/tool-result status is source-capability-aware (server-direct PASS/PARTIAL evidence; #137 = Codex worker-source gap only); (4) protocol layering corrected (recommendation/machine/control = consumers of the ONE selection core, not execution-protocol adapters); (5) model-level aggregation of per-route ceilings explicitly deferred to #136; (6) additive scope comment on existing #139 carries the container tier (no new issue). **Gates at `5d415be`:** `uv sync --only-dev` ✓; `make check` — one transient failure on the first run that did not reproduce in two consecutive full-suite runs (both 2100 OK; consistent with filed load flake #143); `uv run basedpyright` 0/0/0 ✓; `git diff --check` ✓. **Independent exact-head review (GLM-5.3-Flash, reviewed `5d415be` vs `a573af5`):** `VERDICT: APPROVED`. All five remediation points verified against code (gateway_server bind/TLS refusal, control_server composed listener, openai_http_evidence PASS/PARTIAL cells, codex evidence UNSUPPORTED cells, m10 Docker notes); all original settlements confirmed present; no new overclaims. UX impact: none. **DEFER findings (recorded for the human merge gate, no action required on this PR):** 1. D-044's own Status line lacks a one-line amendment annotation pointing at D-056 (repo precedent: D-017/D-028) — suggested for a follow-up docs commit. 2. Issue #136's frozen scope should name the model-level aggregation settlement explicitly when next amended, so the D-056 delegation is traceable on both ends. 3. Observation only: promoting loopback plaintext from "bounded exception" to "supported local transport" is a recorded relaxation of a D-044 rule — accepted because it documents long-implemented default behavior and the container tier's implementation stays gated at security-critical #139. PR left open, not merged, not self-approved; awaiting the owner's independent GPT-5.6 Sol review at the human merge gate.

adrian.tkacz referenced this pull request from a commit 2026-09-27 18:55:06 +00:00

docs: resolve D-056 review round-2 consistency findings (PR #147)

adrian.tkacz added 1 commit 2026-09-27 18:55:06 +00:00

docs: resolve D-056 review round-2 consistency findings (PR #147) ... 7d7ce94486

1. Transport-policy contradiction resolved: the D-056 local-transport
   tiers govern CLIENT traffic to the COMPOSED HTTP listener as a whole.
   In an explicit host-local mode the listener may carry execution,
   machine, control and admin HTTP over plaintext inside the accepted
   local trust boundary, with authentication, authorization, CSRF and
   credential separation fully enforced — only transport confidentiality
   changes inside that boundary. Cross-host stays TLS-required; worker
   protocol transport completely unchanged; no per-surface TLS/plaintext
   mixing on one listener. (D-056 pt 8E + Reconciliation; security.md
   intro + tier E; worker-protocol.md note.)
2. Docker bridge loopback publication made unambiguous: a host
   publication bound ONLY to loopback (-p 127.0.0.1:8787:8787) is an
   allowed target of the explicit local-container mode; LAN/WAN-reachable
   plaintext publication is forbidden; the implementation may bind a
   non-loopback address INSIDE the container in that explicit mode; what
   is forbidden is a general/default non-loopback plaintext bind with no
   deployment-boundary enforcement. #139 must prove host-locality and
   fail closed when the bounded mode is not explicitly selected.
   (D-056 pt 8B; security.md tier B.)
3. Tool-cell evidence statement narrowed to evidence-backed server-direct
   presets; the evidence-free generic preset defaults to UNKNOWN.
   (D-056 pt 5; execution-surface.md lifecycle paragraph.)
4. D-044 status line annotated (additive, no history rewrite): non-loopback
   transport rule amended in part by D-056 for the explicit bounded
   same-host composed-listener tier only.

kilo commented 2026-09-27 19:06:01 +00:00

Author

Member

Copy link

Round-2 remediation + fresh exact-head review record (bounded review lifecycle: round-2 REQUEST_CHANGES → narrow remediation → fresh exact-head review).

Remediation commit: 7d7ce94 (5d415be → 7d7ce94) — three consistency fixes plus the recommended D-044 annotation:

1. Transport-policy contradiction resolved. The D-056 local-transport tiers now govern CLIENT traffic to the COMPOSED HTTP listener as a whole (D-056 pt 8(E) + Reconciliation, security.md intro + tier E, worker-protocol.md). In an explicit host-local mode the listener may carry execution, machine, control and admin HTTP over plaintext inside the accepted local trust boundary, with authentication, authorization, CSRF and credential separation fully enforced — only transport confidentiality changes inside that bounded boundary. Cross-host remains TLS-required; worker protocol transport completely unchanged; no per-surface TLS/plaintext mixing on one listener. The old "control API and machine interfaces keep their existing transport rules" / "inference-API client traffic" phrasings are gone.
2. Docker bridge loopback publication unambiguous. A host publication bound ONLY to loopback (-p 127.0.0.1:8787:8787) is an allowed target of the explicit local-container mode; LAN/WAN-reachable plaintext publication is forbidden; the implementation may bind a non-loopback address INSIDE the container in that explicit mode (today's code refuses any non-loopback plaintext bind); what is forbidden is a general/default non-loopback plaintext bind with no deployment-boundary enforcement. #139 must prove host-locality and fail closed when the bounded mode is not explicitly selected. The contradictory "no externally published plaintext application port" phrase is removed (D-056 pt 8(B); security.md tier B).
3. Tool-cell claim narrowed to evidence-backed server-direct presets, noting the evidence-free generic OpenAI-compatible preset defaults every cell to UNKNOWN (D-056 pt 5; execution-surface.md).
4. D-044 Status line annotated additively (dated, no history rewrite): non-loopback transport rule amended in part by D-056 for the explicit bounded same-host composed-listener tier only.

Gates at 7d7ce94: uv sync --only-dev ✓; make check — one transient failure on the first run that did not reproduce in two consecutive full-suite runs (both 2100 OK), consistent with filed load flake #143; uv run basedpyright 0/0/0 ✓; git diff --check ✓.

Independent exact-head review (GLM-5.3-Flash, reviewed 7d7ce94 vs a573af5): VERDICT: APPROVED. All three fixes verified against code (control_server composed listener + separate optional worker listener; gateway_server non-loopback plaintext refusal; openai_http_evidence GENERIC_PRESET all-UNKNOWN); the delta vs round 2 is exactly the three fixes plus the additive D-044 annotation; D-056 substance otherwise unchanged; deferred-vs-implemented honesty confirmed across all documents. UX impact: none.

DEFER findings (recorded for the human merge gate; no action on this PR):

1. D-044's historical rule-2 example ("a loopback or LAN Ollama endpoint" for bounded plain-HTTP origins) is narrower in the amended security.md bullet; the implemented outbound-origin discipline already refuses non-loopback plaintext, so this is pre-existing docs-vs-code drift correctly left unrewritten — a one-clause extension of the D-044 annotation in a future docs touch would reconcile it fully.
2. Cosmetic: execution-surface.md status-table cell "same-host container tier accepted, implemented under #139" can be misread in isolation; disambiguated by the Status column and security.md tier B.

PR left open at 7d7ce94, not merged, not self-approved; awaiting the owner's independent GPT-5.6 Sol review at the human merge gate.

Round-2 remediation + fresh exact-head review record (bounded review lifecycle: round-2 REQUEST_CHANGES → narrow remediation → fresh exact-head review). **Remediation commit:** `7d7ce94` (`5d415be` → `7d7ce94`) — three consistency fixes plus the recommended D-044 annotation: 1. **Transport-policy contradiction resolved.** The D-056 local-transport tiers now govern CLIENT traffic to the COMPOSED HTTP listener as a whole (D-056 pt 8(E) + Reconciliation, security.md intro + tier E, worker-protocol.md). In an explicit host-local mode the listener may carry execution, machine, control and admin HTTP over plaintext inside the accepted local trust boundary, with authentication, authorization, CSRF and credential separation fully enforced — only transport confidentiality changes inside that bounded boundary. Cross-host remains TLS-required; worker protocol transport completely unchanged; no per-surface TLS/plaintext mixing on one listener. The old "control API and machine interfaces keep their existing transport rules" / "inference-API client traffic" phrasings are gone. 2. **Docker bridge loopback publication unambiguous.** A host publication bound ONLY to loopback (`-p 127.0.0.1:8787:8787`) is an allowed target of the explicit local-container mode; LAN/WAN-reachable plaintext publication is forbidden; the implementation may bind a non-loopback address INSIDE the container in that explicit mode (today's code refuses any non-loopback plaintext bind); what is forbidden is a general/default non-loopback plaintext bind with no deployment-boundary enforcement. #139 must prove host-locality and fail closed when the bounded mode is not explicitly selected. The contradictory "no externally published plaintext application port" phrase is removed (D-056 pt 8(B); security.md tier B). 3. **Tool-cell claim narrowed to evidence-backed server-direct presets**, noting the evidence-free generic OpenAI-compatible preset defaults every cell to UNKNOWN (D-056 pt 5; execution-surface.md). 4. **D-044 Status line annotated** additively (dated, no history rewrite): non-loopback transport rule amended in part by D-056 for the explicit bounded same-host composed-listener tier only. **Gates at `7d7ce94`:** `uv sync --only-dev` ✓; `make check` — one transient failure on the first run that did not reproduce in two consecutive full-suite runs (both 2100 OK), consistent with filed load flake #143; `uv run basedpyright` 0/0/0 ✓; `git diff --check` ✓. **Independent exact-head review (GLM-5.3-Flash, reviewed `7d7ce94` vs `a573af5`):** `VERDICT: APPROVED`. All three fixes verified against code (control_server composed listener + separate optional worker listener; gateway_server non-loopback plaintext refusal; openai_http_evidence GENERIC_PRESET all-UNKNOWN); the delta vs round 2 is exactly the three fixes plus the additive D-044 annotation; D-056 substance otherwise unchanged; deferred-vs-implemented honesty confirmed across all documents. UX impact: none. **DEFER findings (recorded for the human merge gate; no action on this PR):** 1. D-044's historical rule-2 example ("a loopback or LAN Ollama endpoint" for bounded plain-HTTP origins) is narrower in the amended security.md bullet; the implemented outbound-origin discipline already refuses non-loopback plaintext, so this is pre-existing docs-vs-code drift correctly left unrewritten — a one-clause extension of the D-044 annotation in a future docs touch would reconcile it fully. 2. Cosmetic: execution-surface.md status-table cell "same-host container tier accepted, implemented under #139" can be misread in isolation; disambiguated by the Status column and security.md tier B. PR left open at `7d7ce94`, not merged, not self-approved; awaiting the owner's independent GPT-5.6 Sol review at the human merge gate.

adrian.tkacz approved these changes 2026-09-27 20:49:46 +00:00

adrian.tkacz merged commit f574e51ce5 into develop 2026-09-27 20:49:51 +00:00

adrian.tkacz deleted branch arch/harness-independent-backend 2026-09-27 20:49:52 +00:00

adrian.tkacz referenced this pull request from a commit 2026-09-27 20:49:52 +00:00

Merge pull request 'Harness-independent execution backend architecture (D-056, #146)' (#147) from arch/harness-independent-backend into develop

kilo referenced this pull request 2026-09-27 21:01:22 +00:00

Representative-client compatibility + operations hardening program (ZCode acceptance follow-up) #132

kilo referenced this pull request 2026-09-27 21:01:48 +00:00

D. Gateway effective limits + Codex output-token semantics #136

kilo referenced this pull request 2026-09-27 21:01:48 +00:00

E. Codex client tool-call round trip (security-critical) #137

kilo referenced this pull request 2026-09-27 21:01:49 +00:00

F. First-class worker service lifecycle (systemd) #138

kilo referenced this pull request 2026-09-27 21:01:49 +00:00

G. TLS + certificate lifecycle (public-CA direction, private-CA fallback) #139

kilo referenced this pull request 2026-09-27 21:01:49 +00:00

H. Administrator UX cleanup (derived resources, honest counters, revoked-worker deletion, limits + client integration) #140

kilo referenced this pull request 2026-09-27 21:01:49 +00:00

I. Integrated acceptance + coherent install/first-run + program closeout #141

kilo referenced this pull request 2026-10-04 22:16:05 +00:00

Align Scarcity Router with the Creatidy system architecture and register actionable gaps #173

Sign in to join this conversation.

Reviewers

No reviewers

adrian.tkacz

Labels

Clear labels   

accessibility

Barrier affecting people with disabilities

  

bug

Something isn't working

  

documentation

Improvements or additions to documentation

  

duplicate

This issue or pull request already exists

  

enhancement

New feature or request

  

good first issue

Good for newcomers

  

help wanted

Extra attention is needed

  

invalid

This doesn't seem right

  

P1

Owner-declared highest explicit implementation priority (loop ordering rank)

  

P2

Required product and integration work

  

P3

Deferred or producer-dependent work

  

question

Further information is requested

  

wontfix

This will not be worked on

No labels

accessibility

bug

documentation

duplicate

enhancement

good first issue

help wanted

invalid

P1

P2

P3

question

wontfix

Milestone

Clear milestone

No items

No milestone

Projects

Clear projects

No items

No project

Assignees

Clear assignees

No assignees

2 participants

Notifications

Subscribe

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

Creatidy/scarcity-router!147

No description provided.