Keep regather in fixed coordinator lanes
This commit is contained in:
Meta Vens
parent
cda1b26416
commit
c6201363f4
7 changed files with 271 additions and 39 deletions
|
|
@ -35,7 +35,7 @@ The spine exists in twenty-five live paths:
|
|||
- 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, recognizes already accepted reorientation findings so an accepted result is not recommended for acceptance again, 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/acceptance: app-server declares experimental `thread/epiphany/roleLaunch`, read-only `thread/epiphany/roleResult`, and write-gated `thread/epiphany/roleAccept`. `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, optional modeling `statePatch`, raw result, and errors without mutation, promotion, scheduling, or hidden continuation. `roleAccept` currently accepts only completed modeling/checkpoint findings with a valid modeling-safe `statePatch`, applies it through the existing Epiphany state validator, appends audit observation/evidence, and emits `stateUpdated` with source `roleAccept`. Implementation stays owned by the main coding agent, verification output remains review-only, and reorientation stays on the CRRC-specific `reorientLaunch` / `reorientResult` / `reorientAccept` path.
|
||||
- fixed-lane coordinator: app-server declares experimental read-only `thread/epiphany/coordinator`, composes role, pressure, reorient, CRRC, role result, and reorient result signals, and returns one current action without mutation. Review acceptance and implementation clearance are intentionally different: an accepted verification finding clears implementation only when the verifier verdict is `pass`; accepted `needs-evidence`, `needs-review`, or `fail` findings route back toward modeling/checkpoint strengthening before code work resumes.
|
||||
- fixed-lane coordinator: app-server declares experimental read-only `thread/epiphany/coordinator`, composes role, pressure, reorient, CRRC, role result, and reorient result signals, and returns one current action without mutation. Review acceptance and implementation clearance are intentionally different: an accepted verification finding clears implementation only when the verifier verdict is `pass`; accepted `needs-evidence`, `needs-review`, or `fail` findings route back toward modeling/checkpoint strengthening before code work resumes. `regatherManually` is now fallback, not a supervisor-puppeteering excuse: completed unaccepted modeling findings stop at `reviewModelingResult`, modeling acceptance precedes verification, and fixed lane actions are considered before manual regather.
|
||||
- MVP operator view: `tools/epiphany_mvp_status.py` launches app-server over stdio, starts or reads one thread, gathers scene, pressure, reorient, jobs, roles, modeling/verification role results, reorient result, and CRRC responses, and renders them as text or JSON so the loop can be dogfooded without reading Rust.
|
||||
- MVP dogfood runner: `tools/epiphany_mvp_dogfood.py` launches app-server over stdio, drives a bounded state/role/CRRC/reorientation loop, and writes operator-safe JSON/text artifacts plus sealed transcript diagnostics under `.epiphany-dogfood/mvp-loop` so a human can audit the run through projections and compare it against a vanilla Codex reference.
|
||||
- Agent function/API telemetry: `tools/epiphany_agent_telemetry.py` mechanically parses sealed AppServerClient JSONL transcripts into operator-safe telemetry with method counts, request shape, job/status/path fields, and visible function/tool names while sealing text, turn items, raw results, and direct agent messages.
|
||||
|
|
@ -1448,6 +1448,7 @@ stored Epiphany state
|
|||
-> fixed modeling/checkpoint and verification/review role workers can be launched explicitly
|
||||
-> fixed role worker results can be read back for review without mutation
|
||||
-> reviewed modeling/checkpoint statePatch findings can be accepted explicitly to grow graph/checkpoint state
|
||||
-> completed unaccepted modeling findings stop for review before verification
|
||||
-> accepted verifier findings clear implementation only on verifier verdict pass; non-pass findings send the loop back toward modeling/checkpoint repair
|
||||
-> can be consumed explicitly by one bounded reorient-worker launch surface
|
||||
-> retrieval/indexing can inform future work
|
||||
|
|
|
|||
|
|
@ -44,8 +44,9 @@ Do not trust this file for the exact live HEAD. Always check git.
|
|||
- `tools/epiphany_agent_telemetry.py` is the safe instrument panel for sealed runs. Status/coordinator/GUI/dogfood/live-specialist tools generate telemetry JSON from sealed transcripts that preserves method names, call shape, job/status/path counts, and any visible function/tool names while sealing text, direct messages, and raw results.
|
||||
- Latest Aetheria supervised dogfood thread: `019ddc52-4ee8-7203-b6c0-106a9c270067` with codex-home `.epiphany-dogfood/aetheria-supervised/codex-home`. CRRC compact/reorient/reorientAccept reached Epiphany state revision 16 after rollout replay fixes. Later sanity review showed the modeler and verifier were doing the right work: modeling emitted source-grounded graph/checkpoint/frontier artifacts, while verification returned `needs-evidence` and explicitly did not bless implementation.
|
||||
- Coordinator policy now treats accepted verifier results as implementation clearance only when the verifier verdict is `pass`. Accepting a `needs-evidence`, `needs-review`, or `fail` verifier finding records the review but routes the loop back toward modeling/checkpoint strengthening or reorientation. The policy was tightened after a too-permissive coordinator path treated reviewed verifier output as a green light.
|
||||
- A later dogfood attempt exposed another supervision scar: manually launching modeling, reorientation, and verification from the supervisor is not clean dogfood. The coordinator now treats `regatherManually` as fallback only after fixed lanes cannot advance, stops at `reviewModelingResult` for completed unaccepted modeling findings, and the CLI runner's explicit test-only `--auto-review` mode can accept modeling `statePatch` findings before launching verification. Production semantic findings remain review-gated.
|
||||
- The implementation lane is the current product blocker, but it is now harness-visible. The latest audited artifact is `.epiphany-dogfood/aetheria-supervised/gui-actions/continueImplementation-1777528376766401900-25000`: it shows `workspaceChanged: false`, wrote `implementation-audit.md`, applied a no-diff `thread/epiphany/update`, advanced Epiphany state to revision 17, and marked the investigation checkpoint `regather_required`. The next clean run should retry through the operator-safe coordinator lanes after the pre-compaction handoff repair, not by reading sealed worker thoughts.
|
||||
- Follow-up coordinator plans at `.epiphany-dogfood/aetheria-supervised/coordinator-after-nodiff-state-20260430-plan` and `.epiphany-dogfood/aetheria-supervised/coordinator-after-verification-verdict-policy-20260430-plan` return `regatherManually` with implementation blocked, instead of continuing fake progress.
|
||||
- Follow-up coordinator plans at `.epiphany-dogfood/aetheria-supervised/coordinator-after-nodiff-state-20260430-plan` and `.epiphany-dogfood/aetheria-supervised/coordinator-after-verification-verdict-policy-20260430-plan` returned `regatherManually` with implementation blocked. After the latest coordinator fix, `.epiphany-dogfood/aetheria-supervised/supervisor-runs/dogfood-20260430-094721/coordinator-post-puppet-fix-plan` returns `reviewModelingResult`: the harness now surfaces the unaccepted modeling finding instead of asking the supervisor to improvise.
|
||||
- The GUI parses `implementation-result.json` into artifact metadata, surfaces the latest implementation diff/no-diff outcome, and pauses immediate `Continue Implementation` repeats when the newest artifact is a no-diff implementation audit.
|
||||
- The repo is an Epiphany fork of Codex, not a Codex preset.
|
||||
- `vendor/codex` is tracked directly, not a submodule.
|
||||
|
|
@ -359,11 +360,14 @@ bounded browser-fallback controls. A live bridge probe also proved
|
|||
The next real move is not more direct Aetheria implementation. The clean
|
||||
supervised lane reached implementation too early because the coordinator treated
|
||||
an accepted verifier finding as a green light even when the verifier verdict was
|
||||
`needs-evidence`. That policy is now tightened: only an accepted verifier
|
||||
`pass` clears implementation. The current Aetheria thread should therefore
|
||||
start with bounded regather/modeling repair for the `regather_required`
|
||||
checkpoint and verifier evidence gaps, then run verification again before any
|
||||
implementation retry. Read only operator-safe projections; do not
|
||||
`needs-evidence`. The next dogfood attempt then proved that supervisor-clicked
|
||||
lane launches are theatre, not harness autonomy. Both policies are now
|
||||
tightened: only an accepted verifier `pass` clears implementation, and
|
||||
`regatherManually` no longer preempts fixed modeling/verification review lanes.
|
||||
The current Aetheria thread should stop at `reviewModelingResult` because the
|
||||
latest modeler produced a useful source-grounded finding but no reviewable
|
||||
`statePatch`; fix that modeling output contract before another implementation
|
||||
retry. Read only operator-safe projections; do not
|
||||
open raw worker transcripts, direct
|
||||
worker messages, or `rawResult` payloads unless the user explicitly asks for
|
||||
forensic debugging. Use generated function/API telemetry for call-shape
|
||||
|
|
|
|||
Reference in a new issue