Add Epiphany role specialist launch
This commit is contained in:
Meta Vens
parent
a36e2b4a4b
commit
0e7f8a7685
16 changed files with 1790 additions and 56 deletions
|
|
@ -32,6 +32,7 @@ The spine exists in twenty-four live paths:
|
|||
- reorientation policy: app-server declares experimental read-only `thread/epiphany/reorient`, reuses the same live/stored thread view plus mapped pressure/freshness/watcher signals, and derives a bounded `resume`/`regather` verdict with checkpoint-path/frontier reasons without mutating, compacting, scheduling, notifying, or continuing work in [common.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/common.rs:365), [v2.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/v2.rs:4514), [codex_message_processor.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server/src/codex_message_processor.rs:4394), and [codex_message_processor.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server/src/codex_message_processor.rs:11424).
|
||||
- CRRC coordinator recommendation: app-server declares experimental read-only `thread/epiphany/crrc`, reuses the same pressure/freshness/watcher/reorientation mapping plus the fixed `reorient-worker` binding and result read-back helper, and returns one bounded recommendation such as continue, prepare checkpoint, launch worker, wait, review, accept, or regather manually without mutating, launching, accepting, compacting, notifying, scheduling, or silently continuing work in [common.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/common.rs:370), [v2.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/v2.rs:4661), and [codex_message_processor.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server/src/codex_message_processor.rs:4546).
|
||||
- role ownership projection: app-server declares experimental read-only `thread/epiphany/roles`, reuses the same state/jobs/pressure/reorient/result/CRRC signals as the coordinator loop, and projects implementation, modeling/checkpoint, verification/review, and reorientation lanes with owner roles, status, linked jobs, authority scopes, and recommended scene actions without launching, scheduling, accepting, compacting, notifying, or mutating state in [common.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/common.rs:350), [v2.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/v2.rs:4161), and [codex_message_processor.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server/src/codex_message_processor.rs:4310).
|
||||
- role specialist launch/read-back: app-server declares experimental `thread/epiphany/roleLaunch` and read-only `thread/epiphany/roleResult`. `roleLaunch` accepts only fixed modeling/checkpoint and verification/review role ids, builds bounded specialist payloads from authoritative Epiphany state, routes them through the existing `epiphany_launch_job` job-control seam, and emits the normal `jobLaunch` state update for `jobBindings` only. `roleResult` resolves the fixed or requested binding through the current `agent_jobs` backend and projects structured output as a reviewable finding with verdict, summary, next safe move, inspected files, frontier/evidence ids, raw result, and errors without mutation, promotion, scheduling, or hidden continuation. Implementation stays owned by the main coding agent, and reorientation stays on the CRRC-specific `reorientLaunch` / `reorientResult` path.
|
||||
- MVP operator view: `tools/epiphany_mvp_status.py` launches app-server over stdio, starts or reads one thread, gathers scene, pressure, reorient, jobs, roles, reorient result, and CRRC responses, and renders them as text or JSON so the loop can be dogfooded without reading Rust.
|
||||
- reorientation launch: app-server now declares experimental `thread/epiphany/reorientLaunch`, requires a loaded thread plus durable investigation checkpoint, consumes the live reorientation verdict, and launches one fixed `reorient-worker` binding through the existing job-control seam with explicit resume-versus-regather scope, checkpoint-derived payload, and structured-output expectations instead of pretending CRRC suddenly grew a scheduler in [common.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/common.rs:370), [v2.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/v2.rs:4575), and [codex_message_processor.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server/src/codex_message_processor.rs:4519).
|
||||
- reorientation result read-back: app-server now declares experimental read-only `thread/epiphany/reorientResult`, defaults to the fixed `reorient-worker` binding, resolves that binding through the current `agent_jobs` backend when available, reads the single backend item result, and projects a reviewable finding with mode, summary, next safe move, checkpoint-validity flag, inspected files, frontier ids, evidence ids, raw result, and any job/item errors without mutating state, promoting evidence, or scheduling follow-up work in [common.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/common.rs:374), [v2.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server-protocol/src/protocol/v2.rs:4610), and [codex_message_processor.rs](E:/Projects/EpiphanyAgent/vendor/codex/codex-rs/app-server/src/codex_message_processor.rs:4684).
|
||||
|
|
@ -1309,7 +1310,7 @@ These are deliberately not shipped yet:
|
|||
- no automatic Compact-Rehydrate-Reorient-Continue coordinator; current code preserves and rehydrates valid Epiphany snapshots across resume/rollback/compaction, exposes context pressure, carries a durable investigation checkpoint packet, returns a read-only reorientation verdict, can explicitly launch one bounded reorientation worker, read that worker's result back through a reviewable surface, and explicitly accept completed findings into typed state, but runtime launch/acceptance policy and next-step execution are still explicit rather than ambient scheduler behavior.
|
||||
- no automatic watcher-driven graph or semantic invalidation.
|
||||
- no code-intelligence graph.
|
||||
- no specialist-agent scheduler.
|
||||
- no broad specialist-agent scheduler or marketplace; only fixed manual modeling/checkpoint and verification/review role templates exist.
|
||||
- no Epiphany-owned long-running job executor beyond the current runtime `agent_jobs` seam.
|
||||
- no GUI implementation beyond the read-only scene, jobs, freshness, context, and pressure projection surfaces.
|
||||
- no durable retrieval-summary write from `thread/epiphany/retrieve`.
|
||||
|
|
@ -1329,6 +1330,8 @@ stored Epiphany state
|
|||
-> reflected as a read-only resume/regather verdict from checkpoint + freshness + watcher + pressure
|
||||
-> reflected as a read-only CRRC recommendation over verdict + worker/result state
|
||||
-> reflected as read-only MVP role ownership over implementation, modeling/checkpoint, verification/review, and reorientation lanes
|
||||
-> fixed modeling/checkpoint and verification/review role workers can be launched explicitly
|
||||
-> fixed role worker results can be read back for review without mutation
|
||||
-> can be consumed explicitly by one bounded reorient-worker launch surface
|
||||
-> retrieval/indexing can inform future work
|
||||
-> explicit distillation can draft observation/evidence patches
|
||||
|
|
@ -1338,9 +1341,7 @@ stored Epiphany state
|
|||
-> successful writes emit thread/epiphany/stateUpdated
|
||||
```
|
||||
|
||||
The current remaining missing organ is not the red pen, not the first chewing motion, not the first state-change bell, not the first client-shaped scene, not the first job/progress reflection, not the first live bound-job notification, not the first freshness lens, not the first watcher-backed invalidation read, not the first targeted graph/evidence context read, not the first context-pressure gauge, not the first durable investigation checkpoint packet, not the first bounded reorientation verdict, not the first thin launcher seam, not the first explicit launch/interrupt authority surface over that seam, not the first runtime consumer that can act on the reorientation verdict without becoming a hidden tyrant, not the first read-back surface for that worker's findings, not the first explicit acceptance write for completed findings, not the first read-only CRRC coordinator recommendation, and no longer the first dogfood-facing operator view over the landed loop. Those exist. Proposal now has the first bits of map memory, focus, and selection hygiene: it can auto-select a bounded evidence-backed observation cluster when ids are omitted, requires accepting recent evidence behind selected observations, prioritizes stronger selected observations for proposal wording, reuses existing architecture nodes by concrete code-ref/path/id checks before creating new path nodes, can rescue unanchored graph nodes through strict unique semantic overlap, follows graph links and incident edges into the frontier, and reports match-kind-aware churn pressure from the actual proposal shape. Distillation now also has the first source-output-aware teeth for tool/model summaries. Promotion now notices risky deltas instead of just checking the shape of the envelope, treats graph expansion as risky even if the patch tries to mumble "low pressure" while dragging a new wall into the blueprint, token-matches strong verifier kinds so `contest` cannot sneak past as `test`, and validates investigation checkpoints against known evidence before the packet can be inked. Successful update/promote writes now emit `thread/epiphany/stateUpdated` with `source: "update"` or `source: "promote"`, event-level revision, and typed changed fields so clients no longer need to poll blindly, infer the cause from timing, or diff the whole state to know which top-level sections moved. Runtime `agent_job_progress` background events now also emit `thread/epiphany/jobsUpdated` with changed bound-job payloads so clients no longer need to poll `thread/epiphany/jobs` just to watch one real owner move, the bound jobs now reflect launcher id, authority scope, and backend kind/job id instead of treating raw runtime pointers as the architecture, `jobLaunch` / `jobInterrupt` can create or tear down those bindings without pretending reflection is authority, `reorientLaunch` can launch one fixed `reorient-worker` job from the checkpoint verdict without turning the verdict itself into a writer, `reorientResult` can read the worker's structured output back as a reviewable finding without promoting it, `reorientAccept` can explicitly bank a completed finding into typed observation/evidence plus optional scratch/checkpoint state, and `crrc` can recommend the next explicit action over that loop without doing the action itself. The scene, jobs, freshness, context, pressure, reorient, crrc, reorient-result, and MVP status surfaces give clients compact or bounded read-only views without making the GUI/client authoritative, scene/context surface the checkpoint packet back out without becoming writers, freshness surfaces watcher telemetry without becoming a hidden worker, reorient turns those signals into an explicit wakeup verdict that one explicit launch surface can consume on purpose, crrc tells the operator whether to continue, checkpoint, launch, wait, review, accept, or regather manually, and the status tool prints the answer without requiring a Rust spelunk. The remaining missing organs are a second backend if `agent_jobs` stops being enough and eventually role-scoped specialist ownership without silently auto-promoting anything.
|
||||
|
||||
Update: first role-scoped specialist ownership is now landed as read-only `thread/epiphany/roles`. The next missing organ is explicit role-scoped launch/read-back templates for modeling/checkpoint and verification/review, still manual and review-gated.
|
||||
The current remaining missing organ is not the red pen, not the first chewing motion, not the first state-change bell, not the first client-shaped scene, not the first job/progress reflection, not the first live bound-job notification, not the first freshness lens, not the first watcher-backed invalidation read, not the first targeted graph/evidence context read, not the first context-pressure gauge, not the first durable investigation checkpoint packet, not the first bounded reorientation verdict, not the first thin launcher seam, not the first explicit launch/interrupt authority surface over that seam, not the first runtime consumer that can act on the reorientation verdict without becoming a hidden tyrant, not the first read-back surface for that worker's findings, not the first explicit acceptance write for completed findings, not the first read-only CRRC coordinator recommendation, not the first role ownership board, not the first fixed role specialist launch/read-back templates, and no longer the first dogfood-facing operator view over the landed loop. Those exist. Proposal now has the first bits of map memory, focus, and selection hygiene: it can auto-select a bounded evidence-backed observation cluster when ids are omitted, requires accepting recent evidence behind selected observations, prioritizes stronger selected observations for proposal wording, reuses existing architecture nodes by concrete code-ref/path/id checks before creating new path nodes, can rescue unanchored graph nodes through strict unique semantic overlap, follows graph links and incident edges into the frontier, and reports match-kind-aware churn pressure from the actual proposal shape. Distillation now also has the first source-output-aware teeth for tool/model summaries. Promotion now notices risky deltas instead of just checking the shape of the envelope, treats graph expansion as risky even if the patch tries to mumble "low pressure" while dragging a new wall into the blueprint, token-matches strong verifier kinds so `contest` cannot sneak past as `test`, and validates investigation checkpoints against known evidence before the packet can be inked. Successful update/promote writes now emit `thread/epiphany/stateUpdated` with `source: "update"` or `source: "promote"`, event-level revision, and typed changed fields so clients no longer need to poll blindly, infer the cause from timing, or diff the whole state to know which top-level sections moved. Runtime `agent_job_progress` background events now also emit `thread/epiphany/jobsUpdated` with changed bound-job payloads so clients no longer need to poll `thread/epiphany/jobs` just to watch one real owner move, the bound jobs now reflect launcher id, authority scope, and backend kind/job id instead of treating raw runtime pointers as the architecture, `jobLaunch` / `jobInterrupt` can create or tear down those bindings without pretending reflection is authority, `reorientLaunch` can launch one fixed `reorient-worker` job from the checkpoint verdict without turning the verdict itself into a writer, `reorientResult` can read the worker's structured output back as a reviewable finding without promoting it, `reorientAccept` can explicitly bank a completed finding into typed observation/evidence plus optional scratch/checkpoint state, `roleLaunch` can launch fixed modeling/checkpoint and verification/review specialists over that same bounded seam, `roleResult` can read their structured outputs back without mutation, and `crrc` can recommend the next explicit action over that loop without doing the action itself. The scene, jobs, roles, role-result, freshness, context, pressure, reorient, crrc, reorient-result, and MVP status surfaces give clients compact or bounded views without making the GUI/client authoritative. The remaining missing organs are a second backend if `agent_jobs` stops being enough, automatic scheduling if dogfood proves it is needed, and richer product ergonomics after the MVP loop has been tested by a human.
|
||||
|
||||
```text
|
||||
model/tool observations
|
||||
|
|
@ -1385,7 +1386,9 @@ Current tests cover the landed flows at useful seams:
|
|||
- live app-server smoke for the first bounded reorient-launch runtime consumer plus CRRC coordinator recommendations in `tools/epiphany_phase6_reorient_launch_smoke.py`.
|
||||
- live app-server smoke for the first MVP operator status view in `tools/epiphany_mvp_status_smoke.py`.
|
||||
- focused app-server mapper coverage for `thread/epiphany/roles`, plus MVP status smoke coverage that consumes the roles surface.
|
||||
- focused app-server/protocol coverage for `thread/epiphany/roleLaunch` and `thread/epiphany/roleResult`, plus live role smoke coverage that launches modeling and verification workers and reads completed results without state mutation.
|
||||
- live app-server smoke for the first Phase 6 job-control boundary in `tools/epiphany_phase6_job_control_smoke.py`.
|
||||
- live app-server smoke for explicit Phase 6 role launch/read-back in `tools/epiphany_phase6_role_smoke.py`.
|
||||
|
||||
Useful commands:
|
||||
|
||||
|
|
|
|||
Reference in a new issue