Changelog

0.1.0 (2026-06-25)

Features

/effort command + reasoning_effort levels (none..xhigh), learn=high
borg learn → combined BORG.md project context (folding in existing project guides)
add /upgrade slash command for plan upgrades
agent,eval: last-lever model tiering + edits-landed/cache-floor eval signals
agent: ask_user — "something else" free-text option + force the tool over prose
agent: ask_user tool — agent-initiated multiple-choice prompt
agent: auto-escalate reasoning effort when a turn is struggling
agent: auto-verify edits before a turn ends (small-model backstop)
agent: auto-verify runs the project's own test command (BORG.md "## Verify")
agent: bias the harness toward autonomy + document the principle
agent: climb the reasoning ladder (none→medium→high→xhigh) when stuck
agent: convergence brakes to cut token waste + honest feedback errors
agent: default reasoning effort to medium
agent: default tool_choice=auto, escalate to required only on a leak
agent: explore-then-implement prompt + 3 discovery/adaptation evals
agent: make /learn reliable on big repos — batch, higher caps, artifact flush
agent: much deeper borg learn — read all docs, compare dirs, find decoys
agent: no-progress guard — stop a model stuck repeating the same call
agent: nudge model to verify a query before concluding absence
agent: nudge the model to batch independent tool calls (parallelism)
agent: nudge the model to use glob + background bash
agent: nudge token-efficient exploration (grep-first, ranged reads)
agent: opt-in model tiering (chuppa->axiom) after effort tops out
agent: per-model system-prompt addendum (Floko-specific guidance)
agent: polish system prompt (agent best practices) + Floko direct-finish
agent: push Floko to batch independent reads (parallel tool exec)
agent: resumed sessions adopt the current system prompt
agent: self-heal retrospective — learn to BORG.md or report harness problems
agent: show parallel tool batches in the UI (⚡ N tools in parallel)
agent: teach the harness to batch git and discover PRs
agent: tool-using coding-agent loop
agent: turborg co-author attribution on commits/PRs (on by default)
agent: wire borg to the metered LLM proxy + one-shot ask
auth: always print the auth URL and honor $BROWSER
auth: BORG_ACCESS_TOKEN env to inject a bearer (PAT) for headless use
auth: brand the PKCE callback page (xShellz logo + theme)
auth: implement PKCE + device login flows with silent refresh
auth: send device-flow users to the app's /device page
auth: show configured endpoints in auth status
brand as Turborg, run as turborg or borg, co-author commits as Turborg
cache plan/catalog for instant startup, refresh in background
config,tui: ~/.config/borg/env loader + show model/effort in footer
debug: add debug plumbing to LLM/UI, TUI, and CLI flag
directory trust matrix (scoped edits) + /learn REPL command
docs: generate a CLI reference + publish it and the changelog to R2
docs: publish per-version reference + versions.json for versioned docs
eval: 60m timeout + model dropdown (nightly = chuppa only)
eval: efficiency-aware baseline — flag step/token drift, not just pass/fail
eval: gated Axiom measurement path (opt-in, subset-capped)
eval: GofmtClean oracle + auto-format guard tasks; seed efficiency baseline
eval: N-run measurement harness + parallel-batch metric; discard batching prompt
eval: parallel + token-frugal evals, cache-ratio report
eval: regression baseline, run-summary graphs, +4 coding tasks
eval: report token totals in repeated-run aggregate; doc OpenRouter billing
eval: rich deterministic report — graphs, per-task metrics, winner
eval: run floko + chuppa nightly and compare results
eval: two harder edge-case corpus tasks (22 total)
eval: two hermetic discovery/search corpus tasks (20 total)
guided tool-calls for gemma — finish tool + finish_reason handling
learn: cover deploy/ops gotchas + verify-before-asserting (small-model tuning)
live /usage from GET /v1/llm/usage (account-wide rolling-24h budget)
llm,agent: retry transient mid-stream errors + per-session debug logs
llm: bounded retry-with-backoff for transient request failures
llm: send a stable prompt_cache_key derived from the system prompt
llm: send X-Turborg-Version so the proxy can gate old clients
llm: stream-friendly timeout + per-turn token/timing footer
models: rename to chuppa/axiom + budget-burn switch hint
parallel read-only tool execution + ranged/bounded reads
release: automate versioning with release-please
release: GoReleaser + R2 publish pipeline
rename --resume to --attach, in-REPL /attach, cwd-scoped sessions
scaffold borg authenticated coding-agent CLI
sessions: /sessions picker, dir-scoped listing, titles + relative time
settings: typed settings.json + /settings command
surface prompt-cache hits + guard prefix determinism
tools,agent: bash_output blocks for completion; polling never escalates
tools: add glob + background bash (bash_output / kill_shell)
tools: auto-format edits (project's own formatter) + finish /usage fix
tools: broaden verify to all popular languages (safe compile/parse-only)
tools: numbered reads + edit_lines + verify (small-model file editing)
tools: read_file reports total line count
tools: sharper file edits/writes for small-model reliability
tui: /context usage bar, /compact, and a near-full context warning
tui: /login, /logout, /status REPL commands
tui: /purge (rename), /privacy, /usage + plan-correct availability
tui: borderless shaded input box, ~/dir footer, paste placeholders
tui: Esc dismisses the slash menu + windowed (paged) menu
tui: Esc-interrupt, paste echo expansion, markdown headings, box polish
tui: frame the input/working line in a rounded box
tui: growing multi-line input box (textarea), pronounced shade, padding
tui: inline Charm REPL + tests, enforce >=90% coverage
tui: interactive /effort + /attach pickers, narrow-width clipping
tui: keep input live while a turn streams; queue follow-ups
tui: live 'working…' indicator during silent model output
tui: live elapsed timer in working line + /learn no-re-read nudge
tui: live per-turn token counter on the working indicator
tui: migrate to Bubble Tea v2 for a native, non-clobbering cursor
tui: navigable slash menu + per-session input history
tui: nudge to re-run /learn when BORG.md drifts behind HEAD (git-only, configurable)
tui: refresh footer git branch on focus/submit/turn-end, not just startup
tui: replay transcript on resume + live slash-command menu
tui: resumable sessions + inline Bubble Tea REPL with live markdown
tui: richer status footer + highlighted prompt echo
tui: startup plan/version/model banner + interactive model picker
tui: Visor Humanoid launch mascot with a settling blink
ui: clean tool-call lines + green/red result status
update: self-updater + startup version nudge
usage: show credit budget instead of input/output token meters

Bug Fixes

agent,tools: make git commits shell-safe — heredoc nudge + deny -m with substitution
agent: /learn back to high effort — BORG.md is a write-once, high-leverage artifact
agent: /learn prompt favors depth over speed (generic, any repo)
agent: /learn uses medium effort + a timed step-budget smoke check
agent: arm the /learn write-guard in the REPL + kill CLI/REPL config drift
agent: bound /learn exploration in the harness, not just the prompt
agent: break degenerate prose loops and force the model to act
agent: bulletproof the loop — cycle detection, run-on/step-cap, bash timeout
agent: declared verify command always runs before finishing
agent: escalate + nudge on in-turn re-edit thrash
agent: guard the reasoning channel for loops too + fix lint
agent: harden the loop against token-thrash, self-kill, and edit blindness
agent: instruct a proportional commit-message body, not just a subject
agent: recover from text-leaked tool calls; learn enriches existing BORG.md
agent: reply in the user's language, answer directly
agent: require the deliverable to be write_file'd before finishing
agent: revert default reasoning to none (cost)
agent: salvage truncated finish calls, flag output-limit cutoffs
agent: verify-command guidance is language-agnostic ("run this")
auth: preserve token environment on silent refresh; clearer 401
debug: complete + green the debug feature
eval: close live client idle conns so goleak stays green
llm: decode /users/me plan as plan_code (was planCode → always "free")
settings: attribution is on/off only, drop name/email from /settings
test: check os.Unsetenv return (errcheck) in config test
tools: grep uses extended regex (-E) so a|b alternation works
tools: hermetic subprocess env + whitespace-tolerant edit_file
tools: never emit empty tool output (defense in depth for the 422)
tools: verify schema breaks Floko guided decoding (empty properties)
tui: align picker columns to the widest label, not a hardcoded width
tui: bound verbose debug output so it can't flood/corrupt the inline render
tui: clear the tool label when the next step starts thinking
tui: echo the invoked name (turborg/borg) in the REPL banner and prompts
tui: extend row highlight to pickers, chevron affordance, drop /borg-purge
tui: highlight selected slash-command description too
tui: slash menu vanished when filtering (inline-render shrink)
tui: tighter input padding, paste \r detection, catalog spacing

Documentation

/effort + reasoning levels, learn=high, per-model token windows
add BORG.md — project context for AI coding agents
add dev/PLANS/PLAN-harness-gemma4 (harness Floko to the limit)
committed loop SVG + standing harness-state snapshot
correct Floko model id + realistic harness positioning
deepinfra: add full verbatim DeepInfra gemma-4-26B-A4B-it API schemas
document the automated release pipeline (release-please → GoReleaser → R2)
document the full built feature set + models/metering + POC status
eval: document running the eval + regression checks locally
how to run the live eval against the LOCAL env (host go test + PAT)
record the slim authorize contract + large-context chain caps
require Docker for build/test; add docker-bin extract target
standing rule — ship to staging, never main (don't ask to promote)
sync project docs with the per-account credit budget