v0.8.5

v0.8.5 - 2026/06/24

fix(configure-cli): robotic humor thinking budget 0→4096 — thinking re-enabled with minimal budget instead of disabled; reasoning effort mapped to low instead of none

fix(configure-cli): introvert thinking budget 10240→8192 — aligned with industry effort-level estimates for hard tasks

feat(configure-cli): vary top_p by humor — robotic 0.7, introvert 0.75, pragmatic 0.8, sympathetic 0.85, extrovert 0.85; deterministic roles get lower top_p for code precision, exploratory roles keep higher for design flexibility

feat(configure-cli): block destructive git commands on all personas — git clean, git reset, git rebase, git push --force, git push -f explicitly denied in bash permission profiles for all personas; prevents accidental history rewrites, untracked file deletion, and force pushes

refactor(personas): bump coder modelTier from tier-1 to tier-2 — fast model with no thinking insufficient for coding tasks that require reasoning about context, dependencies, and test design

v0.8.4 - 2026/06/24

fix(self-review): adapt SHIELD dimensions for all pass types — S-dimension simplified, H-dimension generalized, E/L/D dimensions made pass-agnostic (grounding, evidence, external factors) instead of pass-specific; all dimensions now work for both full and focused reviews

v0.8.3 - 2026/06/24

fix(self-review): SHIELD S-dimension handles focused reviews — rubric now explicitly scores based on required passes (focused pass only when task specifies focus, all three when no focus); eliminates ambiguity when task scopes to single pass

v0.8.2 - 2026/06/24

fix(reviewer): restructure playbook for progress file discipline — step 3 reads skill(s) in full before proceeding, step 4 initializes progress files, step 5 executes each phase thoroughly with calming guidance; Identity adds depth-over-completeness principle; Red Line enforces progress file timing

v0.8.1 - 2026/06/24

refactor(personas): align identities with main framework — all 5 persona identities trimmed to single paragraphs; operational guidance moved to Playbook/Red Lines, only character and principles remain

v0.8.0 - 2026/06/24

refactor(reviewer): safety net identity — removed "three critics" framing; Reviewer is now the safety net that catches what was dropped, runs all three lenses (coherence, quality, security)

refactor(reviewer): trimmed Red Lines to persona-specific only — removed process-related prohibitions already covered by review skills and SHIELD rubric

refactor(review-loop): restore multi-dispatch tiers with focus — Standard (2 dispatches) and Full (3 dispatches) tiers restored for MoE context window management; focus now stated in itself, no separate <review-focus> block

refactor(self-review): remove <review-focus> references — SHIELD rubric updated for task-driven focus

v0.7.9 - 2026/06/23

refactor(configure-cli): improve agentBindingBuilder() readability — eliminated nested if/else, consolidated jq calls into linear flow with separate conditionals for temperature, top_p, and thinking; each step builds on previous result for clear data flow

feat(configure-cli): emit reasoningEffort alongside reasoning.effort — flat reasoningEffort field added for opencode pass-through compatibility per docs; both nested and flat formats emitted for max provider compatibility

v0.7.8 - 2026/06/23

feat(configure-cli): dual-format thinking config — agent bindings emit both Anthropic (thinking: {type, budgetTokens}) and OpenAI (reasoning: {effort}) formats; disable works universally across SDKs

fix(configure-cli): replace enable_thinking with proper thinking/reasoning formats — enable_thinking was wrong parameter for both Anthropic and OpenAI SDK providers

v0.7.7 - 2026/06/23

feat(personas): add Observations convention — dispatch notes instruct all personas to optionally include a ## Observations section in their handoff for honest opinions, concerns, or patterns outside their deliverable; Maestro scans and saves to long-term memory, feeds relevant observations into subsequent dispatch context

feat(memory): add Observations section to long-term memory schema — dedicated section for persona specialist feedback that falls outside deliverables but may matter later

refactor(contextualizer): humor changed from introvert to robotic — thinking disabled, same temperature/topP as introvert (0.2/0.85)

v0.7.6 - 2026/06/23

fix(memory): pre-create all .memory subdirectories in agent-memory skill — plan/, todo/, reviews/ added alongside session/; prevents failures when architect, task-tracking, or code-quality-review skills write before dirs exist

v0.7.5 - 2026/06/22

fix(configure-cli): simplify .memory permissions to single glob pattern — .memory/* and .memory/**/* replaced with .memory/** for consistent matching; prevents permission prompts for build agent editing .memory/ files

fix(configure-cli): replace jq deep merge with shallow merge — * operator preserved stale budgetTokens when thinking disabled; + replaces agent objects entirely so disabled thinking has no budget field

v0.7.4 - 2026/06/22

feat(configure-cli): add robotic humor — same temperature/topP as introvert (0.2/0.85) but thinking explicitly disabled; agentBindingBuilder handles "disabled" string for thinkingBudget

refactor(coder): tier-2 to tier-1, pragmatic to robotic — fast model with no thinking for explicit linear playbooks

v0.7.3 - 2026/06/18

refactor(architect): reduce per-phase LOC target from 1000 to 600 with 800 hard cap — aligns with Standard review tier threshold, reduces need for Full tier reviews

feat(review): update adversarial review phase size threshold from 1000 to 800 LOC — enforces new per-phase hard cap

feat(review): add LOC threshold to DRAFT self-review — F criterion now checks phase LOC against 600 soft cap and 800 hard cap

fix(review): use git diff HEAD instead of git diff for LOC measurement — captures both staged and unstaged changes, prevents staged work from bypassing review

v0.7.2

fix(coder): make style absorption deterministic — run ls on target directory, read exactly two sibling files most similar in function, match structure/patterns/conventions exactly

feat(rules): add native tooling rule — coders should use Edit/Read/Write/Grep/Glob directly, not write scripts for file operations

feat(review): add style proximity verification — reviewer reads sibling files and compares against changed code before checking rules, project style takes precedence

refactor(review): restructure code-quality-review around progress file breadcrumbs — phases tracked in .memory/reviews/, findings written incrementally, style proximity after rule walk (context window management), dedup final step

refactor(review): lower review tier LOC thresholds — Unified <300, Standard 300-600, Full 600-1000, over 1000 must split via Contextualizer

v0.7.1

fix(configure-cli): add .memory/* alongside .memory/**/* in edit and external_directory for all personas — glob ** may not match direct children of .memory/; files like settings.conf and MEMORY.md were falling through to wildcard ask/deny

feat(rules): add method granularity rule — trivial wrapper functions (~5 LOC or fewer) are indirection without value; functions must do meaningful work

feat(rules): add method ordering rule — callees must be defined above callers for top-down readability (with constructor exception)

feat(rules): clarify data trust boundary — no separate parse functions when value object constructor already validates

fix(configure-cli): add rg (ripgrep) to bash allow list for all personas — file search is a read operation, same category as grep

v0.7.0

feat(configure-cli): add text processing and file manipulation utils to persona bash allow lists — sed, awk, tr, cut, uniq, wc on all profiles; touch, cp, mv, tee, xargs, ln on write-enabled personas

fix(configure-cli): add deny guards for read-only personas — sed -i and file manipulation tools explicitly denied on architect and reviewer to prevent bypassing edit permissions

fix(tests): update stale test assertions for permission defaults and model version checks

v0.6.4

refactor(architect): reduce per-phase LOC target from 1500 to 1000 — smaller phases reduce cognitive burden and review complexity

feat(review): add phase dependency and size checklists to adversarial plan review — enforces 1000 LOC limit per phase, flags missing estimates as Blockers

v0.6.3

feat(architect): version plan files — each revision produces a new file (YYYY-MM-DD---v.md) instead of overwriting; first version is v0

v0.6.1

fix(configure-cli): change all personas from edit/bash deny to ask — write tools hidden when deny matches project root; compound commands (pipes, redirects) now prompt instead of silently failing
feat(boot): add ensureHiddenDirectoriesAreSearchable to maestro-boot-configure-cli.sh — creates .ignore at project root with !.agents/ and !.memory/ entries so AI tooling (glob/grep) can search hidden gitignored directories; ensureGitignoreEntry '.ignore' added to .gitignore

v0.6.0

Changes

• fix(personas/coder.md): restore corrupted file — all newlines were stripped into single line, adapted for starter kit
• fix(personas/maestro.md): break dense Parse step into numbered sub-list, add dispatch skill path to every dispatch reference
• fix(personas/maestro.md): add sub-agent output context to Review loop step, add task-tracking skill reference to Deliver step
• fix(personas/architect.md): remove Confidence score section — redundant with DRAFT self-review
• fix(skills/reviewer-handoff.md): remove Coverage checklist — already covered by SHIELD self-review, focus on handoff format only
• delete(skills/reviewer-scoring.md): vestigial, checklist inlined into reviewer-handoff.md
• fix(skills/assets/maestro-boot-configure-cli.sh): replace sed with awk for frontmatter and YAML extraction — macOS/BSD sed incompatibility
• docs(readme): fix duplicate preferredModel paragraph, consolidate persona section
• docs(personas/readme): bump schema to v0.1.1 for humor field addition

v0.5.9

• fix(skills/maestro-boot-configure-cli.sh): accept host modelId as $1 to resolve correct provider when multiple providers share the same CLI (deepseek/qwen both use opencode), fallback to first match when omitted
• fix(skills/boot.md): pass modelId to configure-cli.sh script
• fix(skills/maestro-boot-configure-cli_test.sh): update personas directory paths from personas to .agents/personas

What's Changed

• feat/include providers in dispatchers by @GabrielDomiciano in #3

New Contributors

• @GabrielDomiciano made their first contribution in #3

Full Changelog: https://github.com/ntorga/agent-starter-kit/commits/v0.5.9