Epic: surpass Claude Code on agentic coding (subagents, modes, checkpoints, background) #196

Activity

1. Delqhi

   added

   enhancementNew feature or requestNew feature or request

   epicUmbrella trackerUmbrella tracker

   loop-systemAgent loop core: budget, gates, delegationAgent loop core: budget, gates, delegation

   on Jun 16, 2026
2. Delqhi commented on Jun 16, 2026

   

   Delqhi

   on Jun 16, 2026

   CollaboratorAuthor

   More actions

   Implementation Roadmap (v0.1 draft — open for feedback)

   Each issue below maps to a single PR with a clear scope, mandate-compliance check, and acceptance criteria. The four sub-issues build on each other: #192 is the foundation, the rest compose it.

   ---

   PR 1: #192 Subagents (foundation, 1-2 weeks)

   Scope: spawn isolated child context, return summary, propagate stop-gate + lessons.

   // internal/subagent/spawn.go
   type SpawnOpts struct {
       Prompt    string
       Profile   string  // agent profile name
       MaxTurns  int     // child-side budget (defense in depth)
       VerifyCmd string  // child runs verify-gate on its result
   }
   
   type SpawnResult struct {
       Summary  string
       Verdict  string  // verified | unverified | failed
       Tokens   int
       CostUSD  float64
   }
   
   func Spawn(ctx context.Context, opts SpawnOpts) (SpawnResult, error)

   Key constraints:

   • Each subagent gets a fresh agentloop.Loop (no shared state with parent)
   • The parent's verification gate is RE-RUN on the child's summary (defense in depth)
   • Lessons are READ-ONLY for the child (no writes — they're the parent's domain)
   • The child's transcript is saved but never sent back to the parent (only the summary)
   • Spawning is bounded: max 4 concurrent subagents per parent turn

   Acceptance:

   • sin subagent 'review the auth module' returns a 200-word summary + a verdict
   • The summary passes the parent's verify-gate (so the parent can trust it)
   • Lessons are filtered: the child sees parent's pre-existing lessons but its own attempts don't pollute them
   • Test coverage ≥ 80%; race-clean

   ---

   PR 2: #193 Permission modes (additive, 2-3 days)

   Scope: three session-wide modes layered on the existing per-tool rule engine.

   Mode	Behavior
   plan (default)	Read-only tools allowed; mutating tools require explicit approval
   acceptEdits	Edit/Write allowed; Bash requires approval; deny list still hard-blocks
   bypass	All allow-list tools allowed; deny list still hard-blocks

   Key constraints:

   • Layered: bypass mode does NOT override the deny list (M4 mandate preserved)
   • The mode is set per session, not per turn (no flip-flopping)
   • The mode is logged in the session transcript for audit

   Acceptance:

   • sin-code chat --mode=plan shows mutating tools as 'ask' in the rule engine
   • sin-code chat --mode=acceptEdits allows Edit/Write without per-tool approval
   • sin-code chat --mode=bypass still blocks 'Bash:rm -rf' via the deny list
   • All three modes pass existing tests (additive only)

   ---

   PR 3: #194 Checkpoints (safety net, 3-5 days)

   Scope: auto-snapshot before mutating tools; sin checkpoint / sin rewind.

   Storage: content-addressed via modernc.org/sqlite. Each checkpoint = a Merkle root of the workspace files. The Checkpoint table is keyed by session-id + turn-id.

   type Checkpoint struct {
       ID        string  // session-id:turn-id
       Merkle    [32]byte
       CreatedAt time.Time
       Files     int  // count
       Size      int  // total bytes
   }

   Key constraints:

   • Snapshot ONLY on mutating tools (Edit, Write, Bash with non-readonly verbs)
   • Snapshot is incremental (only changed files; baseline = last checkpoint)
   • Rewind is per-turn (not per-session) — no global undo
   • Rewind is logged; the operator must explicitly confirm

   Acceptance:

   • After 3 Edit operations, sin checkpoint list shows 3 entries
   • sin rewind <id> restores the workspace to that snapshot
   • sin rewind on a non-mutating turn is a no-op
   • Rewinding doesn't delete the lesson log (the agent can learn from the rewind)

   ---

   PR 4: #195 Background tasks (compose, 3-5 days)

   Scope: sin_background CLI + sin bg list|logs|cancel. Each background task is a subagent.Spawn on a goroutine.

   type BackgroundTask struct {
       ID        string
       Prompt    string
       StartedAt time.Time
       Status    string  // running | verified | failed | cancelled
       Result    *SpawnResult
   }
   
   func Background(ctx context.Context, opts SpawnOpts) (string, error)  // returns task id
   func List(filter TaskFilter) []BackgroundTask
   func Logs(id string, lines int) string
   func Cancel(id string) error

   Key constraints:

   • Each background task is a fresh agentloop.Loop (no shared state with parent)
   • The parent turn can return immediately after Spawn; the task runs concurrently
   • The result is stored in SQLite; the parent reads it on demand via sin bg logs
   • The on_complete hook fires the parent's lesson-recorder automatically (closed learning loop)

   Acceptance:

   • sin_background 'refactor auth module' returns a task id immediately
   • sin bg list shows the task as 'running' then 'verified' when done
   • sin bg logs <id> shows the child's transcript (truncated to last 100 lines by default)
   • sin bg cancel <id> kills the goroutine cleanly (defer-based cleanup)
   • The parent's session does not block on the background task

   ---

   Sequencing rationale

   1. Subagents: isolated-context delegation (sin_subagent) #192 first — every other feature builds on it. Without isolated contexts, the other three are unsafe to use.
   2. Session-wide permission modes (plan / acceptEdits / bypass) #193 second — small, immediately useful, no architectural risk. Unlocks the demo: sin-code chat --mode=acceptEdits is a much better UX than today's per-tool approval.
   3. Workspace checkpointing + rewind (sin checkpoint / sin rewind) #194 third — provides the safety net that makes --mode=bypass actually safe. Without it, bypass is just 'turn off safety'.
   4. Background tasks (sin_background / sin bg list|logs|cancel) #195 last — composes Subagents: isolated-context delegation (sin_subagent) #192 onto goroutines; the moment we have a working subagent.Spawn, background tasks are a 50-line wrapper.

   Total estimate

   • 3-4 weeks for the full sequence
   • 1 dev (no parallel work needed because Subagents: isolated-context delegation (sin_subagent) #192 blocks the rest)
   • The internal/agentloop package is the load-bearing primitive: every PR touches it

   What this epic does NOT do

   • Network-driven subagent delegation (cross-process) — out of scope, see issue on Federation
   • Multi-user permission modes (operator A vs operator B) — out of scope, M4 is single-user
   • CRDT-based workspace sync — out of scope (Federation in [v3.20 rag] RAG over instinct store: semantic search for top-N selection #160 v2)

   ---

   This is a draft. Open for review before PR 1 starts.
3. Delqhi commented on Jun 16, 2026

   

   Delqhi

   on Jun 16, 2026

   CollaboratorAuthor

   More actions

   All four Gaps closed (sequencing as suggested in body):

   • Subagents: isolated-context delegation (sin_subagent) #192 (subagents) -> PR feat(subagent): sin-code subagent CLI (issue #192, wraps #153) #231 (sin-code subagent CLI) merged 2026-06-16
   • Session-wide permission modes (plan / acceptEdits / bypass) #193 (perm modes) -> PR feat(permission): session-wide modes (issue #193, plan/acceptEdits/bypass) #234 (plan/acceptEdits/bypass) merged 2026-06-16
   • Workspace checkpointing + rewind (sin checkpoint / sin rewind) #194 (rewind) -> PR feat(checkpoint): workspace checkpointing + rewind (issue #194) #235 (content-addressed snapshots) merged 2026-06-16
   • Background tasks (sin_background / sin bg list|logs|cancel) #195 (background) -> PR feat(agentloop): background task registry (issue #195, v0 in-process) #233 (in-process TaskRegistry) merged 2026-06-16

   Sub-issues are all closed. The non-negotiable constraints held:

   • M2 preserved: no new runtime dependency. modernc.org/sqlite
     (already in go.sum) for the index, stdlib for blob I/O.
   • Default behavior is byte-for-byte identical: --mode empty,
     --no-auto-checkpoint, no --background. Every new feature
     is opt-in via flag.
   • All five loop-engineering issues (Loop: Stagnations-Erkennung + adaptives Stop-Budget #150, Loop: Token-/Kosten-Budget mit Hard-Cap #151, Loop: Self-Reflection-Turn vor Completion #152, Loop: Sub-Agent / Tool-Delegation #153, Verify: Sprach-agnostische Predicates (Polyglot) #154)
     that composed into the subagents surface are also merged +
     closed.

   Closing this epic. The roadmap item 'surpass Claude Code on
   agentic coding' is satisfied for this cycle.
4. Delqhi

   closed this as completedon Jun 16, 2026