feat: inherit parent context for sub-agent sessions

[NicholaiVogel]

Summary

• promote the sub-agent context continuity spec to approved and document the API, MCP, hook, harness, and config surfaces
• add daemon parent-session inference for explicit lineage, OpenClaw session keys, and Claude Code sub-agent hook payloads
• wire Claude Code, OpenCode, and CLI hook payloads so harness sub-agent IDs do not become Signet agent scopes

PR Readiness Checklist

• Spec alignment validated (INDEX.md + dependencies.yaml)
• Agent scoping verified on all new/changed data queries
• Input/config validation and bounds checks added
• Error handling and fallback paths tested (no silent swallow)
• Security checks applied to admin/mutation endpoints
• Docs updated for API/spec/status changes
• Regression tests added for each bug fix
• Lint/typecheck/tests pass locally

Validation

• bun scripts/spec-deps-check.ts
• ./node_modules/.bin/biome check docs/API.md docs/CONFIGURATION.md docs/HARNESSES.md docs/HOOKS.md docs/MCP.md docs/specs/INDEX.md docs/specs/dependencies.yaml docs/specs/approved/sub-agent-context-continuity.md integrations/claude-code/connector/src/index.ts integrations/opencode/plugin/src/index.ts platform/core/src/types.ts platform/daemon/src/hooks.test.ts platform/daemon/src/hooks.ts platform/daemon/src/mcp/tools.test.ts platform/daemon/src/mcp/tools.ts platform/daemon/src/memory-config.ts platform/daemon/src/routes/session-routes.ts platform/daemon/src/subagent-context.ts surfaces/cli/src/commands/hook.test.ts surfaces/cli/src/commands/hook.ts
• git diff --check HEAD~1..HEAD
• bun run build:core
• bun test platform/daemon/src/hooks.test.ts platform/daemon/src/mcp/tools.test.ts surfaces/cli/src/commands/hook.test.ts
• cd integrations/opencode/plugin && bun run typecheck && bun run build
• cd integrations/claude-code/connector && bun run typecheck && bun run build
• cd platform/daemon && bun run build.ts
• cd surfaces/cli && bun run build:cli

Notes

• Full bun run lint still reports pre-existing package.json formatting issues outside this patch; scoped Biome is clean for the touched files.
• No admin endpoint was added; the new session transcript endpoints use existing recall permission middleware.

Closes #315

[PR-Reviewer-Ant]

Review posted through GitHub's PR review UI for commit 6db56882.

[PR-Reviewer-Ant]

Review metadata

• Reviewer: pr-reviewer
• Model: gpt-5.5
• Commit: c15f71cc

I found one correctness issue in the fallback transcript search path: the SQL parameter order does not match the generated placeholders, so non-FTS searches can fail or return no results even when matching transcripts exist.

Confidence: High [sufficient_diff_evidence, targeted_context_included] - The bug is directly visible in the new searchSessionTranscripts fallback query: placeholders in SELECT come before the WHERE agent_id placeholder, but args pushes LIKE parameters first and agentId afterward. This affects the new /api/sessions/search and session_search MCP behavior added by the PR.

[PR-Reviewer-Ant]

The fallback LIKE query binds parameters in the wrong order. The generated SQL starts with ${score} AS rank, which contains the first N ? placeholders, then WHERE st.agent_id = ?; however args is initialized with the LIKE patterns and only then pushes params.agentId. That part is fine for the score placeholders, but after appending AND (${any}), the code pushes the second set of LIKE patterns at the very end, after optional session_key, project, and limit placeholders have already been added. With a session/project filter, those filters receive LIKE values and the final LIKE clauses receive session/project/limit values, so searches can silently miss valid transcript hits. Build the SQL and args in exact placeholder order, or split score/where args so optional filters don't get interleaved incorrectly.

[PR-Reviewer-Ant]

Review metadata

• Reviewer: pr-reviewer
• Model: gpt-5.5
• Commit: 02788fec

I found one OpenCode lifecycle bug: inherited session-start context is queued too late to be injected into the child session's first prompt, which undercuts the PR's main continuity claim for that harness.

Confidence: Medium [sufficient_diff_evidence, missing_runtime_repro] - The OpenCode hook path reads and deletes pending injection before calling ensureSessionStarted, while ensureSessionStarted writes the inherited context back into pendingInject. That ordering is visible in integrations/opencode/plugin/src/index.ts, though I did not run OpenCode to confirm the exact hook timing.

Since my last review:

• Fixed or likely addressed: 1
• Still blocking: 0
• Newly found: 1
• Suppressed due to human rebuttal/out-of-scope: 0

[PR-Reviewer-Ant]

This posts the per-session session-start request after the hook has already read and deleted pendingInject for input.sessionID, so any inherited parent context returned by ensureSessionStarted() is queued for a later prompt instead of the child session's first prompt. For sub-agent continuity, the first prompt is the important one: the sub-agent starts working before the inherited context is visible. The fix should make ensureSessionStarted(input.sessionID) run before consuming pendingInject, or have it return the inject text directly so this same hook invocation can append it.

[PR-Reviewer-Ant]

Review metadata

• Reviewer: pr-reviewer
• Model: gpt-5.5
• Commit: 6db56882

I found one correctness issue in the OpenCode lifecycle path: the newly fetched session-start context for a per-session start is queued after the prompt handler has already consumed pending injections, so the first prompt of a child session misses the inherited parent context.

Confidence: Medium [sufficient_diff_evidence, missing_cross_module_context] - The diff shows prompt.before deleting pendingInject for input.sessionID before calling ensureSessionStarted, while ensureSessionStarted queues the session-start result.inject back into pendingInject. That makes the inherited context unavailable to the prompt currently being handled. I could not inspect the full local repository in this read-only review environment, so the confidence is medium rather than high, but the ordering problem is visible in the changed OpenCode plugin code.

Since my last review:

• Fixed or likely addressed: 1
• Still blocking: 0
• Newly found: 1
• Suppressed due to human rebuttal/out-of-scope: 0

[PR-Reviewer-Ant]

ensureSessionStarted() can enqueue the per-session session-start injection, including inherited parent context, but this runs after prompt.before has already read and deleted pendingInject for this sessionID. That means a newly created OpenCode child session will not receive the inherited parent context on its first prompt; it only becomes pending for a later prompt. Since this PR’s stated behavior is sub-agent context continuity at session start, the session-start call needs to happen before consuming pending injections, or its returned inject needs to be appended directly to the current output.context path.