# Telegram Dropbear Context Engineering RCA Work Order Date: 2026-07-01 Observed run: `/home/roko/Documents/Projects/Adjacent/telegram_test/.omnius`, run id `1782873796963-i5r7mv`. Task under observation: get a meaningful gear-sonic training run with observed progress on the Dropbear MJCF/URDF after addressing closed-loop knee components. ## Evidence Snapshot - The run made one successful mutation: `dropbear_mjcf/dropbear_mjcf.xml` changed `RL_Revolute28` to ``. - Immediately afterward, the model retried the pre-mutation `old_string` and received repeated `old_string not found` failures. - The focus supervisor then forced `required_next_action=update_todos`, which blocked useful recovery reads, patches, and verification attempts. - The run kept emitting blocked edits, sed/python rewrites, a partial `task_complete`, and more blocked reads through at least turn 28. - Main context grew from about 73k estimated tokens after the successful edit to about 92.7k estimated tokens by turn 28. - System prompt content grew from about 142k chars to about 230k chars. - The final main dump had dozens of `[RUN EVIDENCE]` and focus-supervisor system messages, including repeated synthetic failures. - The workboard existed but had no cards, decisions, or diagnostics. - The completion ledger remained `open`; no meaningful gear-sonic training verification was run. ## Work Orders ### WO-TD-01: Stale Edit Recovery Must Read Current Truth, Not Replan Forever Status: implemented Root cause: repeated stale edit failures currently force `update_todos` for non-shell tools. For an `old_string not found` family, this is the wrong recovery state. Anchors: - `packages/orchestrator/src/focusSupervisor.ts` - `packages/orchestrator/src/agenticRunner.ts` - `packages/orchestrator/tests/focusSupervisor.test.ts` Implementation: - Classify stale edit failure samples in the focus supervisor. - Repeated stale `file_edit`, `file_patch`, or `batch_edit` failures must set `requiredNextAction=read_authoritative_target`. - After the authoritative read succeeds, clear that directive and permit a fresh target-specific edit or verification. - Do not use `update_todos` as the default escape hatch for stale edit families. Acceptance: - Two `old_string not found` failures produce a `read_authoritative_target` directive. - A `file_read` of the target satisfies the directive. - A different edit target in the same file is not blocked just because an old edit target failed. Verification: - Covered by `tests/focusSupervisor.test.ts`. ### WO-TD-02: Directive Escalation Must Terminate Cleanly After Repeated Ignoring Status: implemented Root cause: the observed run ignored 27 focus directives and kept accumulating blocked tool calls. The supervisor continued asking for the same action without forcing a terminal incomplete/blocker path. Anchors: - `packages/orchestrator/src/focusSupervisor.ts` - `packages/orchestrator/src/agenticRunner.ts` - `packages/orchestrator/tests/focusSupervisor.test.ts` Implementation: - Add an ignored-directive escalation threshold. - Once the threshold is crossed with no satisfying mutation/read/verification, force `requiredNextAction=report_incomplete`. - In terminal incomplete state, permit `task_complete` only as an incomplete/blocker report. - Stop appending broad blocked-family lists indefinitely. Acceptance: - Repeatedly ignoring a recovery directive transitions to `terminal_incomplete`. - After terminal incomplete, edit/read/shell variants are blocked. - A reporting tool is allowed so the run can end truthfully. Verification: - Covered by `tests/focusSupervisor.test.ts`. ### WO-TD-03: Focus Action Families Must Be Target-Specific For Edits Status: implemented Root cause: focus supervisor action families collapse all edits to `tool:path`. In the observed run, a stale `RL_Revolute28` edit contaminated later distinct joint edits and line patches in the same XML. Anchors: - `packages/orchestrator/src/focusSupervisor.ts` - `packages/orchestrator/src/agenticRunner.ts` - `packages/orchestrator/tests/focusSupervisor.test.ts` Implementation: - Build edit action families from tool name, path, normalized edit target, patch mode/offset/limit, and error class where available. - Keep shell action families command-based, but trim and hash long commands rather than injecting huge command prefixes. - Preserve a separate coarse path lock only for full-file overwrite repair blocks. Acceptance: - A repeated stale edit for `RL_Revolute28` blocks only that edit family. - A fresh `file_patch` for a different offset/target in the same XML can pass after the required authoritative read. - The directive frame no longer grows with long sed/python command bodies. Verification: - Covered by `tests/focusSupervisor.test.ts`. ### WO-TD-04: Context Evidence Must Fold Repeats And Demote Synthetic Failures Status: implemented Root cause: the context engine injects every recorded tool event as high-authority `[RUN EVIDENCE]`, including repeated focus blocks and stale discovery reads. Anchors: - `packages/orchestrator/src/contextEngine.ts` - `packages/orchestrator/src/agenticRunner.ts` - `packages/orchestrator/tests/contextEngine.test.ts` Implementation: - Fold tool evidence by family before rendering system evidence. - Keep successes, latest mutations, latest verification results, and first/last unresolved failures. - Collapse repeated focus-supervisor synthetic failures into one counted summary. - Cap evidence injection separately from raw conversation/tool results. Acceptance: - Repeated identical focus blocks render as one summary with a repeat count. - Evidence diagnostics report retained and dropped/folded counts. - Large early discovery listings do not stay as repeated high-authority system evidence once more specific evidence exists. Verification: - Covered by `tests/contextEngine.test.ts`. ### WO-TD-05: Context Engine Compaction Must Affect Live Requests Status: implemented Root cause: `AgenticRunner` calls `contextEngine.compact()` but discards its output, while `contextEngine.build()` still prepends extra evidence to the live request. Anchors: - `packages/orchestrator/src/agenticRunner.ts` - `packages/orchestrator/src/contextEngine.ts` - `packages/orchestrator/tests/agenticRunner-context-behavior.test.ts` - `packages/orchestrator/tests/contextEngine.test.ts` Implementation: - Apply context-engine compaction output to the live message list when it actually compacts. - Avoid duplicating engine-added evidence already present in the active context. - Carry latest SNR/evidence metrics into focus-supervisor proposed-call snapshots. Acceptance: - A compacted context-engine output reduces the live message list before backend calls. - `_buildFocusContextSnapshot()` includes raw discovery chars, active evidence chars, active frame chars, and signal-to-noise ratio. - Context dumps and focus decisions see the same pressure/noise picture. Verification: - Covered by `tests/contextEngine.test.ts`, `tests/agenticRunner-context-behavior.test.ts`, and package build. ### WO-TD-06: Completion Ledger Must Finalize Open Failure Runs Truthfully Status: implemented Root cause: the ledger contained evidence of mutation, stale edit failures, and no final verification, but stayed `open`. Anchors: - `packages/orchestrator/src/completionLedger.ts` - `packages/orchestrator/src/agenticRunner.ts` - `packages/orchestrator/tests/completionLedger.test.ts` Implementation: - Finalize open ledgers using evidence when a run exits or hits a terminal incomplete state. - Treat unresolved stale edit families, mutations after last verification, and missing verification after code mutations as incomplete verification. - Record focus terminal-incomplete events as unresolved ledger evidence. Acceptance: - A ledger with an unverified code mutation finalizes as `incomplete_verification`. - A ledger with unresolved stale edit blocks finalizes as `incomplete_verification`. - A run that reaches terminal incomplete writes the ledger status before exit. Verification: - Covered by `tests/completionLedger.test.ts`; runner bridge typechecked by package build. ### WO-TD-07: Complex Tasks Must Seed A Durable Workboard Or Todo Tree Status: implemented Root cause: this robotics/training task had an empty workboard despite requiring discovery, model repair, integration, training, and verification. Anchors: - `packages/orchestrator/src/agenticRunner.ts` - `packages/execution/src/tools/workboard.ts` - `packages/execution/tests/workboard.test.ts` - `packages/orchestrator/tests/agenticRunner-context-behavior.test.ts` Implementation: - Detect complex implementation/training tasks and seed a minimal board or nested todo skeleton before freeform tool use. - Require cards or nested todos for repair, integration, training run, and observed verification. - Attach tool evidence to active cards so completion review has structured state. Acceptance: - A multi-step training/robotics task starts with non-empty durable decomposition. - Workboard cards or nested todos update as evidence is collected. - Completion review sees unresolved cards/blockers when verification has not happened. Verification: - Runner seed-card path guarded by `tests/agenticRunner-context-behavior.test.ts`. ## Verification Checklist - [x] `pnpm --filter @omnius/orchestrator test -- tests/focusSupervisor.test.ts` - [x] `pnpm --filter @omnius/orchestrator test -- tests/contextEngine.test.ts` - [x] `pnpm --filter @omnius/orchestrator test -- tests/completionLedger.test.ts` - [x] `pnpm --filter @omnius/orchestrator test -- tests/agenticRunner-context-behavior.test.ts` - [x] `pnpm --filter @omnius/orchestrator build`