ceos/opencode
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

feat(core): make path-local instruction discovery durable

Browse source 
Replace synthetic-message instruction injection with a durable
discovery projection so discovered AGENTS.md files survive compaction,
forks, reverts, and restarts.

- Add the instruction_file table and the durable
  session.instructions.discovered event; projection stores each
  discovered path and content with its owning assistant-message
  boundary and durable discovery order.
- Fold discovered files into the core/instructions source through
  InstructionDiscovery, absorbing SessionInstructions. Completed
  compaction rebaselines restate discovered instructions instead of
  summarizing them away.
- Re-read discovered files live at each observation so mid-session
  edits reach the model; the frozen discovery content stands in only
  when a file becomes unreadable.
- Narrate instruction file changes as per-file deltas via diffByKey,
  restating the full set only for pure reorderings.
- Fork copies discoveries within the inherited transcript, committed
  revert removes discoveries past the revert boundary, and Session
  movement clears them so the destination initializes a complete
  baseline.
This commit is contained in:
Kit Langton 2026-07-06 14:22:53 -04:00
commit 0726f09b05
28 changed files with 1166 additions and 598 deletions
Download patch file Download diff file Expand all files Collapse all files

2
AGENTS.md
View file

@ -161,5 +161,5 @@ const table = sqliteTable("session", {
- Keep delivery vocabulary explicit. Prompts steer by default and promote at the next safe step boundary while the current drain requires continuation. An explicit `queue` input remains pending until the Session would otherwise become idle; promote one queued input at that boundary, then reevaluate continuation before promoting another. Promoting any new user input resets the selected agent's step allowance; a batch of steers resets it once.
- One step is one logical LLM call; its durable record covers only the model-visible span. Do not write "provider turn", and do not use bare "turn" for a single call: "turn" is reserved for the future assistant-turn unit containing all steps from prompt promotion until the session would go idle.
- Keep EventV2 replay owner claims separate from clustered Session execution ownership.
- Keep the Instructions algebra and built-ins in `src/instructions`; keep instruction producers with their observed domains, and keep Session History selection plus `InstructionCheckpoint` and `InstructionEntry` persistence Session-owned. `InstructionDiscovery` observes ambient global and upward-project instructions. The runner composes built-ins, discovery, guidance, and entries explicitly in `loadInstructions`; there is no instruction registry.
- Keep the Instructions algebra and built-ins in `src/instructions`; keep instruction producers with their observed domains, and keep Session History selection plus `InstructionCheckpoint` and `InstructionEntry` persistence Session-owned. `InstructionDiscovery` combines ambient global/upward-project instructions with durable path-local instructions discovered after successful reads. The runner composes built-ins, discovery, guidance, and entries explicitly in `loadInstructions`; there is no instruction registry.
- The durable `Instructions.Applied` record is what the model was last told, per instruction source. Reconciliation narrates drift through `session.instructions.updated` and never rewrites the baseline; only completed compaction rebaselines, while Session movement or committed revert resets the `InstructionCheckpoint`. Unavailable sources keep the model's prior belief, blocking only a Session's first instruction baseline.