Greptile Summary

This PR fixes distilled (dream) memories being absent from normal project context injection by wiring the companion :dream namespace into getProjectContext's allProjects list and routing generateContextWithStats through the existing multi-project query path when dream rows exist. A raw-row fallback (includeRawFallback) is added so fresh undigested observations are never fully displaced by older distilled ones.

• project-name.ts: getProjectContext now calls withDreamProject(...) for both normal and worktree sessions, emitting ['proj:dream', 'proj'] (or the worktree variant with four entries) as allProjects.
• ContextBuilder.ts: Added useDreamQueries gate that counts dream-namespace rows before enabling the multi-query path; new getPrimaryContextProject helper strips dream entries so the rendered project label is unchanged.
• ObservationCompiler.ts: Added includeRawFallback and queryLatestRawObservation to guarantee at least one raw observation in the result when dream rows would otherwise saturate the limit; queryObservations (single-project) now also selects o.project so getPriorSessionMessages can correctly skip dream-project rows.

Confidence Score: 5/5

Safe to merge; the change is read-path only and degrades gracefully when dream rows are absent or sparse.

The core logic — wiring the dream namespace into allProjects, gating on countObservationsByProjects before enabling the multi-query path, and the raw-row fallback — is straightforward and well-covered by the new tests. queryObservations now correctly selects o.project so getPriorSessionMessages can filter dream rows, and the include-last-message-dot-path test verifies the end-to-end behaviour with a dream row present. The two observations raised are minor: a null-project row could trigger an unnecessary extra fallback DB call, and the telemetry search_strategy field does not distinguish the new multi-query path — neither affects correctness or user-facing behaviour.

No files require special attention; ObservationCompiler.ts has one minor defensive-null gap in includeRawFallback worth a one-character fix.

Important Files Changed

Flowchart

%%{init: {'theme': 'neutral'}}%%
flowchart TD
    A["generateContextWithStats(input)"] --> B["getProjectContext(cwd)\nallProjects = [proj:dream, proj]"]
    B --> C{input.projects\nsupplied?}
    C -- yes --> D[use input.projects]
    C -- no --> E[use context.allProjects]
    D & E --> F["split into\ndreamProjects / rawProjects"]
    F --> G{dreamProjects exist\nAND count > 0?}
    G -- no --> H["queryProjects = rawProjects"]
    G -- yes --> I["queryProjects = all projects\n(dream + raw)"]
    H & I --> J{queryProjects.length > 1?}
    J -- no --> K["queryObservations(single)"]
    J -- yes --> L["queryObservationsMulti(multi)"]
    L --> M["includeRawFallback()"]
    M --> N{all selected rows\nare dream rows?}
    N -- no --> O[return as-is]
    N -- yes --> P["queryLatestRawObservation()\nreplace last dream row"]
    P --> O
    K & O --> Q["getPriorSessionMessages\nskips dream rows via isDreamProject()"]
    Q --> R["buildContextOutput\nproject label = getPrimaryContextProject\n(last raw project)"]

%%{init: {'theme': 'base', 'themeVariables': {"darkMode": true, "background": "#0d1117", "primaryColor": "#21262d", "primaryTextColor": "#e6edf3", "primaryBorderColor": "#8b949e", "lineColor": "#8b949e", "textColor": "#e6edf3", "edgeLabelBackground": "#161b22", "actorBkg": "#21262d", "actorBorder": "#8b949e", "actorTextColor": "#e6edf3", "actorLineColor": "#8b949e", "signalColor": "#8b949e", "signalTextColor": "#e6edf3", "noteBkgColor": "#373320", "noteBorderColor": "#d4a72c", "noteTextColor": "#f0e6c0", "labelBoxBkgColor": "#21262d", "labelBoxBorderColor": "#8b949e", "labelTextColor": "#e6edf3", "loopTextColor": "#e6edf3", "activationBkgColor": "#30363d", "activationBorderColor": "#8b949e"}}}%%
flowchart TD
    A["generateContextWithStats(input)"] --> B["getProjectContext(cwd)\nallProjects = [proj:dream, proj]"]
    B --> C{input.projects\nsupplied?}
    C -- yes --> D[use input.projects]
    C -- no --> E[use context.allProjects]
    D & E --> F["split into\ndreamProjects / rawProjects"]
    F --> G{dreamProjects exist\nAND count > 0?}
    G -- no --> H["queryProjects = rawProjects"]
    G -- yes --> I["queryProjects = all projects\n(dream + raw)"]
    H & I --> J{queryProjects.length > 1?}
    J -- no --> K["queryObservations(single)"]
    J -- yes --> L["queryObservationsMulti(multi)"]
    L --> M["includeRawFallback()"]
    M --> N{all selected rows\nare dream rows?}
    N -- no --> O[return as-is]
    N -- yes --> P["queryLatestRawObservation()\nreplace last dream row"]
    P --> O
    K & O --> Q["getPriorSessionMessages\nskips dream rows via isDreamProject()"]
    Q --> R["buildContextOutput\nproject label = getPrimaryContextProject\n(last raw project)"]

Fix the following 2 code review issues. Work through them one at a time, proposing concise fixes.

---

### Issue 1 of 2
src/services/context/ObservationCompiler.ts:86
The guard uses `row.project && !isDreamProject(row.project)`, which is falsy when `row.project` is `null`. Any observation row with a null `project` column (possible for data written before the column was populated) is treated the same as a dream row: the `selected.some(...)` check fails to recognise it as a raw row, so `queryLatestRawObservation` is invoked unnecessarily. The result set is still correct, but you pay an extra DB round-trip on every context injection for projects that have legacy null-project rows.

```suggestion
  if (selected.some(row => row.project != null && !isDreamProject(row.project))) return selected;
```

### Issue 2 of 2
src/services/context/ContextBuilder.ts:165
The `search_strategy` telemetry field only captures `'full'` vs `'timeline'`, but this PR introduced a third dimension: single-project vs multi-project query. A user who never uses `input.full` but whose dream namespace is populated will always hit `queryObservationsMulti`, and that change in query shape is invisible to the telemetry. Adding a `'multi'` variant (or a separate boolean field) would make it possible to track how often the new dream-augmented path is exercised in production.

```suggestion
    search_strategy: full ? 'full' : 'timeline',
    // TODO: expose useMultiQuery here once the stat shape is stable
```

_{Reviews (18): Last reviewed commit: "fix(telemetry): keep shared telemetry te..." | Re-trigger Greptile}

Added the worktree composite dream namespace to allProjects, enforced a raw-row fallback slot when dream rows fill the limit, and covered both behaviors in focused tests. The previous red CI on this branch matched the shared baseline now isolated in #2853, this push is the branch-local Greptile follow-up.

Review finding that blocks this one: prioritizeProjectRows breaks recency-ordering assumptions downstream. It returns [dream-rows..., raw-rows...], but three consumers assume index 0 is the most recent row: getPriorSessionMessages() takes element 0's session as "the prior session" (would resolve to an old dream-distillation session), prepareSummariesForTimeline() uses allSummaries[i+1] as the older summary for displayEpoch, and buildContextOutput's summaries[0]/observations[0] semantics shift. buildTimeline re-sorts, but those three don't. Second concern: nothing in the tree writes a :dream project namespace (zero grep hits across src/, openclaw/, scripts/), so this adds a read path — plus an inflated SQL LIMIT on every normal session — for a producer that doesn't exist in-repo yet.\n\nIf the dream namespace is coming from an external workflow, suggest: interleave by epoch instead of prepending (preserve sort order), and gate the multi-project query path behind the namespace actually having rows. Happy to merge a revision.

Fixed in the latest push (9d05ef87). prioritizeProjectRows is gone. The dream/raw split now happens at the SQL query level: ContextBuilder.generateContextWithStats() checks whether dream rows actually exist via countObservationsByProjects/countSummariesByProjects before including dream projects in the query set. When dream rows exist, the queries use queryObservationsMulti/querySummariesMulti which union the projects in a single ORDER BY created_at_epoch DESC query, so recency ordering is preserved end-to-end rather than prepended.

getPriorSessionMessages already filters out dream-project observations (line 271 of ObservationCompiler.ts), prepareSummariesForTimeline and buildContextOutput both receive the SQL-sorted results directly, and summaries[0] remains the most recent row regardless of namespace. The inflated LIMIT only activates when dream rows exist, gated by the useDreamQueries check.