Tune role-specific self-improvement prompts

This commit is contained in:
Meta Vens 2026-05-04 23:08:54 +01:00
commit 1a2f40f2ec
7 changed files with 46 additions and 11 deletions

View file

@ -39,7 +39,7 @@ Do not trust this file for the exact live HEAD. Always check git.
- Internal `agent_job:` workers now get the reporting tool independent of the user-facing CSV spawn feature, and ephemeral worker sessions can initialize the sqlite state runtime on demand so specialist reports land in the shared backend.
- Durable in-flight investigation checkpointing is now landed in authoritative typed state, writable through `thread/epiphany/update` or accepted `thread/epiphany/promote`, rendered into the prompt, and reflected through scene/context.
- The prompt doctrine pass is landed. Shared Epiphany prompts now carry distilled memory/evidence discipline. Rendered state intro/doctrine text lives in `epiphany-core/src/prompts/`, and lane/control prompt text lives in `vendor/codex/codex-rs/app-server/src/prompts/epiphany_specialists.toml`: modeling is the Body, implementation is the Hands and GUI-launched main coding lane, verification is the Soul, reorientation is Life, coordinator remains the read-only Self, and CRRC owns the pre-compaction intervention template.
- The Ghostlight memory pass is landed. `epiphany_specialists.toml` now has a shared persistent-memory projection prepended by the harness to fixed role specialists, reorientation workers, coordinator notes, and CRRC checkpoint interventions; `tools/epiphany_gui_action.py` prepends the same memory to GUI-launched implementation prompts. The rendered base doctrine also states the Perfect Machine rule directly: prompt is projection, durable typed state is the mind, and every lane must improve its own memory/model/prompt/evidence habit or name the repair.
- The Ghostlight memory pass is landed. `epiphany_specialists.toml` now has a shared persistent-memory projection prepended by the harness to fixed role specialists, reorientation workers, coordinator notes, and CRRC checkpoint interventions; `tools/epiphany_gui_action.py` prepends the same memory to GUI-launched implementation prompts. The rendered base doctrine also states the Perfect Machine rule directly: prompt is projection, durable typed state is the mind, every lane must improve its own memory/model/prompt/evidence habit or name the repair, and each lane phrases that duty in its own organ language so the salience sticks.
- The Aetheria dogfood run has a contamination scar. The supervising Codex session directly edited and committed target-repo work on `E:\Projects\Aetheria-Economy` instead of only driving Epiphany lanes. Treat those Aetheria commits as supervisor-seeded implementation, not clean evidence that Epiphany coordinated the work. Future dogfood must run through the GUI/coordinator/fixed role lanes with auditable artifacts unless the user explicitly authorizes an operator intervention.
- The dogfood quarantine now has a direct-thought boundary. The supervisor may read coordinator actions, role/reorient statuses, structured finding summaries, reviewed state patches, rendered status snapshots, and artifact manifests. It must not read raw worker transcripts, direct worker messages, full turn logs, or `rawResult` payloads during normal dogfood. Those artifacts are sealed for explicit forensic debugging only.
- `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.