Implement Hackpert active engine orchestration and canonical state mutation

nsaspy commented

2026-09-02 17:24:37 +00:00

Owner

Outcome

Add Hackpert active mode on top of the passive Prolog foundation merged in #25.

Active mode is not just recommendation. It owns the expert reasoning/orchestration loop and must be able to dispatch typed Hackmode capabilities, observe results, update operation state/graph/operational KB through canonical Common Lisp APIs, and continue reasoning.

Parent tracker: #24
Foundation: #25

Required mode split

Passive

observes canonical state/tool results/IPX evidence;
produces reasoning, recommendations, plans, and graph/advice deltas;
cannot dispatch tools;
cannot mutate canonical operation state.

Active

may select and dispatch registered capabilities/providers;
may produce typed canonical state mutations;
may persist tool-call/result graph evidence;
may persist discovered assets/findings through canonical APIs;
may update operational KB state;
may advance plan/playbook state;
loops on new evidence until stop conditions are reached.

Active action protocol

Define a typed Common Lisp boundary for Prolog-emitted actions. At minimum the protocol needs actions equivalent to:

dispatch capability/provider with typed input;
add/update graph nodes/edges;
persist discovered asset/finding;
assert/retract/update operational-KB entries;
advance/fail/succeed a plan step;
stop/yield/wait for more evidence.

Do not expose arbitrary Prolog shell execution or raw Tek9 mutation. Prolog decides what action/state delta it wants; Common Lisp validates and applies it through Hackmode's canonical interfaces.

Execution loop

snapshot canonical state/graph/KB/providers
        -> Prolog active expert
        -> typed action(s)
        -> Common Lisp validation
        -> provider dispatch / canonical mutation
        -> persist tool call + result + state delta graph
        -> refresh snapshot
        -> Prolog active expert
        -> ...

Mutation semantics

Active-mode mutation is intentional and first-class, not an accidental side effect.

Every mutation must be:

operation-scoped;
typed;
attributed to expert/rule/version/run;
connected to evidence/tool-call graph nodes where applicable;
auditable/replayable enough to explain why it happened;
rejected cleanly if stale/invalid/out of policy.

Operational KB writes are allowed from active mode. Long-term/global KB promotion remains a separate explicit lifecycle decision so one noisy run does not automatically poison reusable knowledge.

Provider orchestration

Active Hackpert must use Hackmode's existing capability/provider actor runtime rather than reintroducing Hackpert's old direct shell runner.

This preserves:

typed provider inputs/outputs;
concurrency/lifecycle controls;
graphable execution results;
deterministic capability discovery/ranking;
one canonical execution path for LISH, Hackpert, and other clients.

LISH integration

Add first-class Hackpert controls to LISH:

inspect current engine mode;
select passive/active mode;
start/stop active runs;
show current expert/run/plan step;
show reasoning/action history;
show tool calls/results initiated by Hackpert;
inspect operational KB mutations and graph deltas;
show rejected active actions and why they were rejected.

LISH is the user-facing shell for Hackpert; do not add a separate Hackpert REPL.

Dependencies / sequencing

typed tool-call/result execution graph;
active action/state-delta protocol;
operational KB mutation API;
active orchestration loop;
LISH controls/inspection;
recon expert using active loop;
fuzzing/SQLi/OOB/XSS/blind-XSS experts on the same substrate.

Acceptance proof

passive mode remains unable to dispatch providers or mutate canonical state;
active mode can emit a typed provider action;
Common Lisp validates and dispatches through the provider actor runtime;
call/result are persisted into the execution graph;
active mode can persist a discovered asset/finding through canonical APIs;
active mode can update operational KB state;
the next Prolog iteration receives the updated graph/KB state;
stale/invalid action deltas are rejected without corrupting state;
switching the same engine to passive removes mutation/dispatch authority;
LISH can start/stop/inspect the active run and its mutations.

## Outcome Add Hackpert **active mode** on top of the passive Prolog foundation merged in #25. Active mode is not just recommendation. It owns the expert reasoning/orchestration loop and must be able to dispatch typed Hackmode capabilities, observe results, update operation state/graph/operational KB through canonical Common Lisp APIs, and continue reasoning. Parent tracker: #24 Foundation: #25 ## Required mode split ### Passive - observes canonical state/tool results/IPX evidence; - produces reasoning, recommendations, plans, and graph/advice deltas; - cannot dispatch tools; - cannot mutate canonical operation state. ### Active - may select and dispatch registered capabilities/providers; - may produce typed canonical state mutations; - may persist tool-call/result graph evidence; - may persist discovered assets/findings through canonical APIs; - may update operational KB state; - may advance plan/playbook state; - loops on new evidence until stop conditions are reached. ## Active action protocol Define a typed Common Lisp boundary for Prolog-emitted actions. At minimum the protocol needs actions equivalent to: - dispatch capability/provider with typed input; - add/update graph nodes/edges; - persist discovered asset/finding; - assert/retract/update operational-KB entries; - advance/fail/succeed a plan step; - stop/yield/wait for more evidence. Do **not** expose arbitrary Prolog shell execution or raw Tek9 mutation. Prolog decides what action/state delta it wants; Common Lisp validates and applies it through Hackmode's canonical interfaces. ## Execution loop ```text snapshot canonical state/graph/KB/providers -> Prolog active expert -> typed action(s) -> Common Lisp validation -> provider dispatch / canonical mutation -> persist tool call + result + state delta graph -> refresh snapshot -> Prolog active expert -> ... ``` ## Mutation semantics Active-mode mutation is intentional and first-class, not an accidental side effect. Every mutation must be: - operation-scoped; - typed; - attributed to expert/rule/version/run; - connected to evidence/tool-call graph nodes where applicable; - auditable/replayable enough to explain why it happened; - rejected cleanly if stale/invalid/out of policy. Operational KB writes are allowed from active mode. Long-term/global KB promotion remains a separate explicit lifecycle decision so one noisy run does not automatically poison reusable knowledge. ## Provider orchestration Active Hackpert must use Hackmode's existing capability/provider actor runtime rather than reintroducing Hackpert's old direct shell runner. This preserves: - typed provider inputs/outputs; - concurrency/lifecycle controls; - graphable execution results; - deterministic capability discovery/ranking; - one canonical execution path for LISH, Hackpert, and other clients. ## LISH integration Add first-class Hackpert controls to LISH: - inspect current engine mode; - select passive/active mode; - start/stop active runs; - show current expert/run/plan step; - show reasoning/action history; - show tool calls/results initiated by Hackpert; - inspect operational KB mutations and graph deltas; - show rejected active actions and why they were rejected. LISH is the user-facing shell for Hackpert; do not add a separate Hackpert REPL. ## Dependencies / sequencing 1. typed tool-call/result execution graph; 2. active action/state-delta protocol; 3. operational KB mutation API; 4. active orchestration loop; 5. LISH controls/inspection; 6. recon expert using active loop; 7. fuzzing/SQLi/OOB/XSS/blind-XSS experts on the same substrate. ## Acceptance proof - [ ] passive mode remains unable to dispatch providers or mutate canonical state; - [ ] active mode can emit a typed provider action; - [ ] Common Lisp validates and dispatches through the provider actor runtime; - [ ] call/result are persisted into the execution graph; - [ ] active mode can persist a discovered asset/finding through canonical APIs; - [ ] active mode can update operational KB state; - [ ] the next Prolog iteration receives the updated graph/KB state; - [ ] stale/invalid action deltas are rejected without corrupting state; - [ ] switching the same engine to passive removes mutation/dispatch authority; - [ ] LISH can start/stop/inspect the active run and its mutations.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Auto-RAGE Hackpert slice merged via #40 at dec92721cd32334850d0dec2915c2787dbe95fc7.

Completed the explicit engine-authority foundation:

expert-engine has closed :passive / :active authority modes;
construction defaults to passive;
passive admits reasoning but fails closed for :provider-dispatch and :canonical-mutation;
active may request those effect classes, but this slice deliberately adds no executor or mutation primitive;
denial is typed via expert-effect-denied;
unknown effects fail closed.

RED: 920784a566f4f5d271fc22578fbf65c985e93708 added tests against the missing mode/effect API.
GREEN: e64de42b9e07072bfc00cdeef2695577fb76e4eb; exact-head core and monorepo Actions both passed before expected-head squash merge.

Next owned slice remains the typed active action/state-delta protocol. Database graph/KB persistence stays fenced to the sibling worker.

Auto-RAGE Hackpert slice merged via #40 at `dec92721cd32334850d0dec2915c2787dbe95fc7`. Completed the explicit engine-authority foundation: - `expert-engine` has closed `:passive` / `:active` authority modes; - construction defaults to passive; - passive admits reasoning but fails closed for `:provider-dispatch` and `:canonical-mutation`; - active may request those effect classes, but this slice deliberately adds no executor or mutation primitive; - denial is typed via `expert-effect-denied`; - unknown effects fail closed. RED: `920784a566f4f5d271fc22578fbf65c985e93708` added tests against the missing mode/effect API. GREEN: `e64de42b9e07072bfc00cdeef2695577fb76e4eb`; exact-head `core` and `monorepo` Actions both passed before expected-head squash merge. Next owned slice remains the typed active action/state-delta protocol. Database graph/KB persistence stays fenced to the sibling worker.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Hackpert Auto-RAGE progress: merged typed active action protocol via #44 at master commit 71b3f5d817eb61adebf0f90d6eb2e47929bda083.

RED: 53b62aff3f1faa3e3ad8839d1dc7312e849851a2 specified the missing typed action API.
Exact GREEN PR head: 6f6956af5b6b9623ba2c9c8a2a0902efbde68c90; exact-head core and monorepo both passed twice (the second pair after reopening the identical head as non-draft PR #44 because the connector's ready-for-review mutation failed internally).

Protocol now provides one operation/run/expert/version/evidence-scoped expert-active-action envelope with closed payloads for provider dispatch, graph delta, discovery, operational-KB delta, plan transition, and active-run control. Validation is pure and fail-closed: passive denies every active action; active dispatch requires :provider-dispatch; canonical deltas require :canonical-mutation; run control requires :active-control; operation/run mismatch and kind/payload mismatch are rejected. No provider execution, Tek9 mutation, second scheduler, or database internals were added.

Next Hackpert-owned dependency-ready slice: consume validated :dispatch actions through Hackmode's existing canonical provider/capability runtime, while preserving a single executor/effect boundary. Graph/KB persistence consumer work remains fenced to the database worker's typed storage APIs.

Hackpert Auto-RAGE progress: merged typed active action protocol via #44 at master commit `71b3f5d817eb61adebf0f90d6eb2e47929bda083`. RED: `53b62aff3f1faa3e3ad8839d1dc7312e849851a2` specified the missing typed action API. Exact GREEN PR head: `6f6956af5b6b9623ba2c9c8a2a0902efbde68c90`; exact-head `core` and `monorepo` both passed twice (the second pair after reopening the identical head as non-draft PR #44 because the connector's ready-for-review mutation failed internally). Protocol now provides one operation/run/expert/version/evidence-scoped `expert-active-action` envelope with closed payloads for provider dispatch, graph delta, discovery, operational-KB delta, plan transition, and active-run control. Validation is pure and fail-closed: passive denies every active action; active dispatch requires `:provider-dispatch`; canonical deltas require `:canonical-mutation`; run control requires `:active-control`; operation/run mismatch and kind/payload mismatch are rejected. No provider execution, Tek9 mutation, second scheduler, or database internals were added. Next Hackpert-owned dependency-ready slice: consume validated `:dispatch` actions through Hackmode's existing canonical provider/capability runtime, while preserving a single executor/effect boundary. Graph/KB persistence consumer work remains fenced to the database worker's typed storage APIs.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Hackpert provider-orchestration slice merged as #50 at 9b8cbafa95310ae803c25553ef46aab8a65807f4.

Acceptance now covered:

passive mode rejects before provider dispatch;
active typed :dispatch actions validate exact operation/run scope;
non-dispatch mutation actions cannot enter the provider path;
accepted dispatches delegate to canonical dispatch-capability with no second executor/direct Tek9 write;
exact PR head 2e2191e168bced54dd0a74046bcc8cf44100c573 passed core, monorepo, and agent-framework-boundary CI before expected-head squash merge.

Database #45 is also now merged, so the next dependency-ready Hackpert slice can consume the operation-scoped operational-KB API and execution-graph evidence from the sibling boundary rather than implementing storage internals.

Hackpert provider-orchestration slice merged as #50 at `9b8cbafa95310ae803c25553ef46aab8a65807f4`. Acceptance now covered: - passive mode rejects before provider dispatch; - active typed `:dispatch` actions validate exact operation/run scope; - non-dispatch mutation actions cannot enter the provider path; - accepted dispatches delegate to canonical `dispatch-capability` with no second executor/direct Tek9 write; - exact PR head `2e2191e168bced54dd0a74046bcc8cf44100c573` passed core, monorepo, and agent-framework-boundary CI before expected-head squash merge. Database #45 is also now merged, so the next dependency-ready Hackpert slice can consume the operation-scoped operational-KB API and execution-graph evidence from the sibling boundary rather than implementing storage internals.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Hackpert Auto-RAGE progress: merged #52 at 7a9edae29772845f5a916df244cfee23246e5888.

Concrete provider slice now available through Hackmode's canonical capability/provider runtime:

subdomain-enumerate / subfinder
subdomain-enumerate / crtsh
http-probe / curl

The stale green #48 branch was not merged because it had diverged behind current master; the executable provider delta was forward-reconciled onto current master instead. During reconciliation, exact-head core exposed a real load-order bug where HTTP registration ran before run-http-probe existed. Registration is now deferred until concrete runners load, with a state-independent regression covering runner availability + baseline provider registration.

Exact PR head d4ea5326ff4b57c0ab71e73ca83a4d19b810e1f5 passed core, monorepo, and agent-framework-boundary, with zero unresolved review threads, before expected-head squash merge.

No database internals, direct Tek9 mutation, second executor, or StarIntel product work was introduced. Next dependency-ready Hackpert work is to consume execution evidence/operational-KB state in the expert reasoning loop through the typed database APIs, then advance plan/playbook/recon-expert behavior.

Hackpert Auto-RAGE progress: merged #52 at `7a9edae29772845f5a916df244cfee23246e5888`. Concrete provider slice now available through Hackmode's canonical capability/provider runtime: - `subdomain-enumerate / subfinder` - `subdomain-enumerate / crtsh` - `http-probe / curl` The stale green #48 branch was not merged because it had diverged behind current master; the executable provider delta was forward-reconciled onto current master instead. During reconciliation, exact-head core exposed a real load-order bug where HTTP registration ran before `run-http-probe` existed. Registration is now deferred until concrete runners load, with a state-independent regression covering runner availability + baseline provider registration. Exact PR head `d4ea5326ff4b57c0ab71e73ca83a4d19b810e1f5` passed core, monorepo, and agent-framework-boundary, with zero unresolved review threads, before expected-head squash merge. No database internals, direct Tek9 mutation, second executor, or StarIntel product work was introduced. Next dependency-ready Hackpert work is to consume execution evidence/operational-KB state in the expert reasoning loop through the typed database APIs, then advance plan/playbook/recon-expert behavior.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Hackpert Auto-RAGE slice merged via #54 at dda01c91636afdd8f22f4db2afdbfe304b19d30b.

Hackpert can now project canonical typed execution records and operational-KB entries into deterministic Prolog facts for reasoning. The layer is projection-only: no DB writes, direct Tek9 access, provider dispatch, or shadow graph/KB authority.

Exact PR head 4b1fa96793edced8d34c6cb2bc71634152074c1a passed core, monorepo, and agent-framework-boundary before expected-head squash merge.

Database sibling handoff / next dependency: expose operation-scoped typed enumeration APIs for persisted execution records and operational-KB entries (optionally run-scoped) so Hackpert can reconstruct these snapshots from canonical persisted state without owning storage traversal. Acceptance: deterministic typed results, operation isolation, no raw Tek9 objects, replay-safe ordering/identity. Until then Hackpert accepts canonical typed records supplied by the DB boundary.

Hackpert Auto-RAGE slice merged via #54 at `dda01c91636afdd8f22f4db2afdbfe304b19d30b`. Hackpert can now project canonical typed execution records and operational-KB entries into deterministic Prolog facts for reasoning. The layer is projection-only: no DB writes, direct Tek9 access, provider dispatch, or shadow graph/KB authority. Exact PR head `4b1fa96793edced8d34c6cb2bc71634152074c1a` passed core, monorepo, and agent-framework-boundary before expected-head squash merge. Database sibling handoff / next dependency: expose operation-scoped typed enumeration APIs for persisted execution records and operational-KB entries (optionally run-scoped) so Hackpert can reconstruct these snapshots from canonical persisted state without owning storage traversal. Acceptance: deterministic typed results, operation isolation, no raw Tek9 objects, replay-safe ordering/identity. Until then Hackpert accepts canonical typed records supplied by the DB boundary.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Hackpert plan/playbook slice merged via #56 at d3ffd6df26009b3f31f5dfd40005cb211fb48b64.

Completed in this slice:

generic expert-playbook-step, expert-playbook, and operation/run-scoped expert-plan primitives;
duplicate step IDs, dangling branch targets, invalid terminal branches, and unsupported stop/terminal states fail closed;
steps declare required capabilities without invoking them;
deterministic success/failure branch selection;
explicit stop signals for goal satisfied, budget exhaustion, policy denial, no viable extension, and explicit stop;
plan transitions are emitted as the existing typed :plan-transition active action and remain non-mutating until the canonical effect boundary accepts/applies them;
no second scheduler, direct Tek9 writes, or fixed recon -> LLM pipeline was introduced.

RED head: 15e608aa24b77291229794397ccdf5c35121cd43.
Exact GREEN PR head: 82643e394a00baecdab400e324dd38e408506274; core, monorepo, and agent-framework-boundary all passed before expected-head squash merge.

Database handoff still outstanding: the current hackmode-database public package exports point fetches/persistence but no operation-scoped enumeration API for execution records or operational-KB entries. Hackpert therefore still cannot rebuild a full persisted reasoning snapshot directly from canonical storage without violating the sibling fence. Needed contract: bounded operation-scoped enumeration, optionally run-filtered, returning typed execution-record / operational-kb-entry values in deterministic order and performing no mutation.

Hackpert plan/playbook slice merged via #56 at `d3ffd6df26009b3f31f5dfd40005cb211fb48b64`. Completed in this slice: - generic `expert-playbook-step`, `expert-playbook`, and operation/run-scoped `expert-plan` primitives; - duplicate step IDs, dangling branch targets, invalid terminal branches, and unsupported stop/terminal states fail closed; - steps declare required capabilities without invoking them; - deterministic success/failure branch selection; - explicit stop signals for goal satisfied, budget exhaustion, policy denial, no viable extension, and explicit stop; - plan transitions are emitted as the existing typed `:plan-transition` active action and remain non-mutating until the canonical effect boundary accepts/applies them; - no second scheduler, direct Tek9 writes, or fixed recon -> LLM pipeline was introduced. RED head: `15e608aa24b77291229794397ccdf5c35121cd43`. Exact GREEN PR head: `82643e394a00baecdab400e324dd38e408506274`; `core`, `monorepo`, and `agent-framework-boundary` all passed before expected-head squash merge. Database handoff still outstanding: the current `hackmode-database` public package exports point fetches/persistence but no operation-scoped enumeration API for execution records or operational-KB entries. Hackpert therefore still cannot rebuild a full persisted reasoning snapshot directly from canonical storage without violating the sibling fence. Needed contract: bounded operation-scoped enumeration, optionally run-filtered, returning typed `execution-record` / `operational-kb-entry` values in deterministic order and performing no mutation.

nsaspy commented

2026-09-02 17:24:45 +00:00

Author

Owner

Database dependency update: #58 merged as 78d24647b52805bf47f172d17914651fce4e4ef4.

Active-loop storage now has a typed read side matching the persisted execution/operational-KB data: operation-scoped execution enumeration, run/kind filters, operational-KB assertion/retraction enumeration, validated reconstruction from Tek9 nodes, deterministic ordering, and fail-closed scope checks.

This is database-only plumbing. The consumer that refreshes the next Hackpert iteration from these reads remains in the Hackpert lane.

Database dependency update: #58 merged as `78d24647b52805bf47f172d17914651fce4e4ef4`. Active-loop storage now has a typed read side matching the persisted execution/operational-KB data: operation-scoped execution enumeration, run/kind filters, operational-KB assertion/retraction enumeration, validated reconstruction from Tek9 nodes, deterministic ordering, and fail-closed scope checks. This is database-only plumbing. The consumer that refreshes the next Hackpert iteration from these reads remains in the Hackpert lane.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert canonical snapshot-refresh slice merged via #60 at 9e80edbe8affef1a862e144cb67d3fa4c6c075c0.

Database #58 removed the prior read-boundary blocker. Hackpert now has expert-operation-snapshot, which consumes the sibling database package's typed operation-scoped execution-record and operational-KB enumeration APIs, then delegates to the existing deterministic Prolog projection. Repeated calls re-read canonical state, so a subsequent reasoning iteration observes newly persisted execution/KB evidence.

RED: 166b3558767f9c290afb581d5f123bfb83622b87 — exact-head core failed on the missing snapshot-refresh API while monorepo/boundary remained green.
GREEN exact head: f6b60c875447d0cf2d27b24d9f4b7bc04646e19b — core, monorepo, and agent-framework-boundary all passed twice across the draft/non-draft identical-head PR transition; zero review threads/reviews before expected-head squash merge.

No direct Tek9 traversal/mutation, persistence internals, provider executor, or shadow graph/KB authority was added. This closes the specific #27 acceptance gap where the next Prolog iteration must see updated graph/operational-KB state. Next dependency-ready Hackpert work is the first recon expert over the shared plan/provider/snapshot substrate, then active-loop stop/escalation behavior.

Hackpert canonical snapshot-refresh slice merged via #60 at `9e80edbe8affef1a862e144cb67d3fa4c6c075c0`. Database #58 removed the prior read-boundary blocker. Hackpert now has `expert-operation-snapshot`, which consumes the sibling database package's typed operation-scoped execution-record and operational-KB enumeration APIs, then delegates to the existing deterministic Prolog projection. Repeated calls re-read canonical state, so a subsequent reasoning iteration observes newly persisted execution/KB evidence. RED: `166b3558767f9c290afb581d5f123bfb83622b87` — exact-head core failed on the missing snapshot-refresh API while monorepo/boundary remained green. GREEN exact head: `f6b60c875447d0cf2d27b24d9f4b7bc04646e19b` — core, monorepo, and agent-framework-boundary all passed twice across the draft/non-draft identical-head PR transition; zero review threads/reviews before expected-head squash merge. No direct Tek9 traversal/mutation, persistence internals, provider executor, or shadow graph/KB authority was added. This closes the specific #27 acceptance gap where the next Prolog iteration must see updated graph/operational-KB state. Next dependency-ready Hackpert work is the first recon expert over the shared plan/provider/snapshot substrate, then active-loop stop/escalation behavior.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert first recon-expert slice merged via #63 at ae09008407040a86f4f338b85ec54217f7a17f31.

RED exact head e9357fca5977585e884c91a08b08f7f5188014ba failed specifically because HACKMODE:EXPERT-RECON-NEXT-ACTION did not exist.

GREEN exact head 659b4b70cda300fed78f12a9fb4543defc992d55 passed core, monorepo, and agent-framework-boundary twice across the draft/non-draft identical-head transition, with zero reviews or unresolved review threads before expected-head squash merge.

Completed behavior:

fixed Prolog recon capability policy currently admits subdomain-enumerate and http-probe;
one deterministic compatible provider is selected by priority/capability/provider ordering;
generic compatible non-recon providers are excluded;
Common Lisp converts the recommendation into the existing typed :dispatch action carrying operation/run/expert/version/evidence provenance;
action selection performs zero provider effects and no persistence mutation;
execution remains exclusively behind the existing typed active-action -> canonical provider boundary.

No database internals, direct Tek9 access, second executor/scheduler, or StarIntel product work was added.

Next dependency-ready Hackpert slice: explicit active-loop stop/escalation behavior over plan + refreshed canonical snapshot evidence, including no-viable-action/stall handling without hard-coding a fixed recon->model pipeline.

Hackpert first recon-expert slice merged via #63 at `ae09008407040a86f4f338b85ec54217f7a17f31`. RED exact head `e9357fca5977585e884c91a08b08f7f5188014ba` failed specifically because `HACKMODE:EXPERT-RECON-NEXT-ACTION` did not exist. GREEN exact head `659b4b70cda300fed78f12a9fb4543defc992d55` passed core, monorepo, and agent-framework-boundary twice across the draft/non-draft identical-head transition, with zero reviews or unresolved review threads before expected-head squash merge. Completed behavior: - fixed Prolog recon capability policy currently admits `subdomain-enumerate` and `http-probe`; - one deterministic compatible provider is selected by priority/capability/provider ordering; - generic compatible non-recon providers are excluded; - Common Lisp converts the recommendation into the existing typed `:dispatch` action carrying operation/run/expert/version/evidence provenance; - action selection performs zero provider effects and no persistence mutation; - execution remains exclusively behind the existing typed active-action -> canonical provider boundary. No database internals, direct Tek9 access, second executor/scheduler, or StarIntel product work was added. Next dependency-ready Hackpert slice: explicit active-loop stop/escalation behavior over plan + refreshed canonical snapshot evidence, including no-viable-action/stall handling without hard-coding a fixed recon->model pipeline.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert loop-control slice merged via #67 at eeca3cf48c265f5fa31863996d0185cfc8722a77.

RED exact head 07b5ec1bb0fe791f2f47d948c40bf7d2838eb40c: core failed because the new loop-policy API did not exist; monorepo and agent-framework-boundary were green.
GREEN exact head a5697aecee31e951056e1080c28e82913f9fc7cd: core, monorepo, and agent-framework-boundary all passed again on the non-draft PR before expected-head squash merge.

The new pure loop-control boundary keeps reasoning strategy separate from effect authority: symbolic is the default; configurable repeated symbolic non-progress/failure escalates to direct; direct progress returns to symbolic; declared plan stop conditions preserve explicit reasons; operation/run mismatch fails closed. It adds no model/tool executor, scheduler, persistence path, direct Tek9 access, or database internals.

Next dependency-ready Hackpert work is wiring this decision boundary into an inspectable active run/controller without creating a second scheduler, then LISH run/mode/reasoning/action inspection. Prolog-RLM remains later #28 work.

Hackpert loop-control slice merged via #67 at `eeca3cf48c265f5fa31863996d0185cfc8722a77`. RED exact head `07b5ec1bb0fe791f2f47d948c40bf7d2838eb40c`: core failed because the new loop-policy API did not exist; monorepo and agent-framework-boundary were green. GREEN exact head `a5697aecee31e951056e1080c28e82913f9fc7cd`: core, monorepo, and agent-framework-boundary all passed again on the non-draft PR before expected-head squash merge. The new pure loop-control boundary keeps reasoning strategy separate from effect authority: symbolic is the default; configurable repeated symbolic non-progress/failure escalates to direct; direct progress returns to symbolic; declared plan stop conditions preserve explicit reasons; operation/run mismatch fails closed. It adds no model/tool executor, scheduler, persistence path, direct Tek9 access, or database internals. Next dependency-ready Hackpert work is wiring this decision boundary into an inspectable active run/controller without creating a second scheduler, then LISH run/mode/reasoning/action inspection. Prolog-RLM remains later #28 work.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert run-inspection slice merged via #89 at d018476a144db62c776d0f91165aa0a82810d5dd.

Adds a pure typed expert-run-inspection surface for operator/LISH consumers. It exposes operation/run identity, active/passive authority, symbolic/direct reasoning strategy, non-progress count, and last loop reason while keeping authority and reasoning strategy explicitly independent. No provider execution, canonical mutation, persistence, database internals, or scheduler is introduced.

RED 9cbaa3eb56fb00c1c13a25328951f8c88fc93507: core failed against the missing inspection API while monorepo/boundary were green.
GREEN exact head ff34505104981f404afe9d71b8af37f3fd0c5987: core, monorepo, and agent-framework-boundary all passed. PR #89 was mergeable with zero reviews and zero review threads before expected-head squash merge.

Draft #88 used the identical head and was replaced only because the ready-for-review connector mutation hit GitHub's recurring fullDatabaseId GraphQL error.

Hackpert run-inspection slice merged via #89 at `d018476a144db62c776d0f91165aa0a82810d5dd`. Adds a pure typed `expert-run-inspection` surface for operator/LISH consumers. It exposes operation/run identity, active/passive authority, symbolic/direct reasoning strategy, non-progress count, and last loop reason while keeping authority and reasoning strategy explicitly independent. No provider execution, canonical mutation, persistence, database internals, or scheduler is introduced. RED `9cbaa3eb56fb00c1c13a25328951f8c88fc93507`: core failed against the missing inspection API while monorepo/boundary were green. GREEN exact head `ff34505104981f404afe9d71b8af37f3fd0c5987`: core, monorepo, and agent-framework-boundary all passed. PR #89 was mergeable with zero reviews and zero review threads before expected-head squash merge. Draft #88 used the identical head and was replaced only because the ready-for-review connector mutation hit GitHub's recurring `fullDatabaseId` GraphQL error.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert operator-inspection progress merged via #92 at 1ace3c2620de7698d5933afb5f1563a0b7302098.

Added side-effect-free typed plan inspection alongside the existing run inspection: plan/objective identity, current step, current-step required capabilities, success/failure branches, terminal state, and deterministically normalized stop conditions. List-valued inspection fields are defensive copies. This surface performs no provider execution, canonical mutation, persistence, or scheduling.

RED: 8ba7df814811a158c0b1d4459d83469748800c95 — exact-head core failed against the missing plan-inspection API while monorepo and agent-framework-boundary passed.
GREEN: b69cc5603d1a31d07f2048fe1763aa63e75e49c8 — exact-head core, monorepo, and agent-framework-boundary all passed. PR #92 had zero reviews and zero unresolved review threads and was merged with expected-head protection.

Draft #91 carried the same exact green head but GitHub's ready-for-review mutation hit the recurring fullDatabaseId connector error; #92 reused the identical head without code changes.

Hackpert operator-inspection progress merged via #92 at `1ace3c2620de7698d5933afb5f1563a0b7302098`. Added side-effect-free typed plan inspection alongside the existing run inspection: plan/objective identity, current step, current-step required capabilities, success/failure branches, terminal state, and deterministically normalized stop conditions. List-valued inspection fields are defensive copies. This surface performs no provider execution, canonical mutation, persistence, or scheduling. RED: `8ba7df814811a158c0b1d4459d83469748800c95` — exact-head core failed against the missing plan-inspection API while monorepo and agent-framework-boundary passed. GREEN: `b69cc5603d1a31d07f2048fe1763aa63e75e49c8` — exact-head core, monorepo, and agent-framework-boundary all passed. PR #92 had zero reviews and zero unresolved review threads and was merged with expected-head protection. Draft #91 carried the same exact green head but GitHub's ready-for-review mutation hit the recurring `fullDatabaseId` connector error; #92 reused the identical head without code changes.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert action-inspection slice merged via #95 at 842563fd02619bcc306d0642f243d0bd1261edc5.

RED: c681f7dbc8309ca1a0c42f231ab32e79564f4e08 specified bounded, side-effect-free inspection for typed active actions.
GREEN exact head: 98b06522f343ec5da63398789f778114e4883cd9; fresh non-draft core, monorepo, and agent-framework-boundary all passed. PR #95 was mergeable with zero reviews and zero unresolved threads before expected-head squash merge.

Inspection now exposes action/effect kind, operation/run scope, expert/version provenance, normalized evidence IDs, and bounded payload metadata while deliberately excluding provider inputs and graph/KB payload contents. No provider execution, canonical mutation, database internals, second scheduler, or StarIntel product work was introduced.

Draft #94 was closed and reopened as #95 on the identical head because GitHub's ready-for-review mutation hit the recurring fullDatabaseId connector error.

Hackpert action-inspection slice merged via #95 at `842563fd02619bcc306d0642f243d0bd1261edc5`. RED: `c681f7dbc8309ca1a0c42f231ab32e79564f4e08` specified bounded, side-effect-free inspection for typed active actions. GREEN exact head: `98b06522f343ec5da63398789f778114e4883cd9`; fresh non-draft `core`, `monorepo`, and `agent-framework-boundary` all passed. PR #95 was mergeable with zero reviews and zero unresolved threads before expected-head squash merge. Inspection now exposes action/effect kind, operation/run scope, expert/version provenance, normalized evidence IDs, and bounded payload metadata while deliberately excluding provider inputs and graph/KB payload contents. No provider execution, canonical mutation, database internals, second scheduler, or StarIntel product work was introduced. Draft #94 was closed and reopened as #95 on the identical head because GitHub's ready-for-review mutation hit the recurring `fullDatabaseId` connector error.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Hackpert action-admission inspection slice merged via #98 at 7496a31fc476f4d230343953db4bd2714f92da73.

RED exact head 4970d9710d2a51e05bf6eb7404fd0a6a341ed415: core failed because EXPERT-ACTION-ADMISSION-INSPECTION did not exist; monorepo and agent-framework-boundary were green.

GREEN exact head d42c2b57986a24c56d914209794c5cc82426cfc7: fresh non-draft core, monorepo, and agent-framework-boundary all passed before expected-head squash merge. PR #98 was mergeable with zero reviews and zero unresolved review threads.

The new pure inspection wrapper reports :accepted or :rejected action admission, distinguishes authority denial from invalid/stale action scope, and exposes only bounded reason/effect metadata while reusing the existing payload-redacted action inspection. It does not dispatch providers, mutate canonical state, access Tek9 directly, create a scheduler, or touch StarIntel product work.

Draft #97 was closed and reopened as #98 on the identical green head because GitHub's ready-for-review mutation hit the recurring fullDatabaseId connector error.

Hackpert action-admission inspection slice merged via #98 at `7496a31fc476f4d230343953db4bd2714f92da73`. RED exact head `4970d9710d2a51e05bf6eb7404fd0a6a341ed415`: core failed because `EXPERT-ACTION-ADMISSION-INSPECTION` did not exist; monorepo and agent-framework-boundary were green. GREEN exact head `d42c2b57986a24c56d914209794c5cc82426cfc7`: fresh non-draft core, monorepo, and agent-framework-boundary all passed before expected-head squash merge. PR #98 was mergeable with zero reviews and zero unresolved review threads. The new pure inspection wrapper reports `:accepted` or `:rejected` action admission, distinguishes authority denial from invalid/stale action scope, and exposes only bounded reason/effect metadata while reusing the existing payload-redacted action inspection. It does not dispatch providers, mutate canonical state, access Tek9 directly, create a scheduler, or touch StarIntel product work. Draft #97 was closed and reopened as #98 on the identical green head because GitHub's ready-for-review mutation hit the recurring `fullDatabaseId` connector error.

nsaspy commented

2026-09-02 17:24:46 +00:00

Author

Owner

Database worker handoff: PR #121 is blocked before its database regression runs by a sibling-owned compile failure on current master 54e61994188ab99276a64dac7454cd9f01c8dc12. Exact RED head 4f9ff2ae415aa914a9219886fd3194c99df5d7b3 has monorepo + agent-framework-boundary green, but core fails compiling source/hackmode-core/expert/inspection.lisp: duplicate definitions of EXPERT-RUN-INSPECTION-OPERATION and EXPERT-RUN-INSPECTION-RUN-ID clobber the structure accessors, producing UIOP/LISP-BUILD:COMPILE-FILE-ERROR. This is inside the Hackpert sibling fence, so I am not modifying it. PR #121's database implementation remains isolated to the typed operational-KB fetch boundary and will stay unmerged until the sibling regression is repaired and exact-head core reaches the database tests.

Database worker handoff: PR #121 is blocked before its database regression runs by a sibling-owned compile failure on current master `54e61994188ab99276a64dac7454cd9f01c8dc12`. Exact RED head `4f9ff2ae415aa914a9219886fd3194c99df5d7b3` has monorepo + agent-framework-boundary green, but `core` fails compiling `source/hackmode-core/expert/inspection.lisp`: duplicate definitions of `EXPERT-RUN-INSPECTION-OPERATION` and `EXPERT-RUN-INSPECTION-RUN-ID` clobber the structure accessors, producing `UIOP/LISP-BUILD:COMPILE-FILE-ERROR`. This is inside the Hackpert sibling fence, so I am not modifying it. PR #121's database implementation remains isolated to the typed operational-KB fetch boundary and will stay unmerged until the sibling regression is repaired and exact-head core reaches the database tests.

nsaspy referenced this issue

2026-09-04 20:37:18 +00:00

Database: add explicit global KB export lifecycle #124

Rows
Columns

Implement Hackpert active engine orchestration and canonical state mutation #49

Outcome

Required mode split

Passive

Active

Active action protocol

Execution loop

Mutation semantics

Provider orchestration

LISH integration

Dependencies / sequencing

Acceptance proof