Project Tracker: Peggy — personal-assistant agent on glue

[erain]

This pinned issue is the durable progress keeper for Peggy, a long-running personal-assistant agent built on the glue framework. It tracks execution order, status, and acceptance for the OpenClaw/Hermes-inspired local assistant work.

Operating contract: no feature issue is closed without a merged PR that references it. See CONTRIBUTING.md for the per-issue workflow.

Product Goal

Peggy is a local-first personal-assistant agent that is:

• reachable from multiple channels: CLI, local daemon clients, and Telegram today
• durable across sessions through SQLite-backed memory and recall
• capable of permission-gated coding work in trusted local workspaces
• extensible through file-backed skills, roles, and MCP servers
• a testbed for keeping glue clean: product policy lives in agents, not framework defaults

Inspirations: OpenClaw for channel multiplexing and Hermes-Agent for skills-as-files, FTS5 session search, and tools-as-RPC.

Current Phase

M7: Dual-track — Glue coding-agent binary + Peggy OpenClaw polish — in progress as of 2026-06-07.

🚀 Public launch (2026-05-27). The framework now has a homepage — https://glue-framework-site.vercel.app — backed by a first tagged release v0.1.0 (pre-1.0 stance: ADR-0013). Discussions and a security policy (SECURITY.md) are live; issue + PR templates wired. Social announcement still gated on a deliberate decision.

M1–M6 are shipped (see Shipped Milestones). Peggy is in owner/developer dogfooding on the peggy-dogfood Multipass VM. Work now runs on two parallel tracks, kept independently reviewable under the one-issue/one-PR contract:

• Track A — Glue as a coding-agent binary (framework). Make cmd/glue an awesome coding agent in its own right. ADR-0012 established the boundary: Glue owns the reusable coding-tool bundle (tools/coding) and the binary's coding surface (glue run/serve --coding); Peggy is a thin consumer. The coding-tool surface and the interactive TUI are complete — provider choice + Pi-class tools (edit_file, list_dir, find_files, grep) + a bubbletea TUI with streaming text, tool cards, inline permission prompts, edit_file diff preview, and slash commands. glue run --provider codex --coding is now an interactive coding agent by default in a terminal. Only one Track A enhancement remains: the sandboxed Executor backend (container/VM) behind the existing seam.
• Track B — Peggy as an OpenClaw-clone personal assistant (product). Harden the always-on assistant from concrete dogfood findings: proactivity, richer channel UX, and dashboard actions.

The architectural rule still holds (ADR-0005 §1): product concerns live in agents/peggy; reusable primitives live in glue / tools/*.

M7 Work Queue

Track A — Glue coding-agent binary

• SDK coding-agent boundary (ADR-0012). glue: extract reusable coding-agent tool bundle and wire cmd/glue --coding (ADR-0012) #270, PR glue: add tools/coding bundle and cmd/glue --coding (closes #270) #271, commit 867aada. Adds tools/coding reusable bundle and glue run/serve --coding; makes Peggy a thin consumer with a CodingExecutor seam.
• Provider registry in cmd/glue. cmd/glue: select provider from registry, support codex coding agent #272, PR cmd/glue: select provider from registry, support codex coding agent (closes #272) #274, commit 5a0d2ac. glue run/serve --provider selects codex/gemini/nvidia/openrouter from the registry; --model defaults to the provider's default. Enables glue run --provider codex --coding.
• Pi-class edit_file tool. tools/fs: add edit_file tool for surgical edits #275, PR tools/fs: add edit_file tool for surgical edits (closes #275) #276, commit e3bd34a. Permission-gated surgical exact-string replacement with a unique-match guard and replace_all, registered in the tools/coding bundle.
• Read-only code-navigation tools. tools/fs: add read-only navigation tools (ls, find_files, grep) #277, PR tools/fs: add read-only navigation tools (ls, find_files, grep) (closes #277) #278, commit d625a97. Adds list_dir, find_files, and grep (workspace-scoped, escape-safe, .git-skipping, output-capped; grep skips secret/oversized files), registered in the tools/coding bundle.
• Interactive coding-agent TUI. cmd/glue: interactive TUI for the coding agent (closes the M7 Track A 'interactive multi-turn' loophole) #291, PR cmd/glue: interactive TUI for the coding agent (closes #291) #292, commit ccf04fc. glue run opens a bubbletea TUI by default in a terminal (scrollable transcript, streaming text, tool-call cards, inline permission prompts, edit_file diff preview, slash commands); scripted and --prompt paths preserved. Designed in ADR-0014. Charmbracelet deps live strictly under cmd/glue/tui/; the library import graph is unchanged.
• TUI inline / command autocomplete. cmd/glue/tui: inline autocomplete for / slash commands #318, PR cmd/glue/tui: inline autocomplete for / slash commands (closes #318) #319, commit 89d26bf (v1.9.0). Typing / opens a filtering command popup (↑/↓ navigate, Tab complete, Esc keep-text, Enter run-or-complete), mirroring the @-file picker; /help and the picker now share one command list.

Harness quality — be the best Gemini + open-weight coding harness

Source-verified analysis of pi / Cline / Codex CLI / Gemini CLI distilled into docs/coding-harness-roadmap.md (#336). Implementation order (P0 first — all four harnesses converge on these):

• edit_file repair ladder — flexible matching, instructive errors, escape repair. tools/fs: edit_file repair ladder — flexible matching, instructive errors, escape repair #338, PR tools/fs: edit_file repair ladder (closes #338) #346, commit 925092f.
• Structured output truncation — head/tail shell capture + temp-file spool, dual read caps, next-action markers. tools: structured output truncation — head/tail shell capture, dual read caps, next-action markers #339, PR glue, tools: structured output truncation — head/tail capture, spooling, read paging (closes #339) #347, commit 4d24ade.
• History hardening — loop.HardenHistory repairs orphaned tool calls / invalid turns / foreign-model artifacts before every request. loop: history hardening — repair orphaned tool calls and invalid turns before every request #340, PR loop: history hardening — repair orphaned tool calls and invalid turns (closes #340) #348, commit 6364ddd.
• Retry/overflow recovery — classified retries with backoff + retry hints, typed OverflowError → compact-once-and-retry (ADR-0017). loop: retry/overflow recovery — classified retries with backoff, compact-once-and-retry on overflow #341, PR loop, glue: classified retry + overflow recovery (closes #341, ADR-0017) #349, commit 085208b.
• Compaction upgrade — structured snapshot + injection firewall, handoff framing, cumulative file ledger, safe splits, inflation guard. glue: compaction upgrade — structured snapshot, keep-recent verbatim, handoff framing, file ledger #342, PR glue: compaction upgrade — structured snapshot, file ledger, safe splits, inflation guard (closes #342) #350, commit c0bd7e2.
• Gemini next-speaker check + stall recovery — bounded AutoContinue nudges; surrogate item verified moot per providers/gemini: sanitize tool schemas (nullable tuples) + strip invalid Unicode in prompts #313. providers/gemini + loop: next-speaker check and stall recovery for Gemini #343, PR loop, glue, cmd/glue: next-speaker stall recovery (closes #343) #351, commit 7dfe7fe.
• Loop & mistake guardrails — graduated identical-call + consecutive-mistake detectors with typed halt errors. loop: mistake and loop guardrails — consecutive-mistake counter, identical-call detector, no-tool reprompt #344, PR loop: graduated guardrails — identical-call and consecutive-mistake detectors (closes #344) #352, commit cc3f98c.
• Per-model capability registry + tool-owned prompt snippets — providers.Capabilities + assembled coding system prompt (previously none was sent!). providers: per-model capability registry + tool-owned prompt snippets #345, PR providers, tools, cmd/glue: capability registry + tool-owned prompt snippets (closes #345) #353, commit ef3f078. Queue complete — shipped as v1.13.0.

🌐 Coding-agent homepage shipped (2026-06-09). New repo erain/glue-coding-agent-site (Astro+Tailwind, same design system as the framework site, terminal-green accent) → https://glue-coding-agent-site.vercel.app once imported in the Vercel dashboard. The framework site's §06/§07 now cross-link it.

Goal loop — "loop engineering" / /goal (ADR-0016)

• Phase 1 — Agent.PursueGoal primitive. glue: goal loop primitive (Agent.PursueGoal) — loop-engineering /goal, Phase 1 #320, PR glue: goal loop primitive Agent.PursueGoal (closes #320) #321, commit e198da5 (v1.10.0). Headless goal loop: plan a verifiable checklist → maker (fresh Ralph-style session) → separate checker audits via structured output → decide. Guardrails: max iterations, no-progress, token budget. GoalSpec.Emit streams progress.

• Phase 2 — TUI /goal. cmd/glue/tui: /goal — drive Agent.PursueGoal from the TUI (goal loop Phase 2) #322, PR cmd/glue/tui: /goal — drive Agent.PursueGoal from the TUI (closes #322) #323, commit 8ac9044 (v1.11.0). /goal <objective> runs PursueGoal in the background (chat stays usable); live in-place checklist card + ◎ goal status-bar segment; status|pause|resume|clear (resume continues from the verified checklist via new GoalSpec.Checklist seed — no re-planning). Also: GoalSpec.Permission now reaches planner/checker sessions; TUI permission prompts queue FIFO.

• Phase 3 — durable, isolated, schedulable goals (v1.12.0):

  • 3a — durable glue/goal:* state + resume across restarts. glue: durable goal state + resume across restarts (goal loop Phase 3a) #324, PR glue: durable goal state + resume across restarts (closes #324) #325, commit 949c7bd. PursueGoal checkpoints a GoalRecord per verdict (cancel → paused); LoadGoal/ListGoals; StartIteration keeps resumed sessions fresh; TUI /goal resume survives restarts, /goal list.
  • 3b — worktree isolation. cmd/glue: goal worktree isolation — /goal -w runs the maker on a goal/<id> branch (goal loop Phase 3b) #326, PR cmd/glue: goal worktree isolation — /goal -w (closes #326) #327, commit 00088ea. /goal -w runs maker+checker in .glue/worktrees/<goal-id> on branch goal/<id> (reviewable branch, never auto-merged); resume re-attaches via GoalRecord.WorkDir.
  • 3c — headless glue goal subcommand (scheduled/CI goals). cmd/glue: headless 'glue goal' subcommand — scheduled/CI goal runs (goal loop Phase 3c) #328, PR cmd/glue: headless glue goal subcommand (closes #328) #329, commit 10bc6dc. Full loop without a TUI, status-mapped exit codes (0/2/3/4/1), --list/--resume/--worktree/--budget; store shared with the TUI; verified live end-to-end on gemini.

• (planned, follow-up) Daemon goal endpoints — start/inspect goals over glue serve/connect (alongside agents/peggy: surface schedules over daemon, Telegram, and dashboard #280's schedule surfacing).

• (planned, later) Sandboxed Executor backend (container/VM) behind the existing seam.

• (planned, follow-ups to cmd/glue: interactive TUI for the coding agent (closes the M7 Track A 'interactive multi-turn' loophole) #291) TUI on glue connect (daemon-backed shell); per-tool spinners; per-hunk diff approval; glamour markdown rendering for assistant text.

Gemini harness hardening (being a first-class Gemini 3.x coding harness) — shipped in v1.8.0

• Round-trip thoughtSignature + enable includeThoughts. providers/gemini: round-trip thoughtSignature on tool calls + enable includeThoughts (fixes thought_signature 400) #311, PR providers/gemini: round-trip thoughtSignature + enable includeThoughts (closes #311) #315, commit c138c08. Fixes 400 … Function call is missing a thought_signature on the second turn of any Gemini 3.x tool conversation — captures the signature onto loop.ContentPart.Signature and echoes it on replay. Verified live with a two-turn tool loop.
• Synthetic thoughtSignature fallback. providers/gemini: synthetic thoughtSignature fallback for active-loop model turns #312, PR providers/gemini: synthetic thoughtSignature fallback for active-loop turns (closes #312) #316, commit 7c0f6c2. Injects the skip_thought_signature_validator sentinel on active-loop model turns that reach Gemini 3.x without a real signature (compacted / pre-fix history). Verified live.
• GOOGLE_GENAI_API_VERSION env knob. providers/gemini: GOOGLE_GENAI_API_VERSION env knob #314, PR providers/gemini: GOOGLE_GENAI_API_VERSION env knob (closes #314) #317, commit 7275e3a. Pins the API version (e.g. v1alpha) to reach version-gated preview features without a code change.
• Schema sanitization / Unicode stripping — verified unnecessary. providers/gemini: sanitize tool schemas (nullable tuples) + strip invalid Unicode in prompts #313, closed without a PR. Probed live: Gemini 3.x accepts array-typed type and recursive $ref via parametersJsonSchema, and Go's encoding/json auto-sanitizes invalid UTF-8 — so the JS-harness sanitization tricks are no-ops on glue's path. (The "never retry 400" trick is likewise moot: glue has no provider-level retry; the genai SDK owns transport.)

All four landed as v1.8.0 (the gemini default-id fix that shipped as v1.7.0 was also promoted out of the changelog's Unreleased section).

Track B — Peggy OpenClaw polish

• Scheduled / proactive runs. agents/peggy: add scheduled/proactive runs #273, PR agents/peggy: add scheduled/proactive runs (closes #273) #279, commit a3d23b9. peggy serve runs a scheduler; peggy schedule add|remove / peggy schedules manage persisted prompt/skill runs on --every/--at cadence with a per-schedule permission tier (default trusted). Fire-forwards past missed ticks.
• (planned) Surface schedules over daemon / Telegram / dashboard. agents/peggy: surface schedules over daemon, Telegram, and dashboard #280 (follow-up to agents/peggy: add scheduled/proactive runs #273).
• (planned) Telegram streaming / edited replies instead of progress-then-final.
• (planned) Dashboard actions beyond inspection (run skills/roles, forget memory from the UI).
• (open) Findings filed as they surface from dogfooding the peggy-dogfood VM.

Repo hygiene (shipped)

• Repo re-org: cmd/glue de-monolithed. cmd/glue: split the main.go monolith into per-subcommand files #330, PR cmd/glue: split the main.go monolith into per-subcommand files (closes #330) #331, commit b020548. main.go (3,005 lines) split into per-subcommand files (run / usage / serve / connect / connect_inspect; tests along the same seams) — pure declaration movement, signatures byte-identical.
• Version-line drift reconciled in docs. docs: reconcile version-line drift (0.x → 1.x) and refresh stale status sections #332, PR docs: reconcile version-line drift (0.x → 1.x) and refresh stale status (closes #332) #333, commit 8268799. ADR-0013 addendum records the accidental v0.5.0 → v1.2.0 jump and carries the pre-1.0 stance onto the 1.x line; README "Project status" / CHANGELOG header / project-plan Track A brought current; .gitignore covers /glue + /peggy.
• gofmt drift fixed + CI gate. repo hygiene: fix gofmt drift (23 files) and enforce gofmt in CI #334, PR repo: fix gofmt drift in 23 files and enforce gofmt in CI (closes #334) #335, commit 5a90686. 23 files re-formatted; CI test job now fails on gofmt -l output.
• Newcomer docs overhaul. docs: newcomer-focused docs overhaul (README, building-agents guide, examples, godoc) #281, PR docs: newcomer-focused docs overhaul (closes #281) #282. README restructured into a user-first flow, new docs/building-agents.md end-to-end guide, examples/local-agent modernized to glue.NewTool[T], doc.go staleness fixed.
• Opt-in fixture replay. agents/glue-review: make flaky TestFixtureReplay opt-in so it stops gating unrelated PRs #283, PR agents/glue-review: make flaky TestFixtureReplay opt-in (closes #283) #284. Gated glue-review's non-deterministic TestFixtureReplay behind GLUE_REVIEW_LIVE_FIXTURES so free-model variance no longer flakes the per-PR live CI job.
• Project-plan / tracker refresh. docs: refresh project-plan.md current status (and tracker #110 framing) #285. Brought docs/project-plan.md current (dual-goal status; #1 marked historical) and refreshed this phase summary.
• Repo-wide stale-reference sweep. docs: sweep remaining stale references (#1 tracker, design package list, peggy roadmap, gap-analysis) #287, PR docs: sweep remaining stale references (closes #287) #288. CONTRIBUTING + issue-automation point at #110 (#1 historical), design.md package list completed (tools/shell, tools/coding, tools/mcp), peggy README "What's coming" refreshed, gap-analysis marked historical snapshot.
• Official homepage shipped. New repo erain/glue-framework-site — Astro + Tailwind landing page modeled on glue-review-site (hero + 7 numbered sections + FAQ + OG image), deployed to https://glue-framework-site.vercel.app with GitHub→Vercel auto-deploy on push.
• Launch hygiene + v0.1.0 cut. launch hygiene: ADR-0013 stability stance, SECURITY.md, issue/PR templates, v0.1.0 CHANGELOG cut #289, PR launch hygiene: ADR-0013, SECURITY, issue/PR templates, v0.1.0 cut (closes #289) #290, commit 3866291. ADR-0013 records the pre-1.0 stability stance; SECURITY.md provides a reporting channel; .github/ISSUE_TEMPLATE/ + PR template wired; first tagged release v0.1.0; Discussions enabled; repo description / homepage / topics set.

Completed Dogfood Readiness Work

• Dogfood readiness gate. agents/peggy: add dogfood doctor command #252, PR [codex] agents/peggy: add dogfood doctor command #253, commit 267e004. Adds peggy doctor text/JSON checks for provider credentials, SQLite recall, workspace context, skills/roles, coding policy, Telegram config, permissions, and MCP settings.
• First-run dogfood quickstart. agents/peggy: add first-run dogfood quickstart #254, PR [codex] agents/peggy: document dogfood quickstart #261, commit c96cd0d. Documents install/auth, starter workspace, minimal dogfood settings, peggy doctor, daemon smoke, recall smoke, and optional Telegram wiring.
• Persistent permission decisions. daemon: persist permission decisions across restarts #255, PR daemon: persist permission grants #262, commit e58f46a. Persists daemon permission grants across restarts and adds glue connect --permissions / --forget-permission inspect and revoke controls.
• Telegram daemon reply UX. agents/peggy/telegram: improve daemon reply UX #256, PR telegram: improve daemon reply UX #263, commit 23000d3. Adds Telegram /help, immediate progress replies for daemon runs, and long-reply splitting instead of hard truncation.
• Session history browser. agents/peggy: add session history browser #257, PR peggy: add session history listing #264, commit f4a771c. Adds peggy sessions with prefix filtering, JSON output, and SQLite/file-backed session summaries without starting a provider.
• Daemon ops diagnostics. agents/peggy: add daemon ops diagnostics #258, PR daemon: add ops diagnostics #265, commit 36a66cd. Adds authenticated daemon diagnostics and glue connect --diagnose for healthy, missing/stale metadata, bad-token, unreachable, and recent-error states.
• Memory backup and restore. agents/peggy: add memory backup and restore #259, PR agents/peggy: add memory backup and restore #266, commit 64a8fb5. Adds versioned curated-memory JSON export/import, validation, dry-run reports, and duplicate skipping by id or normalized content.
• Local control surface. agents/peggy: add local control surface #260, PR agents/peggy: add local dashboard #267, commit b8da74b. Adds peggy dashboard, a localhost web control surface for daemon health, diagnostics, tools, skills, roles, memories, recall, and recent local sessions.

Dogfood Polish

• Dashboard prompt runs. agents/peggy: let dashboard start prompt runs #268, PR agents/peggy: run prompts from dashboard #269, commit 2e67e1a. Lets the local dashboard start prompt runs via the existing authenticated daemon run/SSE endpoint and render the final assistant response without exposing the bearer token in HTML.

Shipped Milestones

• M1 — Peggy v0.1: "She remembers and you can text her". Shipped 2026-05-16. Foundation ADRs, Codex provider, SQLite/FTS5 store, memory tools, single-prompt CLI, and Telegram channel.
• M2 — Peggy v0.2: "She can code". Shipped 2026-05-24. Executor/Permission/Hook interfaces, shell/fs/git tools, subagents, opt-in Peggy coding mode, and CLI/Telegram permission prompts.
• M3 — Peggy v0.3: "Multi-channel daemon". Shipped 2026-05-24. HTTP+SSE daemon protocol, peggy serve, glue connect, Telegram daemon-client mode, and per-channel permission tiers.
• M4 — Peggy v0.4: "Ecosystem". Shipped 2026-05-24. MCP stdio and Streamable HTTP clients, MCP tools/resources/prompts, daemon status/tools/inspect, and local peggy status.
• M5 — Peggy v0.5: "Workflow/runtime". Shipped 2026-05-25. Workspace context, file-backed skills, roles, starter peggy init, usage summaries, daemon memory/recall controls, and Telegram daemon role/skill/memory/status commands.
• M6 — Dogfood hardening and product shell. Shipped 2026-05-25. Doctor, quickstart, persistent permission grants, Telegram daemon UX, sessions, diagnostics, memory backup/restore, and local dashboard.

See agents/peggy/CHANGELOG.md for detailed release notes and PR history.

Known Product Gaps

The M6 dogfood blockers are complete; the active gaps are now tracked in the M7 Work Queue above. Additional Track B candidates not yet filed: packaged install flows, calendar / external-service triggers, and more realistic long-running personal-assistant evaluations. New gaps from daily dogfooding of peggy-dogfood should be filed as focused follow-up issues and linked under the M7 Work Queue.

Pacing

Sessions should check /status periodically and pause/resume around quota limits rather than forcing oversized turns. Keep PRs small enough to review and merge independently.