Changelog

All notable changes to Tacit, newest first. The format follows Keep a Changelog and the versions follow Semantic Versioning (before 1.0, a minor bump for a feature or a breaking change, a patch for a fix). Each version section is also the text of its GitHub release; node scripts/release-notes.mjs <version> prints it. One line per change; the PR has the detail.

Unreleased

Added

• A weekly log-audit workflow and pnpm check:ci-logs scan the repository's own CI logs for credentials, personal paths and addresses. (#61)

0.6.1 - 2026-09-01

Small fixes to what a profile reads back and to the usage ledger, a live rehearsal script, and a compatibility check against the host's alpha line.

Upgrading

• Restart dsh web after updating; host code only reloads on a restart.
• No new settings, routes or storage changes.

Added

• pnpm rehearse drives real headless turns in a throwaway DSH_HOME and checks what Tacit learned from them. (#59)

Changed

• The weekly compatibility check also runs against @deepseek-ai/dsh@alpha. (#57)

Fixed

• Pattern rows whose kinds spell the same thing fold into one row with summed counts. (#57)
• A profile keeps one directive on trial per scope; extra candidates from an older profile wait in the queue. (#57)
• A run that ends right before the harness exits still counts in the Usage card; the summary is written when the run ends. (#58)

0.6.0 - 2026-09-01

One pass over the whole UI: the same card, badge, chip and row idioms on every surface, fewer visible controls, and a directive receipt you can actually read.

Upgrading

• Restart dsh web after updating; host code only reloads on a restart.
• No new settings, routes or storage changes.

Changed

• Directive cards carry a status-coloured rail and a pill badge; the receipt reads as tiles, a lifecycle timeline, trigger chips and an identity grid. (#54)
• Every settings row shares one shape, and each checkbox setting is a whole-sentence clickable label. (#54)
• Turn rows lead with the heading and two actions; tools, steps, tokens and retries read as one quiet dot-separated line under the prompt. (#54)
• One rotating chevron marks every card head and disclosure; zero counts are hidden, and the Learning and Agent guidance heads summarize while collapsed. (#54)
• The Improve preview leads with what changed and highlights the proposed rewrite; the confirm dialog narrows to fit its two buttons. (#54)
• Long prompts expand by clicking the text itself; empty states read as quiet dashed panels. (#54)

0.5.0 - 2026-08-31

Everything the review on deepseek-harness discussion #5061 asked for (issues #39 to #44), plus the hardening that went with it.

Upgrading

• Restart dsh web after updating; host code only reloads on a restart.
• Turn digests are recomputed on first load, so credential masking covers the retained history too. Nothing to set.
• Profiles migrate on read; existing directives gain their provenance fields.
• New setting reviewCandidates (off by default); new route /api/tacit/directive-receipt.

Added

• Every directive shows a receipt in Settings: where it came from, its trial, what its distillation cost, with Copy as JSON; never prompt text. (#49)
• Review new directives before their trial: with reviewCandidates on, a new directive waits until you press Start trial. (#49)
• A removed directive stays removed; the distiller cannot propose it, or a rewording of it, again. (#49)
• Analyzing a turn that already has a report costs nothing; Re-analyze pays for a fresh one. (#49)
• Move to workspace on every directive, and a not seen since chip on a workspace with no open session. (#48)
• Usage card: spend by route and by trigger, a tile for failed or repair calls, input and output tokens shown apart, the daily cap next to today. (#47)
• The trend chips link to How it works §6 and §10. (#46)
• The privacy page states that Tacit sends no telemetry of any kind. (#49)

Changed

• A workspace is one normalised path; a conversation anywhere inside it gets its directives, deepest workspace first. (#48)
• Same-name workspaces are shown to the distiller as a/web and b/web. (#48)
• At most 12 workspaces keep distilled directives; typed ones are never dropped. (#48)
• Removing or switching off a directive reaches conversations already open; adding one still waits for the next. (#49)
• The trend chips say the numbers are a before/after over your own turns, not a controlled comparison. (#46)
• A retired directive reads "corrections rose 10% → 30% during its trial". (#46)

Fixed

• Unticking a candidate frees its scope's trial slot for the next queued directive. (#49)
• Two workspaces sharing a folder name no longer cross-wire directives. (#48)
• A directive for a workspace the evidence never named is dropped and logged, not applied everywhere. (#48)
• A renamed or moved workspace no longer strands its directives; a candidate there pauses until a session opens in it again. (#48)
• ✨ Improve skips background-job notifications and bare "continue" turns when it gathers context. (#46)
• Escape out of a confirm dialog returns focus to the button that opened it. (#46)

Security

• Credential-shaped strings (API keys, tokens, JWTs, private keys) are masked as a turn is captured, before anything is stored or sent. (#45)
• The steering section states it changes no tool permission, approval or sandbox policy, and a directive about those is refused. (#45)
• Every coach prompt treats its input as evidence, never as instructions. (#45)

0.4.0 - 2026-08-31

Changed

• Directive trials are graded on how often you correct the agent; tool errors alone no longer retire a directive. (#36)
• One directive per scope is on trial at a time; the rest wait as queued. (#36)
• A reworded directive keeps its identity and trial; retired ones are remembered so they are not proposed again. (#36)
• The measured trend leads with your correction rate. (#36)
• The cost panel re-reads nothing that has not changed; per-day buckets stop at 400 days. (#37)

Fixed

• A provider out of credit says so instead of a generic failure. (#37)
• Today no longer loses a run that crossed midnight. (#37)
• The runs filter can ask for running. (#37)
• Learn from my last 20 turns no longer shows a previous batch's figures. (#37)
• The retention selector shows a costHistoryDays set by hand in YAML. (#37)
• The Pricing rate table stays in the page while collapsed. (#37)
• Tab stays inside the confirm dialog; two accessibility fixes in the run list. (#37)

0.3.0 - 2026-08-30

Added

• Every Tacit model call is metered into a local, content-free usage ledger; new costHistoryDays, costWarnDailyUsd, costWarnMonthlyUsd. (#33)
• Routes for the ledger and prices, /api/tacit/bootstrap-preview (what a bootstrap would cost) and /api/tacit/analyze-batch. (#33)
• Usage and Pricing cards in Settings: spend tiles, a daily chart, spend by operation, warnings, a run list, the rate table and the price source. (#33)
• A Data & privacy card with retention, spend warnings and confirm-gated clearing; the bootstrap button shows a live cost estimate. (#33)

Changed

• Settings → Tacit is eight collapsible cards instead of one long page. (#33)
• ✨ Improve converges in one pass; improving a finished prompt again returns it unchanged. (#31)

0.2.3 - 2026-08-29

Added

• Tacit also learns from a clean prompt that follows a messy one (learnFromGood). (#29)
• Directives can be scoped to a workspace. (#27)
• Bootstrap can run several analyses at once (bootstrapConcurrency). (#20)
• README: a Help shape Tacit section. (#10)

Changed

• Distilled directives are one sentence of at most 25 words. (#26)
• The browser bundle is generated from client/src/. (#22)
• Official @deepseek-ai/* packages are peer dependencies. (#23)

Fixed

• The 👍/👎 strip appears on a conversation's first prompt too. (#24)
• Directive trials count only turns from conversations steered by the candidate. (#21)

0.2.2 - 2026-08-29

Added

• A sanitized Settings screenshot and a short demo. (#8, #9)

Fixed

• The Improve preview shows its pending state on the first click. (#19)
• The style-rule distillation call carries the session id. (#11)
• The bootstrap hint shows the documented $0.02 to $0.05 range.

0.2.1 - 2026-08-27

Changed

• The npm page shows the English README; the Chinese one moved to docs/README.zh.md. (#2)
• Documentation overhaul: getting started, how it works, a diagram, a glossary, and claims corrected to match the code. (#3)

0.2.0 - 2026-08-27

First public release. Renamed from dsh-prompt-coach to Tacit; an existing ~/.dsh/storages/prompt-coach directory is adopted automatically. (#1)

Added

• Zero-click learning from messy turns and your own corrections.
• Learned directives injected as a system-prompt section you can read, edit, toggle and delete in Settings → Tacit.
• Directive trials: a candidate is activated or retired after directiveTrialTurns turns.
• Learn from my last 20 turns.
• A measured trend instead of a guessed savings figure.
• ✨ Improve with 👍/👎 feedback and style rules; opt-in pre-send context.

Security

• Tacit's routes refuse cross-site requests.

Removed

• The guessed "token savings %", the good-examples library and the learningThreshold gate.

0.1.x

Published as dsh-prompt-coach: manual prompt coaching gated behind a learning threshold. Superseded.