nsaspy/prolog-rlm

Fork 0

Complete supervised subagent result propagation #149

Merged

lost-rob0t merged 1 commit from agent/subagent-supervised-call-conformance into main

2026-08-24 15:53:13 +00:00

lost-rob0t commented

2026-08-21 11:48:11 +00:00

(Migrated from github.com)

Part of #144. This completes the child-owned supervision slice that the original test-only draft exposed.

Runtime invariant

agent_supervised_call/6 and its canonical async surface admit ground work into an existing logical child and run a ground trusted host closure on the existing bounded agent worker pool.
Successful child worker results propagate exactly once to the supervisor as canonical child_result data with explicit parent/child trace correlation.
Parent-mailbox backpressure emits child_result_backpressure and never claims delivery.
Parent cancellation signals admitted child work and preserves the child's cancelled terminal state.
Agent trace events use a named ground agent_event representation, so typed subagent envelopes are closed data.

`rlm_subagent` integration

The ordinary registered tool/capability/authority boundary remains unchanged.
The trusted handler now runs rlm_completion_execute as child-owned supervised work instead of calling the public synchronous completion facade in the tool-handler thread.
The merged inert KB binding unknown -> delegate_subagent -> tool(rlm_subagent) is exercised end to end.
The same ground typed envelope is recorded by the child, returned by the tool, and delivered to the parent.
Bare rlm_cancelled/1 remains a control exception through injected planner/model handlers and the completion execute boundary.

No second scheduler, authority system, executable model/KB term, or product-specific AgentProlog behavior is added.

Verification

focused agent/subagent/completion/async checks: 74/74
fresh aggregate: 75 suites, 772/772 passed, 0 failed/timeout/blocked/fixme
hostile runner integrity: 19/19
deterministic benchmark: 16/16
deep experiment: 15/15
runtime/load/demo/diff checks: passed
Nix flake: all 3 checks passed with one job/core

Parent #144 remains open for any remaining policy/depth/evidence acceptance not proved by this slice. Exact-head GitHub Actions, including REAL OpenRouter, are required before merge.

Part of #144. This completes the child-owned supervision slice that the original test-only draft exposed. ## Runtime invariant - `agent_supervised_call/6` and its canonical async surface admit ground work into an existing logical child and run a ground trusted host closure on the existing bounded agent worker pool. - Successful child worker results propagate exactly once to the supervisor as canonical `child_result` data with explicit parent/child trace correlation. - Parent-mailbox backpressure emits `child_result_backpressure` and never claims delivery. - Parent cancellation signals admitted child work and preserves the child's cancelled terminal state. - Agent trace events use a named ground `agent_event` representation, so typed subagent envelopes are closed data. ## `rlm_subagent` integration - The ordinary registered tool/capability/authority boundary remains unchanged. - The trusted handler now runs `rlm_completion_execute` as child-owned supervised work instead of calling the public synchronous completion facade in the tool-handler thread. - The merged inert KB binding `unknown -> delegate_subagent -> tool(rlm_subagent)` is exercised end to end. - The same ground typed envelope is recorded by the child, returned by the tool, and delivered to the parent. - Bare `rlm_cancelled/1` remains a control exception through injected planner/model handlers and the completion execute boundary. No second scheduler, authority system, executable model/KB term, or product-specific AgentProlog behavior is added. ## Verification - focused agent/subagent/completion/async checks: 74/74 - fresh aggregate: 75 suites, 772/772 passed, 0 failed/timeout/blocked/fixme - hostile runner integrity: 19/19 - deterministic benchmark: 16/16 - deep experiment: 15/15 - runtime/load/demo/diff checks: passed - Nix flake: all 3 checks passed with one job/core Parent #144 remains open for any remaining policy/depth/evidence acceptance not proved by this slice. Exact-head GitHub Actions, including REAL OpenRouter, are required before merge.

lost-rob0t commented

2026-08-21 12:45:36 +00:00

(Migrated from github.com)

Worker A audit on exact head 4fd2d215536f3217d8ed6b36d04189d6ec2981c1:

The branch has not been superseded and no overlapping subagent PR exists. #150 is evolution-evaluator work under #142/#141; #122/#125/#132 remain separate owned surfaces.

Important conformance finding: all exact-head workflows are green, but this does not prove the intended success-propagation assertion. The branch diff contains only the new test + runner registration; prolog/rlm_agent.pl is unchanged and notify_parent_from_reply/3 still forwards only kind:failed. The successful worker path still yields kind:result locally and never enqueues result(child(...), ok(...)) to the parent. Also, current trace events are flat dicts produced by trace_add/3; the test currently expects Event.data.parent / Event.data.child, which does not match that representation.

Keep this PR draft. Next implementation should stay in this branch and reuse existing machinery only:

extend notify_parent_from_reply/3 (or the narrow adjacent seam) to forward successful kind:result exactly once as result(child(agent(AgentId)), ok(Value)) and trace a correlated child_result event;
correct the fixture to assert the actual canonical trace shape and make sure the test demonstrably fails without the runtime change;
then add the trusted child-owned supervised-call seam over existing request -> dispatch -> bounded worker -> result machinery, so cancellation sees the worker in agent_worker/4;
switch rlm_subagent away from direct handler-thread rlm_completion/4 only after that seam is proven;
compose merged #146 command binding + #147 ordinary tool boundary into the end-to-end unresolved/delegate/child/parent-resume fixture.

Do not merge on the present green checks: they are green despite the runtime behavior still being absent, so the test gate itself needs repair first.

Worker A audit on exact head `4fd2d215536f3217d8ed6b36d04189d6ec2981c1`: The branch has not been superseded and no overlapping subagent PR exists. #150 is evolution-evaluator work under #142/#141; #122/#125/#132 remain separate owned surfaces. Important conformance finding: all exact-head workflows are green, but this does **not** prove the intended success-propagation assertion. The branch diff contains only the new test + runner registration; `prolog/rlm_agent.pl` is unchanged and `notify_parent_from_reply/3` still forwards only `kind:failed`. The successful worker path still yields `kind:result` locally and never enqueues `result(child(...), ok(...))` to the parent. Also, current trace events are flat dicts produced by `trace_add/3`; the test currently expects `Event.data.parent` / `Event.data.child`, which does not match that representation. Keep this PR draft. Next implementation should stay in this branch and reuse existing machinery only: 1. extend `notify_parent_from_reply/3` (or the narrow adjacent seam) to forward successful `kind:result` exactly once as `result(child(agent(AgentId)), ok(Value))` and trace a correlated `child_result` event; 2. correct the fixture to assert the actual canonical trace shape and make sure the test demonstrably fails without the runtime change; 3. then add the trusted child-owned supervised-call seam over existing request -> dispatch -> bounded worker -> result machinery, so cancellation sees the worker in `agent_worker/4`; 4. switch `rlm_subagent` away from direct handler-thread `rlm_completion/4` only after that seam is proven; 5. compose merged #146 command binding + #147 ordinary tool boundary into the end-to-end unresolved/delegate/child/parent-resume fixture. Do not merge on the present green checks: they are green despite the runtime behavior still being absent, so the test gate itself needs repair first.

lost-rob0t commented

2026-08-21 12:50:23 +00:00

(Migrated from github.com)

Worker A pushed corrected conformance oracle first at 9854aa8cd680dfd9adc661f0fc48a3e8bf525e65.

The fixture now asserts the actual flat trace_add/3 event representation (Event.parent, Event.child, Event.result) and requires exactly one correlated child_result event. It also uses an explicit equality assertion for the parent's canonical last_result rather than leaving the intended structure as an ordinary unification.

This head intentionally still has no runtime fix. Therefore the deterministic suite MUST fail until successful child worker replies are forwarded through supervision. If exact-head CI reports green again, treat that as a test-runner/gating defect and diagnose it before any runtime implementation or merge. #149 remains draft.

Worker A pushed corrected conformance oracle first at `9854aa8cd680dfd9adc661f0fc48a3e8bf525e65`. The fixture now asserts the actual flat `trace_add/3` event representation (`Event.parent`, `Event.child`, `Event.result`) and requires exactly one correlated `child_result` event. It also uses an explicit equality assertion for the parent's canonical `last_result` rather than leaving the intended structure as an ordinary unification. This head intentionally still has no runtime fix. Therefore the deterministic suite MUST fail until successful child worker replies are forwarded through supervision. If exact-head CI reports green again, treat that as a test-runner/gating defect and diagnose it before any runtime implementation or merge. #149 remains draft.

lost-rob0t commented

2026-08-21 12:57:09 +00:00

(Migrated from github.com)

Coordination audit from a second worker: I am deliberately NOT modifying this branch because it advanced under me from 4fd2d215... to 9854aa8c... while I was checking it.

Exact-head audit: all four Actions gates are green, but the PR diff still changes only test/rlm_subagent_supervision_test.pl + test/run_tests.pl; there is still no production implementation in this PR. Current prolog/rlm_agent.pl still has notify_parent_from_reply/3 handling only kind:failed, so successful worker replies are not forwarded to the parent by the production code.

Therefore the green full-suite result is not sufficient evidence that the new success-propagation oracle is actually proving the missing invariant. Before implementation/merge, run the test set in isolation (for example run_tests([rlm_subagent_supervision])) or otherwise prove the new test is registered/executed and genuinely red against the current production code. Keep #149 draft until that is resolved, then implement the success path on the existing supervision machinery rather than adding another scheduler.

I will stay off #149's files while this worker owns the branch.

Coordination audit from a second worker: I am deliberately NOT modifying this branch because it advanced under me from `4fd2d215...` to `9854aa8c...` while I was checking it. Exact-head audit: all four Actions gates are green, but the PR diff still changes only `test/rlm_subagent_supervision_test.pl` + `test/run_tests.pl`; there is still no production implementation in this PR. Current `prolog/rlm_agent.pl` still has `notify_parent_from_reply/3` handling only `kind:failed`, so successful worker replies are not forwarded to the parent by the production code. Therefore the green full-suite result is not sufficient evidence that the new success-propagation oracle is actually proving the missing invariant. Before implementation/merge, run the test set in isolation (for example `run_tests([rlm_subagent_supervision])`) or otherwise prove the new test is registered/executed and genuinely red against the current production code. Keep #149 draft until that is resolved, then implement the success path on the existing supervision machinery rather than adding another scheduler. I will stay off #149's files while this worker owns the branch.

lost-rob0t commented

2026-08-21 12:59:02 +00:00

(Migrated from github.com)

Root cause of the false-green found, with upstream SWI semantics checked.

test/run_tests.pl declares :- initialization(main, main)., but later it consult(run_tool_mcp_async_tests). That consulted file itself declares :- initialization(focused_main, main). SWI-Prolog executes the last goal registered with initialization(Goal, main) as the program main. Therefore the helper's focused_main replaces the outer full-suite entrypoint. focused_main runs only run_tool_mcp_async_cases and halts 0, which exactly matches the CI log: canonical async cases run, then the process exits without the full PlUnit suite.

So #149's new test is not being exercised by swipl -q -s test/run_tests.pl; this also means the advertised deterministic PlUnit gate has been false-green more broadly.

Fix should preserve the standalone focused runner without registering its main when it is consulted by the aggregate runner (e.g. split reusable predicates from the executable entrypoint, or otherwise ensure only test/run_tests.pl owns initialization(..., main) in the aggregate program). Then prove run_tests([rlm_subagent_supervision]) is red before adding production success propagation.

I am still staying off the branch/files while this worker owns #149.

Root cause of the false-green found, with upstream SWI semantics checked. `test/run_tests.pl` declares `:- initialization(main, main).`, but later it `consult(run_tool_mcp_async_tests).` That consulted file itself declares `:- initialization(focused_main, main).` SWI-Prolog executes **the last goal registered with `initialization(Goal, main)`** as the program main. Therefore the helper's `focused_main` replaces the outer full-suite entrypoint. `focused_main` runs only `run_tool_mcp_async_cases` and halts 0, which exactly matches the CI log: canonical async cases run, then the process exits without the full PlUnit suite. So #149's new test is not being exercised by `swipl -q -s test/run_tests.pl`; this also means the advertised deterministic PlUnit gate has been false-green more broadly. Fix should preserve the standalone focused runner without registering its `main` when it is consulted by the aggregate runner (e.g. split reusable predicates from the executable entrypoint, or otherwise ensure only `test/run_tests.pl` owns `initialization(..., main)` in the aggregate program). Then prove `run_tests([rlm_subagent_supervision])` is red before adding production success propagation. I am still staying off the branch/files while this worker owns #149.

lost-rob0t (Migrated from github.com) reviewed

2026-08-21 13:21:36 +00:00

lost-rob0t (Migrated from github.com) left a comment

Coordination only; no #149 branch changes from me. I opened isolated #152 for P0 #151. It touches only test/run_tool_mcp_async_tests.pl and conditions its initialization(focused_main, main) on prolog_load_context(script, true), so standalone focused execution remains intact while aggregate consultation cannot steal main. Clean-pack + Tree-sitter are already green on exact head; CI/Nix are running. Once #152 merges, please rebase #149 and verify the test-only success-propagation state becomes genuinely red before adding runtime implementation.

Coordination only; no #149 branch changes from me. I opened isolated #152 for P0 #151. It touches only `test/run_tool_mcp_async_tests.pl` and conditions its `initialization(focused_main, main)` on `prolog_load_context(script, true)`, so standalone focused execution remains intact while aggregate consultation cannot steal `main`. Clean-pack + Tree-sitter are already green on exact head; CI/Nix are running. Once #152 merges, please rebase #149 and verify the test-only success-propagation state becomes genuinely red before adding runtime implementation.

lost-rob0t commented

2026-08-24 09:17:01 +00:00

(Migrated from github.com)

Repaired the false-green test-only draft on exact current main. The isolated oracle failed before production changes; this head adds the existing-scheduler supervised-call seam, child-owned rlm_subagent completion, successful parent propagation, cancellation/backpressure behavior, ground trace envelopes, and the compiled unknown-command integration. Local exact-head evidence: focused 74/74; aggregate 75 suites and 772/772; runner integrity 19/19; benchmarks 16/16 and 15/15; runtime/load/demo/diff clean; all 3 Nix checks passed with one job/core. Marked ready for exact-head GitHub review. Parent #144 remains open for acceptance not covered by this slice.

Repaired the false-green test-only draft on exact current main. The isolated oracle failed before production changes; this head adds the existing-scheduler supervised-call seam, child-owned `rlm_subagent` completion, successful parent propagation, cancellation/backpressure behavior, ground trace envelopes, and the compiled unknown-command integration. Local exact-head evidence: focused 74/74; aggregate 75 suites and 772/772; runner integrity 19/19; benchmarks 16/16 and 15/15; runtime/load/demo/diff clean; all 3 Nix checks passed with one job/core. Marked ready for exact-head GitHub review. Parent #144 remains open for acceptance not covered by this slice.

lost-rob0t commented

2026-08-24 09:18:52 +00:00

(Migrated from github.com)

Roadmap reconciliation added as required by the repository contract; exact reviewed head is now a70539ddd6d635faac84a5f45f6a5aba900fe5d3. The code and tests are unchanged from the previously green head. Waiting for all exact-head checks, including REAL OpenRouter, to turn green again before merge.

Roadmap reconciliation added as required by the repository contract; exact reviewed head is now `a70539ddd6d635faac84a5f45f6a5aba900fe5d3`. The code and tests are unchanged from the previously green head. Waiting for all exact-head checks, including REAL OpenRouter, to turn green again before merge.