v1.13.0 — harness quality

The complete harness-quality queue from docs/coding-harness-roadmap.md — eight improvements distilled from source-verified analysis of pi, Cline, Codex CLI, and Gemini CLI, targeting Gemini 3.x first and open-weight OpenRouter/NVIDIA models second.

P0 — wasted-turn killers

• edit_file repair ladder (#338): exact → whitespace → indentation (re-indented replacement) → punctuation-folding → block-anchor matching, over-escape repair, BOM/CRLF preservation, updated-lines echo, instructive errors, lazy-placeholder guard.
• Structured output truncation (#339): head+tail shell capture with omitted counts and a named full-output spool file; read_file offset/max_lines paging with "Use offset=N to continue".
• History hardening (#340): loop.HardenHistory synthesizes results for dangling tool calls, drops orphans/empty turns, normalizes call IDs, strips foreign-model artifacts — before every request.

P1 — reliability and cost

• Retry/overflow recovery (#341, ADR-0017): classified loop-level retries with backoff and server retry hints; typed OverflowError → compact-once-and-retry in sessions.
• Compaction upgrade (#342): structured state snapshot with injection firewall, handoff framing, cumulative file ledger, tool-pair-safe splits, inflation guard.
• Capability registry + tool-owned prompts (#345): providers.Capabilities per provider; coding.SystemPrompt assembled from the active toolset (the --coding agent previously ran with no system prompt at all).

P2 — polish

• Next-speaker stall recovery (#343): bounded "Please continue." nudges for Gemini's narrate-then-stop stall, auto-enabled per the registry.
• Loop guardrails (#344): graduated identical-call and consecutive-mistake detectors with typed halt errors.

Default-behavior notes: loop-level retries and guardrails are on by default with opt-outs (RetryPolicy.Disabled, GuardrailPolicy.Disabled); a 429 now retries up to 3× before failing. All changes additive otherwise; full changelog in CHANGELOG.md.

v1.12.0 — goal loop Phase 3: durable, isolated, schedulable goals

Completes the goal-loop roadmap from ADR-0016 (Phase 1 Agent.PursueGoal shipped in v1.10.0, Phase 2 TUI /goal in v1.11.0).

Durable goal state + resume across restarts (#324, PR #325)

• Agent.PursueGoal checkpoints a GoalRecord (objective, status, verified checklist, iterations, usage, summary) as glue/goal:* session metadata — after planning, every checker verdict, and at the terminal state. Cancellation persists as the new GoalPaused status; GoalRunning marks in-flight records.
• Agent.LoadGoal / Agent.ListGoals read records back; GoalSpec.StartIteration keeps resumed maker/checker sessions fresh (iter-4, check-4, …).
• TUI: /goal resume continues the most recent unfinished goal even in a new process; /goal list shows recent goals with status, fraction, and age.

Worktree isolation — /goal -w (#326, PR #327)

• /goal -w <objective> runs the loop in .glue/worktrees/<goal-id> on branch goal/<id> — your checkout stays untouched and the result is a reviewable branch, never auto-merged. Resume re-attaches the same worktree (GoalSpec.WorkDir / GoalRecord.WorkDir).

Headless glue goal subcommand (#328, PR #329)

• glue goal "<objective>" runs the loop without a TUI, streaming checklist progress to stdout; exit code maps the outcome (0 achieved · 2 blocked · 3 max-iterations · 4 budget-limited · 1 errored) so cron/CI/peggy schedules can branch on it.
• --list / --resume [id] share the TUI's store; --worktree, --max-iterations, --budget, --yolo for unattended runs.
• Verified live end-to-end (gemini): plan → maker → evidence-backed verdict → achieved, exit 0, checkout untouched.

🤖 Generated with Claude Code

v1.11.0 — TUI /goal (goal loop Phase 2)

TUI /goal: drive the goal loop from inside glue run (ADR-0016 Phase 2, #322, PR #323).

• /goal <objective> runs Agent.PursueGoal in the background on its own session ids while the chat stays usable: a live goal card in the transcript updates in place per iteration (verified [x]/[ ] checklist with evidence, iteration counter, token usage, last checker verdict), and the status bar gains a ◎ goal · iter 2/10 · 1/4 ✓ · 12.3k tok segment.
• Subcommands: /goal status, /goal pause (cancels cleanly, keeps the verified checklist), /goal resume (continues from that checklist without re-planning), /goal clear.
• Goal tool calls flow through the same in-card permission prompt as chat turns — pending permission requests are now a FIFO queue, so a concurrent goal + chat request can no longer drop one. --yolo makes goals fully autonomous.
• GoalSpec.Checklist (library): seed PursueGoal with an existing plan — skips planning, respects done flags and evidence. The /goal resume mechanism, and the hook Phase 3 durable resume will reuse.
• GoalSpec.Permission now applies to the planner and checker sessions too (was maker-only) — without this, the checker under an agent with no default permission policy (e.g. the interactive TUI) was denied every side-effecting tool and could not run builds or tests to verify evidence.

🤖 Generated with Claude Code

v1.10.0 — goal loop (loop engineering / /goal)

Loop engineering comes to glue. Agent.PursueGoal turns one persistent objective into an autonomous loop — the /goal pattern (Codex, pi-goal), as a reusable library primitive.

How it works

1. Plan — a planner decomposes the objective into a verifiable checklist.
2. Maker — a fresh session (Ralph-style, seeded from the durable checklist so memory lives in state, not a growing transcript) works the open items.
3. Checker — a separate session with its own model + verifier prompt audits each item against real evidence via structured output, and decides completion. The writer never grades its own homework.
4. Decide — stop on achieved, blocked (no progress), budget_limited, or max_iterations.

Guardrails are first-class (MaxIterations, NoProgressLimit, TokenBudget); GoalSpec.Emit streams progress.

Built entirely on existing primitives (Session.Prompt, Session.PromptJSON, loop.Run/MaxTurns, Agent.Session) — no new dependencies.

This is Phase 1: the headless primitive. The TUI /goal command (Phase 2) and durable resume + branch isolation (Phase 3) follow. Designed in ADR-0016.

(#320, PR #321)

Verified

go build/vet/test ./... green; 7 tests cover every terminal status, the maker/checker call sequence, and event emission.

Full Changelog: v1.9.0...v1.10.0

v1.9.0 — TUI slash-command autocomplete

Type / and the commands come to you. The coding-agent TUI now has inline autocomplete for slash commands, matching the existing @-file picker.

What's new

• Typing / opens a filtering command popup.
• ↑/↓ navigate · Tab completes · Esc closes (keeping what you typed).
• Enter runs a fully-typed or no-argument command immediately, and completes an argument-taking command to /<name> so you can type the argument (e.g. /model ).
• Opens only on a bare slash token (closes once you start typing an argument); mutually exclusive with the @-file picker.
• The command list is now a single source of truth shared by the picker and /help, so they can't drift.

(#318, PR #319)

Verified

go build/vet/test ./... green; new tests cover trigger detection, prefix/alias matching, stable ordering, Tab-complete, Esc-keeps-text, and run-vs-complete Enter behavior.

Full Changelog: v1.8.0...v1.9.0

v1.8.0 — Gemini 3.x thought-signature fix

The headline: Gemini 3.x tool calls now work across turns. The default gemini-3.1-pro-preview returns an opaque thoughtSignature on its function calls and requires it echoed back on replay — without it, the second turn of any tool-using conversation died with 400 … Function call is missing a thought_signature in functionCall parts. That's fixed, verified live with a two-turn tool loop.

Changes

• Round-trip thoughtSignature + includeThoughts (#311, PR #315) — capture the signature from function-call and thought parts (base64 on loop.ContentPart.Signature) and replay it verbatim; send prior thinking back as real thought parts; enable includeThoughts on Gemini 3.x. This is the actual fix for the 400.
• Synthetic thoughtSignature fallback (#312, PR #316) — when an active-loop model turn reaches Gemini 3.x without a real signature (compacted or pre-fix history), inject the skip_thought_signature_validator sentinel so the request still validates. Scoped to the active loop and Gemini 3.x; real signatures are never overwritten.
• GOOGLE_GENAI_API_VERSION env knob (#314, PR #317) — pin the API version (e.g. v1alpha) to reach version-gated preview features without a code change.

Verified

go build/vet/test ./... green; gated live tests TestLiveToolLoopRoundTripsSignature, TestLiveSyntheticSignatureFallback, TestLiveSmoke all pass against gemini-3.1-pro-preview.

Notes

Investigation also closed #313 (schema/Unicode sanitization) as verified-unnecessary: Gemini 3.x accepts array-typed type and recursive $ref via parametersJsonSchema, and Go's encoding/json auto-sanitizes invalid UTF-8 — those JS-harness tricks are no-ops on glue's path.

Full Changelog: v1.7.0...v1.8.0

v1.7.0 — gemini default id correction (gemini-3.1-pro-preview)

Headline

Fix Gemini default id: gemini-3.1-pro-preview (the v1.4.0 default gemini-3.1-pro 404'd because the public id carries the -preview suffix on generativelanguage.googleapis.com v1beta).

Install

go install github.com/erain/glue/cmd/glue@v1.7.0
glue version
glue run --yolo --coding --work .

Override paths still work

• glue run --model gemini-2.5-pro
• GLUE_MODEL=gemini-2.5-pro
• gemini.Options{DefaultModel: "gemini-2.5-pro"} or loop.ProviderRequest{Model: "gemini-2.5-pro"}

Library API

Additive only.

v1.6.0 — wide-terminal cap + glue version

Daily-driver friction round.

Install

go install github.com/erain/glue/cmd/glue@v1.6.0
glue version
glue run --yolo --coding --work .

Fixes

• TUI: cap transcript at 100 cols, center on wide terminals. Conversation no longer stretches edge-to-edge on a 200-col terminal. The viewport and the input box share a single bodyMaxWidth = 100 constant; on anything wider, the conversation column sits centered via lipgloss.PlaceHorizontal. Glamour markdown width inherits the cap. Header + status bar keep full width — they read better edge-to-edge.

• glue version subcommand. Prints module version, git revision, build time, and Go toolchain pulled from the linker-embedded build info. Same data go version -m $(which glue) exposes, just reachable as a first-class command so a stale binary is a one-line diagnosis:

  $ glue version
  glue v1.6.0 (59b5026706e7) · built 2026-06-08T19:17:43Z · go1.25
  

  Also reachable as --version and -v.

Library API

Additive only.

v1.5.0 — Catppuccin TUI theme (Mocha + Latte)

Headline

Catppuccin TUI theme. Mocha (dark) and Latte (light) flavors, picked automatically based on your terminal background.

The previous palette mixed Tailwind primaries under one purple accent — saturation levels didn't match, so colors shouted. Mocha/Latte are intentionally pastel: every color sits at similar lightness so the chrome reads as a coherent design system instead of an alert tree.

Install

go install github.com/erain/glue/cmd/glue@v1.5.0
glue run --yolo --coding --work .

role	Mocha (dark)	Latte (light)
accent	mauve #cba6f7	mauve #8839ef
ink	text #cdd6f4	text #4c4f69
ok	green #a6e3a1	green #40a02b
err	red #f38ba8	red #d20f39
warn	peach #fab387	peach #fe640b
border	surface1 #45475a	surface1 #bcc0cc

Glamour markdown output uses matching JSON style configs (glamour-mocha.json, glamour-latte.json) with chroma syntax highlighting via catppuccin-mocha / catppuccin-latte. On any JSON parse failure the renderer falls back to glamour.WithAutoStyle so a typo can't break the TUI.

Library API

Additive only. No breaking changes.

v1.4.0 — gemini-3.1-pro is the default

Headline

Default model for the gemini provider is now gemini-3.1-pro.

gemini-2.5-flash was removed from generativelanguage.googleapis.com v1beta on the current key surface and started returning 404 on generateContent. gemini-3.1-pro is the unmetered daily-driver target for this project.

Install

go install github.com/erain/glue/cmd/glue@v1.4.0

Override paths still work

• CLI: glue run --model gemini-2.5-pro
• Env: GLUE_MODEL=gemini-2.5-pro
• Library: gemini.Options{DefaultModel: "gemini-2.5-pro"} or loop.ProviderRequest{Model: "gemini-2.5-pro"}

What's not in this release

The Catppuccin TUI theme (#305) is queued as v1.5.0 — landing right after.