Conversation
…ation.md) Red-first probes A-p1..A-p5 for the v1.0.15 graph-revision view ledger (workflows/dag-engine-optimization.md, Train A). View-layer semantics only: durable data untouched, replan/cancel engine unchanged. RED on the unmodified engine (documented failing assertions): - A-p1/A-p3: replaced failed+cancelled nodes still counted in current-rev summaries (nodeCount 8 vs 6) and re-seeded as required-unsatisfied on rebuild — the wake-up bug chain this train breaks. - A-p5(a): wake digest failure attribution still names superseded replaced failures. - A-p1 status seam: workflow status action surfaces superseded rows. PINS (green before and after): A-p2 current-rev genuine failure visibility, A-p4 legacy no-rev rendering, A-p5(b) reopen wake wash by status.
…revision (dag-engine-optimization.md) Train A implementation for the v1.0.15 rev-view ledger (workflows/dag-engine-optimization.md): durable data untouched, engine replan/cancel semantics unchanged — view-layer wash only. Root cause of the wake-up bug: replaced segment rows (cancelled via replan; terminal failures the fragment bypassed) re-seeded as required-unsatisfied on every runtime rebuild, welding workflows to failure despite the new path succeeding. Mechanism (evidence M1 design, red-first probes landed in d33d4f4): - workflow.graph_rev (audit counter) + workflow_node.superseded (view predicate); migration defaults graph_rev=1 / superseded=false keep every legacy workflow rendering exactly as v1.0.14 (A-p4). - Projector marks superseded on NodeCancelled (+ explicit cancel, U1) and on the new optional WorkflowReplanned.superseded list; graph_rev bumps once per replanning event (leg order guards the double-bump). - Dag._replan publishes the supersede list: terminal-FAILED rows at replan time are the segment the new revision replaces. - DagStore.getCurrentNodes + summary SQL filter; switched view/aggregation seams: loop rebuild/recovery/checkCompletion/wake attribution, step-mode readiness, workflow status action, TUI dag.nodes route. - Engine reads stay unfiltered on purpose: _replan plan input, terminate cleanup, review gate/extend, spawnReady input mapping (completed old-rev outputs survive for A1/A2), result store, recovery reconcile. No HTTP response-shape change => no SDK regen; TUI renders the filtered server seams verbatim (zero client changes, server-side aggregation invariant). Mutation re-proof: four targeted mutations each red the probes; A-p2 pin green throughout. Gates: opencode/core/tui/schema typecheck, core migration --check, root lint 4849<=4850, dag suites green (opencode 725, core dag 94).
…ization.md) Train B probes for the v1.0.15 node-output file-reference ledger (B1-B4). RED on the unmodified engine (5 fail / 8 pass): B-p1(a) and B-p1(a2) fail because non-schema nodes never write captured_output (no submit-time absolute-path detection); B-p3 fails because the workflow result action pages the raw output string and never surfaces content_ref/path/summary; B-p4(a)/(b) fail because nothing touches the project .gitignore on first write into the report area. Guard and pin probes are green before and must stay green after: nonexistent / empty / directory / prose-embedded paths stay inline (B-p1 guards), output_schema payloads containing a path string stay inline JSON (B-p2 dual-track pin), the legacy paged result read is unchanged (B-p3 pin), and refs outside the report area never touch .gitignore (B-p4(c)). The run's evidence.md was absent from the worktree, the config workflow repo, and the opencode data dir at train start — the probe contract is derived from the settled ledger in workflows/dag-engine-optimization.md; the report area convention is .opencode/workflow-reports/ (no better-fit existing .opencode/ directory; naming mirrors .opencode/workflow-drafts/).
…gine-optimization.md) Train B implementation for the v1.0.15 node-output file-reference ledger B1-B4 (red-first probes landed in a1dbaf8; all 13 probes + 6 post-feature units green). Mechanism (dual track, B1): - B2 submit-time detection — new runtime/output-ref.ts. When a non-schema node's final reply IS one absolute path (single token, no prose) and the file exists as a non-empty regular file, capture {kind: file_ref, content_ref, path, size, sha256, summary} into the existing captured_output column — the same durable seam submit_result writes and NodeStarted resets. Any anomaly (relative path, missing / empty file, directory, prose around the path, fs error) keeps the exact legacy inline settlement; the capture is audit metadata and never fails the node. Integrity receipt: sha256 + size at submit time; summary is the first ~200 chars captured at submit time so it stays stable even if the file drifts later. - B3 result action — a captured file ref reads as a durable pointer (content_ref + summary + path + size + sha256, truncated:false, no cursor); the parent agent fetches content itself with the read tool. Inline outputs and output_schema payloads keep the paged read byte-identical — no migration, no shape change for them. The node output column stays the raw path string, so input mapping and wake digests behave exactly as before. - output_schema nodes are untouched (B1): detection runs only in the free-text completion branch; schema settlement keeps inline JSON even when a payload field contains a path (B-p2 pin). - B4 report-area gitignore — first capture into <directory>/.opencode/workflow-reports/ appends the area entry to <directory>/.gitignore (append-only via appendFile, idempotent, skips when the entry or a covering .opencode/ rule exists, best- effort). Cross-worktree refs never touch a project gitignore. The report area convention is .opencode/workflow-reports/ — no better-fit existing .opencode/ directory; naming mirrors the existing .opencode/workflow-drafts/ area. No HTTP response-shape change => no SDK regen; TUI untouched (result view is tool-served, captured_output is not a TUI surface). Crash window between capture and NodeCompleted persists: recovery completes such a node with undefined output while the result seam still serves the durable ref. Mutation re-proof (each reverted, then restored): capture-wiring removal reds B-p1(a)/(a2)+B-p4(a)/(b); result-branch removal reds B-p3 alone; gitignore-call removal reds B-p4(a)/(b) alone with B-p1 still green. Gates: packages/opencode bun typecheck + bun test test/dag test/goal test/project = 744 pass/0 fail/1 skip (62 files). packages/core bun typecheck + bun run migration --check (no schema changes) + license- scope check + bun test test/dag = 94 pass/0 fail. packages/tui bun typecheck + bun test = 244 tests/0 fail. packages/schema typecheck + 16 tests/0 fail. ROOT bun run lint = 4849 warnings (unchanged from baseline), 0 errors.
feat(dag): current-revision view and node-output file refs (v1.0.15 ledger)
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Summary
Promote to main for the next official release. One train merged on dev since v1.0.14 (#274):
feat(dag): current-revision view and node-output file refs (PR #275, implements the settled ledger in
workflows/dag-engine-optimization.md):graph_rev; replan/extend supersedes the replaced segment and the view (status output, httpapi nodes, TUI summaries, fail counts) renders exactly the current revision — historical replaced segments including their old failures are hidden; live current-rev failures (token/API/cap unreachable) stay visible and counted; terminal aggregation no longer counts superseded failures (the old rebuild bug re-seeded replaced nodes as required-unsatisfied and failed workflows); a current-rev tail failure still fails the workflow visibly. Durable rows, event sourcing, completed outputs, and engine replan semantics unchanged — audit stays agent-reachable by id, TUI zero entry. Legacy workflows render byte-identical (migration defaults, pinned).{content_ref, size, sha256, summary}at submit time (anomalies fall back to inline, never fail the node); the result action returns the pointer + summary; inline and output_schema payloads unchanged; report writes under.opencode/workflow-reports/auto-ensure an append-only.gitignoreentry.Dev validation
Release plan
On merge: dispatch release-fork from main (official). Release notes carry the two feature groups above.