Agent Work

Agent Works

A Work is a durable record of something an agent produced or touched — a GitHub PR/issue, a Linear issue, an entity file (pptx/xlsx/docx/pdf), a document, a task. Works render as cards under the assistant message that produced them and accumulate versions across operations, so the same PR edited twice shows one card with history.

Two tables (packages/database/src/schemas/work.ts):

• works — one row per resource. Identity/dedup key: (resourceType, resourceId) within the user/workspace scope. currentVersionId soft-references the latest version.
• work_versions — one row per registration event. Dedup: unique (workId, toolCallId), so a retried registration with the same real tool call id is a no-op. Versions carry provenance (source message, producing tool) and operation-level cumulativeCost / cumulativeUsage.

Type registry

packages/database/src/models/work/registry.ts is the single registry of Work types (document / external / file / task): WORK_TYPE_ADAPTERS is satisfies Record<WorkType, WorkTypeAdapter>, so a type added to @lobechat/types without an adapter is a compile error.

Read-path compatibility gate: OPT_IN_WORK_TYPES (currently {'file'}) hides newer types from clients that did not opt in (includeFileWorks). Released Electron clients lag by weeks and crash on unknown type descriptors (descriptor.getIcon on undefined), so a request without the opt-in receives exactly the pre-file set. Any NEW Work type must ship behind the same kind of opt-in.

Registration write paths

Four write paths, two timing classes:

Execution-time: skill providers

SKILL_TOOL_RESULT_NORMALIZERS in packages/database/src/models/work/index.ts maps WorkSkillProvider (github, linear; vocabulary in packages/types/src/work.ts) to a normalizer. Both satisfies a Record<WorkSkillProvider, …>, so provider list and normalizer map cannot drift.

A normalizer turns one tool result into ExternalToolWorkOperation | null — null means "not Work-worthy", which is the common case and never an error.

Completion-time: the shell Work scan framework

Heterogeneous CLI agents (codex, claude-code) and the device lobe-local-system tool run CLIs like gh through their own shell surfaces, which never pass the skill-tool hook. registerWorksForOperation recovers their Works at completion from the persisted command text + stdout. Layering:

registerWorksForOperation (registerWorksForOperation.ts)
  ├─ collectOperationRecords         one pass over the operation tree, shared by both scans
  ├─ registerShellWorks              (shellWorkRegistration.ts — the ENGINE, command-agnostic)
  │    ├─ SHELL_COMMAND_SOURCES      identifier→apiName scoping of shell surfaces
  │    ├─ success gate               plugin error / state.success===false / state.error → skip
  │    ├─ extract {command, exitCode, output}   from arguments / state / message content
  │    └─ shellWorkScanners/         one file per CLI family
  │         ├─ types.ts              ShellWorkScanner = { matches, name, register }
  │         ├─ github.ts             matches: includes('gh ') → workModel.registerShellGithubResult
  │         └─ index.ts              SHELL_WORK_SCANNERS registry
  └─ file Work scan                  aggregate per-path fold + sandbox export pipeline (NOT a scanner)

Command parsing is split the same way: packages/database/src/models/work/shellCommandParsing.ts holds the command-agnostic layer (POSIX-ish tokenizer, control-operator segmenting, codex /bin/zsh -lc wrapper expansion — parseShellCommandSegments), while githubToolResult.ts holds only the gh-specific parseGhSegment + field mapping.

Cost granularity: execution-time vs completion-time

cumulativeCost / cumulativeUsage are per-version SNAPSHOTS ("spend up to this version"), never deltas — summing them across versions double-counts on every path.

• Execution-time registrations (skill providers) snapshot right after UsageCounter.accumulateTool for that call, so create + edit + edit within one run carries an INCREASING series (e.g. $0.30 → $0.70 → $1.20) and step deltas are recoverable.
• Completion-time registrations (shell scan, file scan) attach one OPERATION-LEVEL figure — the completing run's terminal total plus terminal child-op totals — to EVERY version registered that round: the scan only has persisted command text + stdout, no per-call intermediate snapshots (hetero CLI usage is aggregated per step by the trace recorder, not per tool call). The same create + edit + edit therefore yields three versions all carrying $1.20, indistinguishable per step.
• Cross-operation is independent: a later run editing the same resource appends a version with THAT run's total only.

Why file Works are not a scanner: their unit is a file path folded across ALL records (multi-edit merge, renames, last-edit provenance), the registration is a heavy IO pipeline (idempotency probe → sandbox export → upload → version → redeploy, bounded concurrency), hetero records are deliberately EXCLUDED (the file lives on the executing device, not the exportable sandbox), and the version key is the synthetic op:${operationId} instead of a real toolCallId. Shoehorning that into ShellWorkScanner would break the interface. If a second aggregate-style scan ever appears, extract an aggregate layer then.

Display anchor & completion marker

A Works card renders only under an assistant message stamped with metadata.work.rootOperationId (the "anchor"). Who stamps it:

• In-process runs: callLlmFinalizer stamps the round's final assistant message.
• Hetero runs never pass that executor, so registerWorksForOperation stamps the anchor itself when the shell scan registered anything — using the completion's assistantMessageId, falling back to the parentId of the last registered tool message (hetero single-step runs persist no final-assistant pointer at all; see heteroFinish).

The whole scan returns {attempted, failed} and the completion backstop (CompletionLifecycle) writes its idempotency marker only when failed === 0. Anything that must be retried — a thrown registration, a {success: false} anchor stamp, an unresolvable anchor — counts into failed. Registration is retry-safe end to end (probe short-circuits, (workId, toolCallId) guard backs it up), so counting a failure is always the right move.

Key decisions (from the Work PRs)

• Only successful create/edit results become Works (LOBE-10967). Read-only queries, comments, merges/closes, and branch/repo operations are excluded — gh pr view printing an entity URL must NOT register.
• owner/repo#number is the canonical github identity, not node_id: the gh CLI surface never returns node_id, and the same entity touched via REST tools and CLI must land on one Work row.
• stdout is the source of truth for identity on the CLI path (gh … create/edit prints the entity URL); the command's edit target is the fallback. The LAST gh create/edit segment of a chained command owns the trailing URL.
• Persisted URLs are http(s)-allowlisted (sanitizeExternalUrl): gh stdout / tool results are member-controlled and the URL reaches shell.openExternal on desktop.
• The tokenizer is deliberately hand-rolled (no shell-quote dep): a real parser would expand what must stay literal, and the worst failure mode is skipping a bookkeeping registration.
• Error-terminated runs skip the completion scan (same success gate as file Works): a gh create that succeeds before a later terminal failure registers no Work that round. Accepted — bounded cost, missing card only.
• Hetero Works register after the terminal snapshot (agent_runtime_end publishes before completeOperation), so a hetero-created card appears on the next refetch, not in the terminal snapshot. Accepted for now.
• PRs pushed via codex's own GitHub integration are not covered — no shell record exists; text-mention parsing was rejected as too false-positive-prone.

Extension recipes

New skill provider (structured tools), e.g. Notion:

1. packages/types/src/work.ts — extend WORK_SKILL_PROVIDERS + WORK_PROVIDER_RESOURCE_TYPES (+ resource-type unions).
2. packages/database/src/models/work/<provider>ToolResult.ts — write the normalizer; mirror linearToolResult.ts.
3. packages/database/src/models/work/index.ts — add it to SKILL_TOOL_RESULT_NORMALIZERS (the satisfies forces this).
4. Pick a stable cross-surface resourceId (the github lesson: choose the identity every surface can produce).

New shell CLI scanner, e.g. linear CLI or vercel deploy:

1. Add a normalizer in packages/database/src/models/work/ reusing parseShellCommandSegments, plus a WorkModel.registerShell<X>Result facade method.
2. Add apps/server/src/services/workRegistration/shellWorkScanners/<name>.ts with { matches, name, register } — matches stays a cheap substring check.
3. Append it to SHELL_WORK_SCANNERS in shellWorkScanners/index.ts. The engine needs zero changes.
4. Pin the parsing edge cases in packages/database/src/models/work/__tests__/ (wrapper unwrap, chained segments, failed/read-only commands must not register).

New Work type (new card kind):

1. @lobechat/types — extend WorkType and related unions.
2. packages/database/src/models/work/ — new adapter module + WORK_TYPE_ADAPTERS entry (compile-enforced).
3. Client — add the type descriptor for the card UI.
4. Add the type to the read-path opt-in gate (OPT_IN_WORK_TYPES pattern) — released clients throw on unknown types; never ship a new type into the legacy result set.

File map